merge v0.5.0

.buildkite/nightly-benchmarks/kickoff-pipeline.sh

+#!/usr/bin/env bash
+
+set -euo pipefail
+
+# Install system packages
+apt update
+apt install -y curl jq
+
+# Install minijinja for templating
+curl -sSfL https://github.com/mitsuhiko/minijinja/releases/latest/download/minijinja-cli-installer.sh | sh
+source $HOME/.cargo/env
+
+# If BUILDKITE_PULL_REQUEST != "false", then we check the PR labels using curl and jq
+if [ "$BUILDKITE_PULL_REQUEST" != "false" ]; then
+  PR_LABELS=$(curl -s "https://api.github.com/repos/vllm-project/vllm/pulls/$BUILDKITE_PULL_REQUEST" | jq -r '.labels[].name')
+
+  if [[ $PR_LABELS == *"perf-benchmarks"* ]]; then
+    echo "This PR has the 'perf-benchmarks' label. Proceeding with the nightly benchmarks."
+  else
+    echo "This PR does not have the 'perf-benchmarks' label. Skipping the nightly benchmarks."
+    exit 0
+  fi
+fi
+
+# Upload sample.yaml
+buildkite-agent pipeline upload .buildkite/nightly-benchmarks/sample.yaml

.buildkite/nightly-benchmarks/sample.yaml

+steps:
+  # NOTE(simon): You can create separate blocks for different jobs
+  - label: "A100: NVIDIA SMI"
+    agents:
+      queue: A100
+    plugins:
+    - kubernetes:
+        podSpec:
+          containers:
+          # - image: us-central1-docker.pkg.dev/vllm-405802/vllm-ci-test-repo/vllm-test:$BUILDKITE_COMMIT
+          # TODO(simon): check latest main branch or use the PR image.
+          - image: us-central1-docker.pkg.dev/vllm-405802/vllm-ci-test-repo/vllm-test:45c35f0d58f4508bf43bd6af1d3d0d0ec0c915e6
+            command:
+            - bash -c 'nvidia-smi && nvidia-smi topo -m && pwd && ls'
+            resources:
+              limits:
+                nvidia.com/gpu: 8
+            volumeMounts:
+            - name: devshm
+              mountPath: /dev/shm
+          nodeSelector:
+            nvidia.com/gpu.product: NVIDIA-A100-SXM4-80GB
+          volumes:
+          - name: devshm
+            emptyDir:
+              medium: Memory
+  # TODO(simon): bring H100 online
+  # - label: "H100: NVIDIA SMI"
+  #   agents:
+  #     queue: H100
+  #   plugins:
+  #   - docker#v5.11.0:
+  #       image: us-central1-docker.pkg.dev/vllm-405802/vllm-ci-test-repo/vllm-test:45c35f0d58f4508bf43bd6af1d3d0d0ec0c915e6
+  #       command:
+  #       - bash -c 'nvidia-smi && nvidia-smi topo -m'
+  #       propagate-environment: true
+  #       ipc: host
+  #       gpus: all
+

.buildkite/run-benchmarks.sh

@@ -50,16 +50,16 @@ echo "### Serving Benchmarks" >> benchmark_results.md
 sed -n '1p' benchmark_serving.txt >> benchmark_results.md # first line
 echo "" >> benchmark_results.md
 echo '```' >> benchmark_results.md
-tail -n 20 benchmark_serving.txt >> benchmark_results.md # last 20 lines
+tail -n 24 benchmark_serving.txt >> benchmark_results.md # last 24 lines
 echo '```' >> benchmark_results.md
 
 # if the agent binary is not found, skip uploading the results, exit 0
-if [ ! -f /workspace/buildkite-agent ]; then
+if [ ! -f /usr/bin/buildkite-agent ]; then
     exit 0
 fi
 
 # upload the results to buildkite
-/workspace/buildkite-agent annotate --style "info" --context "benchmark-results" < benchmark_results.md
+buildkite-agent annotate --style "info" --context "benchmark-results" < benchmark_results.md
 
 # exit with the exit code of the benchmarks
 if [ $bench_latency_exit_code -ne 0 ]; then
@@ -75,4 +75,4 @@ if [ $bench_serving_exit_code -ne 0 ]; then
 fi
 
 rm ShareGPT_V3_unfiltered_cleaned_split.json
-/workspace/buildkite-agent artifact upload "*.json"
+buildkite-agent artifact upload "*.json"

.buildkite/run-cpu-test.sh

@@ -10,5 +10,15 @@ remove_docker_container() { docker rm -f cpu-test || true; }
 trap remove_docker_container EXIT
 remove_docker_container
 
-# Run the image and launch offline inference
-docker run --network host --env VLLM_CPU_KVCACHE_SPACE=1 --name cpu-test cpu-test python3 vllm/examples/offline_inference.py
+# Run the image
+docker run -itd -v ~/.cache/huggingface:/root/.cache/huggingface --cpuset-cpus=48-95 --cpuset-mems=1 --network host -e HF_TOKEN --env VLLM_CPU_KVCACHE_SPACE=4 --name cpu-test cpu-test
+
+# offline inference
+docker exec cpu-test bash -c "python3 examples/offline_inference.py"
+
+# Run basic model test
+docker exec cpu-test bash -c "cd tests;
+  pip install pytest Pillow protobuf
+  bash ../.buildkite/download-images.sh
+  cd ../
+  pytest -v -s tests/models --ignore=tests/models/test_llava.py --ignore=tests/models/test_embedding.py --ignore=tests/models/test_registry.py"

.buildkite/test-pipeline.yaml

@@ -37,6 +37,7 @@ steps:
   working_dir: "/vllm-workspace/tests"
   num_gpus: 2
   commands:
+  - VLLM_TEST_SAME_HOST=1 torchrun --nproc-per-node=4 distributed/test_same_node.py
   - TEST_DIST_MODEL=facebook/opt-125m DISTRIBUTED_EXECUTOR_BACKEND=ray pytest -v -s distributed/test_basic_distributed_correctness.py
   - TEST_DIST_MODEL=meta-llama/Llama-2-7b-hf DISTRIBUTED_EXECUTOR_BACKEND=ray pytest -v -s distributed/test_basic_distributed_correctness.py
   - TEST_DIST_MODEL=facebook/opt-125m DISTRIBUTED_EXECUTOR_BACKEND=ray pytest -v -s distributed/test_chunked_prefill_distributed.py
@@ -46,6 +47,7 @@ steps:
   - TEST_DIST_MODEL=facebook/opt-125m DISTRIBUTED_EXECUTOR_BACKEND=mp pytest -v -s distributed/test_chunked_prefill_distributed.py
   - TEST_DIST_MODEL=meta-llama/Llama-2-7b-hf DISTRIBUTED_EXECUTOR_BACKEND=mp pytest -v -s distributed/test_chunked_prefill_distributed.py
   - pytest -v -s spec_decode/e2e/test_integration_dist.py
+  - CUDA_VISIBLE_DEVICES=0,1 pytest -v -s test_sharded_state_loader.py
 
 - label: Distributed Tests (Multiple Groups)
   #mirror_hardwares: [amd]
@@ -62,7 +64,6 @@ steps:
   mirror_hardwares: [amd]
 
   commands:
-  - pytest -v -s test_inputs.py
   - pytest -v -s entrypoints -m llm
   - pytest -v -s entrypoints -m openai
 
@@ -79,6 +80,13 @@ steps:
     - python3 llava_example.py
     - python3 tensorize_vllm_model.py --model facebook/opt-125m serialize --serialized-directory /tmp/ --suffix v1 && python3 tensorize_vllm_model.py --model facebook/opt-125m deserialize --path-to-tensors /tmp/vllm/facebook/opt-125m/v1/model.tensors
 
+- label: Inputs Test
+  #mirror_hardwares: [amd]
+  commands:
+    - bash ../.buildkite/download-images.sh
+    - pytest -v -s test_inputs.py
+    - pytest -v -s multimodal
+
 - label: Kernels Test %N
   #mirror_hardwares: [amd]
   command: pytest -v -s kernels --shard-id=$$BUILDKITE_PARALLEL_JOB --num-shards=$$BUILDKITE_PARALLEL_JOB_COUNT
@@ -87,14 +95,13 @@ steps:
 - label: Models Test
   #mirror_hardwares: [amd]
   commands:
-    - bash ../.buildkite/download-images.sh
-    - pytest -v -s models --ignore=models/test_llava.py
+    - pytest -v -s models -m \"not llava\"
 
 - label: Llava Test
   mirror_hardwares: [amd]
   commands:
     - bash ../.buildkite/download-images.sh
-    - pytest -v -s models/test_llava.py
+    - pytest -v -s models -m llava
 
 - label: Prefix Caching Test
   mirror_hardwares: [amd]
@@ -118,7 +125,10 @@ steps:
 
 - label: Speculative decoding tests
   #mirror_hardwares: [amd]
-  command: pytest -v -s spec_decode
+  commands:
+    # See https://github.com/vllm-project/vllm/issues/5152
+    - export VLLM_ATTENTION_BACKEND=XFORMERS
+    - pytest -v -s spec_decode
 
 - label: LoRA Test %N
   #mirror_hardwares: [amd]
@@ -130,14 +140,7 @@ steps:
   num_gpus: 4
   # This test runs llama 13B, so it is required to run on 4 GPUs.
   commands:
-    # Temporarily run this way because we cannot clean up GPU mem usage
-    # for multi GPU tests.
-    # TODO(sang): Fix it.
-    - pytest -v -s lora/test_long_context.py::test_rotary_emb_replaced
-    - pytest -v -s lora/test_long_context.py::test_batched_rope_kernel
-    - pytest -v -s lora/test_long_context.py::test_self_consistency
-    - pytest -v -s lora/test_long_context.py::test_quality
-    - pytest -v -s lora/test_long_context.py::test_max_len
+    - pytest -v -s -x lora/test_long_context.py
 
 - label: Tensorizer Test
   #mirror_hardwares: [amd]

.buildkite/test-template-aws.j2

+{% set docker_image = "public.ecr.aws/q9t5s3a7/vllm-ci-test-repo:$BUILDKITE_COMMIT" %}
+{% set default_working_dir = "/vllm-workspace/tests" %}
+
+steps:
+  - label: ":docker: build image"
+    agents:
+      queue: cpu_queue
+    commands:
+      - "aws ecr-public get-login-password --region us-east-1 | docker login --username AWS --password-stdin public.ecr.aws/q9t5s3a7"
+      - "docker build --build-arg max_jobs=16 --tag {{ docker_image }} --target test --progress plain ."
+      - "docker push {{ docker_image }}"
+    env:
+      DOCKER_BUILDKIT: "1"
+    retry:
+      automatic:
+        - exit_status: -1  # Agent was lost
+          limit: 5
+        - exit_status: -10  # Agent was lost
+          limit: 5
+  - wait
+
+  {% for step in steps %}
+  - label: "{{ step.label }}"
+    agents:
+      {% if step.label == "Documentation Build" %}
+      queue: small_cpu_queue
+      {% elif step.no_gpu %}
+      queue: cpu_queue
+      {% elif step.num_gpus == 2 or step.num_gpus == 4 %}
+      queue: gpu_4_queue
+      {% else %}
+      queue: gpu_1_queue
+      {% endif %}
+    soft_fail: true
+    {% if step.parallelism %}
+    parallelism: {{ step.parallelism }}
+    {% endif %}
+    retry:
+      automatic:
+        - exit_status: -1  # Agent was lost
+          limit: 5
+        - exit_status: -10  # Agent was lost
+          limit: 5
+    plugins:
+      - docker#v5.2.0:
+          image: {{ docker_image }}
+          always-pull: true
+          propagate-environment: true
+          {% if not step.no_gpu %}
+          gpus: all
+          {% endif %}
+          {% if step.label == "Benchmarks" %}
+          mount-buildkite-agent: true
+          {% endif %}
+          command: ["bash", "-c", "cd {{ (step.working_dir or default_working_dir) | safe  }} && {{ step.command  or (step.commands | join(' && ')) | safe }}"]
+          environment:
+            - VLLM_USAGE_SOURCE=ci-test
+            - HF_TOKEN
+            {% if step.label == "Speculative decoding tests" %}
+            - VLLM_ATTENTION_BACKEND=XFORMERS
+            {% endif %}
+          volumes:
+            - /dev/shm:/dev/shm
+  {% endfor %}

.buildkite/test-template.j2

@@ -28,6 +28,7 @@ steps:
         command: bash .buildkite/run-amd-test.sh "cd {{ (step.working_dir or default_working_dir) | safe  }} ; {{ step.command  or (step.commands | join(" ; ")) | safe }}"
         env:
           DOCKER_BUILDKIT: "1"
+        soft_fail: true
     {% endif %}
     {% endfor %}
 
@@ -36,10 +37,12 @@ steps:
     agents:
       queue: neuron
     command: bash .buildkite/run-neuron-test.sh
-    soft_fail: true
+    soft_fail: false
 
   - label: "Intel Test"
     depends_on: ~
+    agents:
+      queue: intel
     command: bash .buildkite/run-cpu-test.sh
 
   {% for step in steps %}

.github/workflows/mypy.yaml

@@ -37,6 +37,7 @@ jobs:
         mypy vllm/distributed --config-file pyproject.toml
         mypy vllm/entrypoints --config-file pyproject.toml
         mypy vllm/executor --config-file pyproject.toml
+        mypy vllm/multimodal --config-file pyproject.toml
         mypy vllm/usage --config-file pyproject.toml
         mypy vllm/*.py --config-file pyproject.toml
         mypy vllm/transformers_utils --config-file pyproject.toml

CMakeLists.txt

@@ -67,19 +67,6 @@ endif()
 #
 find_package(Torch REQUIRED)
 
-#
-# Normally `torch.utils.cpp_extension.CUDAExtension` would add
-# `libtorch_python.so` for linking against an extension. Torch's cmake
-# configuration does not include this library (presumably since the cmake
-# config is used for standalone C++ binaries that link against torch).
-# The `libtorch_python.so` library defines some of the glue code between
-# torch/python via pybind and is required by VLLM extensions for this
-# reason. So, add it by manually with `find_library` using torch's
-# installed library path.
-#
-find_library(torch_python_LIBRARY torch_python PATHS
-  "${TORCH_INSTALL_PREFIX}/lib")
-
 #
 # Forward the non-CUDA device extensions to external CMake scripts.
 #
@@ -172,7 +159,7 @@ set(VLLM_EXT_SRC
   # "csrc/quantization/fp8/common.cu"
   "csrc/cuda_utils_kernels.cu"
   "csrc/moe_align_block_size_kernels.cu"
-  "csrc/pybind.cpp")
+  "csrc/torch_bindings.cpp")
 
 if(VLLM_GPU_LANG STREQUAL "CUDA")
   include(FetchContent)
@@ -219,6 +206,7 @@ define_gpu_extension_target(
   COMPILE_FLAGS ${VLLM_GPU_FLAGS}
   ARCHITECTURES ${VLLM_GPU_ARCHES}
   INCLUDE_DIRECTORIES ${CUTLASS_INCLUDE_DIR};${CUTLASS_TOOLS_UTIL_INCLUDE_DIR}
+  USE_SABI 3
   WITH_SOABI)
 
 #
@@ -226,7 +214,7 @@ define_gpu_extension_target(
 #
 
 set(VLLM_MOE_EXT_SRC
-  "csrc/moe/moe_ops.cpp"
+  "csrc/moe/torch_bindings.cpp"
   "csrc/moe/topk_softmax_kernels.cu")
 
 define_gpu_extension_target(
@@ -236,6 +224,7 @@ define_gpu_extension_target(
   SOURCES ${VLLM_MOE_EXT_SRC}
   COMPILE_FLAGS ${VLLM_GPU_FLAGS}
   ARCHITECTURES ${VLLM_GPU_ARCHES}
+  USE_SABI 3
   WITH_SOABI)
 
 #
@@ -250,7 +239,7 @@ set(VLLM_PUNICA_EXT_SRC
   "csrc/punica/bgmv/bgmv_fp32_bf16_bf16.cu"
   "csrc/punica/bgmv/bgmv_fp32_fp16_fp16.cu"
   "csrc/punica/punica_ops.cu"
-  "csrc/punica/punica_pybind.cpp")
+  "csrc/punica/torch_bindings.cpp")
 
 #
 # Copy GPU compilation flags+update for punica
@@ -287,6 +276,7 @@ if (VLLM_PUNICA_GPU_ARCHES)
     SOURCES ${VLLM_PUNICA_EXT_SRC}
     COMPILE_FLAGS ${VLLM_PUNICA_GPU_FLAGS}
     ARCHITECTURES ${VLLM_PUNICA_GPU_ARCHES}
+    USE_SABI 3
     WITH_SOABI)
 else()
   message(WARNING "Unable to create _punica_C target because none of the "
@@ -312,6 +302,9 @@ if(VLLM_GPU_LANG STREQUAL "CUDA" OR VLLM_GPU_LANG STREQUAL "HIP")
   message(STATUS "Enabling C extension.")
   add_dependencies(default _C)
 
+  message(STATUS "Enabling moe extension.")
+  add_dependencies(default _moe_C)
+
   # Enable punica if -DVLLM_INSTALL_PUNICA_KERNELS=ON or
   # VLLM_INSTALL_PUNICA_KERNELS is set in the environment and
   # there are supported target arches.
@@ -321,8 +314,3 @@ if(VLLM_GPU_LANG STREQUAL "CUDA" OR VLLM_GPU_LANG STREQUAL "HIP")
     add_dependencies(default _punica_C)
   endif()
 endif()
-
-if(VLLM_GPU_LANG STREQUAL "CUDA")
-  message(STATUS "Enabling moe extension.")
-  add_dependencies(default _moe_C)
-endif()

Dockerfile.cpu

 # This vLLM Dockerfile is used to construct image that can build and run vLLM on x86 CPU platform.
 
-FROM ubuntu:22.04
+FROM ubuntu:22.04 AS cpu-test-1
 
 RUN apt-get update  -y \
     && apt-get install -y git wget vim numactl gcc-12 g++-12 python3 python3-pip \
     && update-alternatives --install /usr/bin/gcc gcc /usr/bin/gcc-12 10 --slave /usr/bin/g++ g++ /usr/bin/g++-12
 
 RUN pip install --upgrade pip \
-    && pip install wheel packaging ninja setuptools>=49.4.0 numpy
+    && pip install wheel packaging ninja "setuptools>=49.4.0" numpy
+
+FROM cpu-test-1 AS build
 
 COPY ./ /workspace/vllm
 
@@ -19,4 +21,6 @@ RUN VLLM_TARGET_DEVICE=cpu python3 setup.py install
 
 WORKDIR /workspace/
 
+RUN ln -s /workspace/vllm/tests  && ln -s /workspace/vllm/examples && ln -s /workspace/vllm/benchmarks
+
 CMD ["/bin/bash"]

Dockerfile.neuron

@@ -28,7 +28,7 @@ COPY ./requirements-neuron.txt /app/vllm/requirements-neuron.txt
 RUN cd /app/vllm \
     && python3 -m pip install -U -r requirements-neuron.txt
 
-ENV VLLM_BUILD_WITH_NEURON 1
+ENV VLLM_TARGET_DEVICE neuron
 RUN cd /app/vllm \
     && pip install -e . \
     && cd ..

Dockerfile.rocm

@@ -106,8 +106,9 @@ RUN --mount=type=cache,target=/root/.cache/pip \
     pip install -U -r requirements-rocm.txt \
     && patch /opt/rocm/include/hip/amd_detail/amd_hip_bf16.h ./rocm_patch/rocm_bf16.patch \
     && python3 setup.py install \
-    && cp build/lib.linux-x86_64-cpython-39/vllm/_C.cpython-39-x86_64-linux-gnu.so vllm/ \
-    && cp build/lib.linux-x86_64-cpython-39/vllm/_punica_C.cpython-39-x86_64-linux-gnu.so vllm/ \
+    && cp build/lib.linux-x86_64-cpython-39/vllm/_C.abi3.so vllm/ \
+    && cp build/lib.linux-x86_64-cpython-39/vllm/_punica_C.abi3.so vllm/ \
+    && cp build/lib.linux-x86_64-cpython-39/vllm/_moe_C.abi3.so vllm/ \
     && cd ..
 
 

README_ORIGIN.md

@@ -16,6 +16,13 @@ Easy, fast, and cheap LLM serving for everyone
 
 ---
 
+**Ray Summit CPF is Open (June 4th to June 20th)!**
+
+There will be a track for vLLM at the Ray Summit (09/30-10/02, SF) this year!
+If you have cool projects related to vLLM or LLM inference, we would love to see your proposals.
+This will be a great chance for everyone in the community to get together and learn.
+Please submit your proposal [here](https://raysummit.anyscale.com/flow/anyscale/raysummit2024/landing/page/eventsite)
+
 **The Fourth vLLM Bay Area Meetup (June 11th 5:30pm-8pm PT)**
 
 We are thrilled to announce our fourth vLLM Meetup!
@@ -107,6 +114,7 @@ vLLM is a community project. Our compute resources for development and testing a
 - Replicate
 - Roblox
 - RunPod
+- Sequoia Capital
 - Trainy
 - UC Berkeley
 - UC San Diego

benchmarks/benchmark_latency.py

@@ -36,7 +36,8 @@ def main(args: argparse.Namespace):
               enable_chunked_prefill=args.enable_chunked_prefill,
               download_dir=args.download_dir,
               block_size=args.block_size,
-              gpu_memory_utilization=args.gpu_memory_utilization)
+              gpu_memory_utilization=args.gpu_memory_utilization,
+              distributed_executor_backend=args.distributed_executor_backend)
 
     sampling_params = SamplingParams(
         n=args.n,
@@ -221,5 +222,12 @@ if __name__ == '__main__':
                         help='the fraction of GPU memory to be used for '
                         'the model executor, which can range from 0 to 1.'
                         'If unspecified, will use the default value of 0.9.')
+    parser.add_argument(
+        '--distributed-executor-backend',
+        choices=['ray', 'mp'],
+        default=None,
+        help='Backend to use for distributed serving. When more than 1 GPU '
+        'is used, will be automatically set to "ray" if installed '
+        'or "mp" (multiprocessing) otherwise.')
     args = parser.parse_args()
     main(args)

benchmarks/benchmark_serving.py

@@ -56,6 +56,9 @@ class BenchmarkMetrics:
     mean_tpot_ms: float
     median_tpot_ms: float
     p99_tpot_ms: float
+    mean_itl_ms: float
+    median_itl_ms: float
+    p99_itl_ms: float
 
 
 def sample_sharegpt_requests(
@@ -200,16 +203,24 @@ def calculate_metrics(
     actual_output_lens = []
     total_input = 0
     completed = 0
+    itls = []
     tpots = []
     ttfts = []
     for i in range(len(outputs)):
         if outputs[i].success:
-            output_len = len(tokenizer(outputs[i].generated_text).input_ids)
+            # We use the tokenizer to count the number of output tokens for all
+            # serving backends instead of looking at len(outputs[i].itl) since
+            # multiple output tokens may be bundled together
+            # Note: this may inflate the output token count slightly
+            output_len = len(
+                tokenizer(outputs[i].generated_text,
+                          add_special_tokens=False).input_ids)
             actual_output_lens.append(output_len)
             total_input += input_requests[i][1]
             if output_len > 1:
                 tpots.append(
                     (outputs[i].latency - outputs[i].ttft) / (output_len - 1))
+            itls += outputs[i].itl
             ttfts.append(outputs[i].ttft)
             completed += 1
         else:
@@ -234,6 +245,9 @@ def calculate_metrics(
         mean_tpot_ms=np.mean(tpots or 0) * 1000,
         median_tpot_ms=np.median(tpots or 0) * 1000,
         p99_tpot_ms=np.percentile(tpots or 0, 99) * 1000,
+        mean_itl_ms=np.mean(itls or 0) * 1000,
+        median_itl_ms=np.median(itls or 0) * 1000,
+        p99_itl_ms=np.percentile(itls or 0, 99) * 1000,
     )
 
     return metrics, actual_output_lens
@@ -333,6 +347,10 @@ async def benchmark(
     print("{:<40} {:<10.2f}".format("Median TPOT (ms):",
                                     metrics.median_tpot_ms))
     print("{:<40} {:<10.2f}".format("P99 TPOT (ms):", metrics.p99_tpot_ms))
+    print("{s:{c}^{n}}".format(s='Inter-token Latency', n=50, c='-'))
+    print("{:<40} {:<10.2f}".format("Mean ITL (ms):", metrics.mean_itl_ms))
+    print("{:<40} {:<10.2f}".format("Median ITL (ms):", metrics.median_itl_ms))
+    print("{:<40} {:<10.2f}".format("P99 ITL (ms):", metrics.p99_itl_ms))
     print("=" * 50)
 
     result = {
@@ -349,6 +367,9 @@ async def benchmark(
         "mean_tpot_ms": metrics.mean_tpot_ms,
         "median_tpot_ms": metrics.median_tpot_ms,
         "p99_tpot_ms": metrics.p99_tpot_ms,
+        "mean_itl_ms": metrics.mean_itl_ms,
+        "median_itl_ms": metrics.median_itl_ms,
+        "p99_itl_ms": metrics.p99_itl_ms,
         "input_lens": [output.prompt_len for output in outputs],
         "output_lens": actual_output_lens,
         "ttfts": [output.ttft for output in outputs],

benchmarks/benchmark_throughput.py

@@ -78,6 +78,7 @@ def run_vllm(
     enable_prefix_caching: bool,
     enable_chunked_prefill: bool,
     max_num_batched_tokens: int,
+    distributed_executor_backend: Optional[str],
     gpu_memory_utilization: float = 0.9,
     download_dir: Optional[str] = None,
 ) -> float:
@@ -100,6 +101,7 @@ def run_vllm(
         download_dir=download_dir,
         enable_chunked_prefill=enable_chunked_prefill,
         max_num_batched_tokens=max_num_batched_tokens,
+        distributed_executor_backend=distributed_executor_backend,
     )
 
     # Add the requests to the engine.
@@ -225,8 +227,8 @@ def main(args: argparse.Namespace):
             args.enforce_eager, args.kv_cache_dtype,
             args.quantization_param_path, args.device,
             args.enable_prefix_caching, args.enable_chunked_prefill,
-            args.max_num_batched_tokens, args.gpu_memory_utilization,
-            args.download_dir)
+            args.max_num_batched_tokens, args.distributed_executor_backend,
+            args.gpu_memory_utilization, args.download_dir)
     elif args.backend == "hf":
         assert args.tensor_parallel_size == 1
         elapsed_time = run_hf(requests, args.model, tokenizer, args.n,
@@ -375,6 +377,13 @@ if __name__ == "__main__":
         type=str,
         default=None,
         help='Path to save the throughput results in JSON format.')
+    parser.add_argument(
+        '--distributed-executor-backend',
+        choices=['ray', 'mp'],
+        default=None,
+        help='Backend to use for distributed serving. When more than 1 GPU '
+        'is used, will be automatically set to "ray" if installed '
+        'or "mp" (multiprocessing) otherwise.')
     args = parser.parse_args()
     if args.tokenizer is None:
         args.tokenizer = args.model

benchmarks/cutlass_benchmarks/w8a8_benchmarks.py

+import argparse
+import copy
+import itertools
+import pickle as pkl
+import time
+from typing import Callable, Iterable, List, Tuple
+
+import torch
+import torch.utils.benchmark as TBenchmark
+from torch.utils.benchmark import Measurement as TMeasurement
+from weight_shapes import WEIGHT_SHAPES
+
+from vllm import _custom_ops as ops
+
+DEFAULT_MODELS = list(WEIGHT_SHAPES.keys())[1:]
+DEFAULT_BATCH_SIZES = [1, 16, 32, 64, 128, 256, 512]
+DEFAULT_TP_SIZES = [1]
+
+# helpers
+
+
+def to_fp8(tensor: torch.tensor) -> torch.tensor:
+    finfo = torch.finfo(torch.float8_e4m3fn)
+    return torch.round(tensor.clamp(
+        min=finfo.min, max=finfo.max)).to(dtype=torch.float8_e4m3fn)
+
+
+def to_int8(tensor: torch.tensor) -> torch.tensor:
+    return torch.round(tensor.clamp(min=-128, max=127)).to(dtype=torch.int8)
+
+
+def make_rand_tensors(dtype: torch.dtype, m: int, n: int,
+                      k: int) -> Tuple[torch.tensor, torch.tensor]:
+
+    a = torch.randn((m, k), device='cuda') * 5
+    b = torch.randn((n, k), device='cuda').t() * 5
+
+    if dtype == torch.int8:
+        return to_int8(a), to_int8(b)
+    if dtype == torch.float8_e4m3fn:
+        return to_fp8(a), to_fp8(b)
+
+    raise ValueError("unsupported dtype")
+
+
+# impl
+
+
+def pytorch_i8_impl(a: torch.tensor, b: torch.tensor, scale_a: torch.tensor,
+                    scale_b: torch.tensor,
+                    out_dtype: torch.dtype) -> torch.tensor:
+    return torch.mm(a, b)
+
+
+def pytorch_fp8_impl(a: torch.tensor, b: torch.tensor, scale_a: torch.tensor,
+                     scale_b: torch.tensor,
+                     out_dtype: torch.dtype) -> torch.tensor:
+    return torch._scaled_mm(a,
+                            b,
+                            scale_a=scale_a,
+                            scale_b=scale_b,
+                            out_dtype=out_dtype)
+
+
+def pytorch_fp8_impl_fast_accum(a: torch.tensor, b: torch.tensor,
+                                scale_a: torch.tensor, scale_b: torch.tensor,
+                                out_dtype: torch.dtype) -> torch.tensor:
+    return torch._scaled_mm(a,
+                            b,
+                            scale_a=scale_a,
+                            scale_b=scale_b,
+                            out_dtype=out_dtype,
+                            use_fast_accum=True)
+
+
+def cutlass_impl(a: torch.tensor, b: torch.tensor, scale_a: torch.tensor,
+                 scale_b: torch.tensor,
+                 out_dtype: torch.dtype) -> torch.tensor:
+    return ops.cutlass_scaled_mm_dq(a,
+                                    b,
+                                    scale_a,
+                                    scale_b,
+                                    out_dtype=out_dtype)
+
+
+# bench
+def bench_fn(a: torch.tensor, b: torch.tensor, scale_a: torch.tensor,
+             scale_b: torch.tensor, out_dtype: torch.dtype, label: str,
+             sub_label: str, fn: Callable, description: str) -> TMeasurement:
+
+    min_run_time = 1
+
+    globals = {
+        "a": a,
+        "b": b,
+        "scale_a": scale_a,
+        "scale_b": scale_b,
+        "out_dtype": out_dtype,
+        "fn": fn,
+    }
+    return TBenchmark.Timer(
+        stmt="fn(a, b, scale_a, scale_b, out_dtype)",
+        globals=globals,
+        label=label,
+        sub_label=sub_label,
+        description=description,
+    ).blocked_autorange(min_run_time=min_run_time)
+
+
+def bench_int8(dtype: torch.dtype, m: int, k: int, n: int, label: str,
+               sub_label: str) -> Iterable[TMeasurement]:
+    assert dtype == torch.int8
+    a, b = make_rand_tensors(torch.int8, m, n, k)
+    scale_a = torch.tensor(1.0, device="cuda", dtype=torch.float32)
+    scale_b = torch.tensor(1.0, device="cuda", dtype=torch.float32)
+
+    timers = []
+    # pytorch impl
+    timers.append(
+        bench_fn(a.to(dtype=torch.bfloat16, device="cuda"),
+                 b.to(dtype=torch.bfloat16, device="cuda"), scale_a, scale_b,
+                 torch.bfloat16, label, sub_label, pytorch_i8_impl,
+                 "pytorch_bf16_bf16_bf16_matmul-no-scales"))
+
+    # cutlass impl
+    timers.append(
+        bench_fn(a, b, scale_a.to(device="cpu"), scale_b.to(device="cpu"),
+                 torch.bfloat16, label, sub_label, cutlass_impl,
+                 "cutlass_i8_i8_bf16_scaled_mm"))
+
+    return timers
+
+
+def bench_fp8(dtype: torch.dtype, m: int, k: int, n: int, label: str,
+              sub_label: str) -> Iterable[TMeasurement]:
+    assert dtype == torch.float8_e4m3fn
+    a, b = make_rand_tensors(torch.float8_e4m3fn, m, n, k)
+    scale_a = torch.tensor(1.0, device="cuda", dtype=torch.float32)
+    scale_b = torch.tensor(1.0, device="cuda", dtype=torch.float32)
+
+    timers = []
+
+    # pytorch impl: bf16 output, without fp8 fast accum
+    timers.append(
+        bench_fn(a, b, scale_a, scale_b, torch.bfloat16, label, sub_label,
+                 pytorch_fp8_impl, "pytorch_fp8_fp8_bf16_scaled_mm"))
+
+    # pytorch impl: bf16 output, with fp8 fast accum
+    timers.append(
+        bench_fn(a, b, scale_a, scale_b, torch.bfloat16, label, sub_label,
+                 pytorch_fp8_impl_fast_accum,
+                 "pytorch_fp8_fp8_bf16_scaled_mm_fast_accum"))
+
+    # pytorch impl: fp16 output, without fp8 fast accum
+    timers.append(
+        bench_fn(a, b, scale_a, scale_b, torch.float16, label, sub_label,
+                 pytorch_fp8_impl, "pytorch_fp8_fp8_fp16_scaled_mm"))
+
+    # pytorch impl: fp16 output, with fp8 fast accum
+    timers.append(
+        bench_fn(a, b, scale_a, scale_b, torch.float16, label, sub_label,
+                 pytorch_fp8_impl_fast_accum,
+                 "pytorch_fp8_fp8_fp16_scaled_mm_fast_accum"))
+
+    # cutlass impl: bf16 output
+    timers.append(
+        bench_fn(a, b, scale_a.to(device="cpu"), scale_b.to(device="cpu"),
+                 torch.bfloat16, label, sub_label, cutlass_impl,
+                 "cutlass_fp8_fp8_bf16_scaled_mm"))
+    # cutlass impl: fp16 output
+    timers.append(
+        bench_fn(a, b, scale_a.to(device="cpu"), scale_b.to(device="cpu"),
+                 torch.float16, label, sub_label, cutlass_impl,
+                 "cutlass_fp8_fp8_fp16_scaled_mm"))
+    return timers
+
+
+def bench(dtype: torch.dtype, m: int, k: int, n: int, label: str,
+          sub_label: str) -> Iterable[TMeasurement]:
+    if dtype == torch.int8:
+        return bench_int8(dtype, m, k, n, label, sub_label)
+    if dtype == torch.float8_e4m3fn:
+        return bench_fp8(dtype, m, k, n, label, sub_label)
+    raise ValueError("unsupported type")
+
+
+# runner
+def print_timers(timers: Iterable[TMeasurement]):
+    compare = TBenchmark.Compare(timers)
+    compare.print()
+
+
+def run(dtype: torch.dtype,
+        MKNs: Iterable[Tuple[int, int, int]]) -> Iterable[TMeasurement]:
+
+    results = []
+    for m, k, n in MKNs:
+        timers = bench(dtype, m, k, n, f"scaled-{dtype}-gemm",
+                       f"MKN=({m}x{k}x{n})")
+        print_timers(timers)
+        results.extend(timers)
+
+    return results
+
+
+# output makers
+def make_output(data: Iterable[TMeasurement],
+                MKNs: Iterable[Tuple[int, int, int]],
+                base_description: str,
+                timestamp=None):
+
+    print(f"== All Results {base_description} ====")
+    print_timers(data)
+
+    # pickle all the results
+    timestamp = int(time.time()) if timestamp is None else timestamp
+    with open(f"{base_description}-{timestamp}.pkl", "wb") as f:
+        pkl.dump(data, f)
+
+
+# argparse runners
+
+
+def run_square_bench(args):
+    dim_sizes = list(
+        range(args.dim_start, args.dim_end + 1, args.dim_increment))
+    MKNs = list(zip(dim_sizes, dim_sizes, dim_sizes))
+    data = run(args.dtype, MKNs)
+
+    make_output(data, MKNs, f"square_bench-{args.dtype}")
+
+
+def run_range_bench(args):
+    dim_sizes = list(range(args.dim_start, args.dim_end, args.dim_increment))
+    n = len(dim_sizes)
+    Ms = [args.m_constant] * n if args.m_constant is not None else dim_sizes
+    Ks = [args.k_constant] * n if args.k_constant is not None else dim_sizes
+    Ns = [args.n_constant] * n if args.n_constant is not None else dim_sizes
+    MKNs = list(zip(Ms, Ks, Ns))
+    data = run(args.dtype, MKNs)
+
+    make_output(data, MKNs, f"range_bench-{args.dtype}")
+
+
+def run_model_bench(args):
+
+    print("Benchmarking models:")
+    for i, model in enumerate(args.models):
+        print(f"[{i}]  {model}")
+
+    def model_shapes(model_name: str, tp_size: int) -> List[Tuple[int, int]]:
+        KNs = []
+        for KN, tp_split_dim in copy.deepcopy(WEIGHT_SHAPES[model_name]):
+            KN[tp_split_dim] = KN[tp_split_dim] // tp_size
+            KNs.append(KN)
+        return KNs
+
+    model_bench_data = []
+    models_tps = list(itertools.product(args.models, args.tp_sizes))
+    for model, tp_size in models_tps:
+        Ms = args.batch_sizes
+        KNs = model_shapes(model, tp_size)
+        MKNs = []
+        for m in Ms:
+            for k, n in KNs:
+                MKNs.append((m, k, n))
+
+        data = run(args.dtype, MKNs)
+        model_bench_data.append(data)
+
+    # Print all results
+    for data, model_tp in zip(model_bench_data, models_tps):
+        model, tp_size = model_tp
+        print(f"== Results {args.dtype} {model}-TP{tp_size} ====")
+        print_timers(data)
+
+    timestamp = int(time.time())
+
+    all_data = []
+    for d in model_bench_data:
+        all_data.extend(d)
+    # pickle all data
+    with open(f"model_bench-{args.dtype}-{timestamp}.pkl", "wb") as f:
+        pkl.dump(all_data, f)
+
+
+if __name__ == '__main__':
+
+    def to_torch_dtype(dt):
+        if dt == "int8":
+            return torch.int8
+        if dt == "fp8":
+            return torch.float8_e4m3fn
+        raise ValueError("unsupported dtype")
+
+    parser = argparse.ArgumentParser(
+        description="""
+Benchmark Cutlass GEMM.
+
+    To run square GEMMs:
+        python3 ./benchmarks/cutlass_benchmarks/w8a8_benchmarks.py --dtype fp8 square_bench --dim-start 128 --dim-end 512 --dim-increment 64
+    
+    To run constant N and K and sweep M:
+        python3 ./benchmarks/cutlass_benchmarks/w8a8_benchmarks.py --dtype fp8 range_bench --dim-start 128 --dim-end 512 --dim-increment 64 --n-constant 16384 --k-constant 16384
+    
+    To run dimensions from a model:
+        python3 ./benchmarks/cutlass_benchmarks/w8a8_benchmarks.py --dtype fp8 model_bench --models meta-llama/Llama-2-7b-hf --batch-sizes 16 --tp-sizes 1
+    
+    Output:
+        - a .pkl file, that is a list of raw torch.benchmark.utils.Measurements for the pytorch and cutlass implementations for the various GEMMs.
+            """,  # noqa: E501
+        formatter_class=argparse.RawTextHelpFormatter)
+
+    parser.add_argument("--dtype",
+                        type=to_torch_dtype,
+                        required=True,
+                        help="Available options are ['int8', 'fp8']")
+    subparsers = parser.add_subparsers(dest="cmd")
+
+    square_parser = subparsers.add_parser("square_bench")
+    square_parser.add_argument("--dim-start", type=int, required=True)
+    square_parser.add_argument("--dim-end", type=int, required=True)
+    square_parser.add_argument("--dim-increment", type=int, required=True)
+    square_parser.set_defaults(func=run_square_bench)
+
+    range_parser = subparsers.add_parser("range_bench")
+    range_parser.add_argument("--dim-start", type=int, required=True)
+    range_parser.add_argument("--dim-end", type=int, required=True)
+    range_parser.add_argument("--dim-increment", type=int, required=True)
+    range_parser.add_argument("--m-constant", type=int, default=None)
+    range_parser.add_argument("--n-constant", type=int, default=None)
+    range_parser.add_argument("--k-constant", type=int, default=None)
+    range_parser.set_defaults(func=run_range_bench)
+
+    model_parser = subparsers.add_parser("model_bench")
+    model_parser.add_argument("--models",
+                              nargs="+",
+                              type=str,
+                              default=DEFAULT_MODELS,
+                              choices=WEIGHT_SHAPES.keys())
+    model_parser.add_argument("--tp-sizes",
+                              nargs="+",
+                              type=int,
+                              default=DEFAULT_TP_SIZES)
+    model_parser.add_argument("--batch-sizes",
+                              nargs="+",
+                              type=int,
+                              default=DEFAULT_BATCH_SIZES)
+    model_parser.set_defaults(func=run_model_bench)
+
+    args = parser.parse_args()
+    args.func(args)

benchmarks/cutlass_benchmarks/weight_shapes.py

+# Weight Shapes are in the format
+# ([K, N], TP_SPLIT_DIM)
+# Example:
+#  A shape of ([14336, 4096], 0) indicates the following GEMM shape,
+#   - TP1 : K = 14336, N = 4096
+#   - TP2 : K = 7168, N = 4096
+#  A shape of ([4096, 6144], 1) indicates the following GEMM shape,
+#   - TP1 : K = 4096, N = 6144
+#   - TP4 : K = 4096, N = 1536
+
+# TP1 shapes
+WEIGHT_SHAPES = {
+    "mistralai/Mistral-7B-v0.1": [
+        ([4096, 6144], 1),
+        ([4096, 4096], 0),
+        ([4096, 28672], 1),
+        ([14336, 4096], 0),
+    ],
+    "meta-llama/Llama-2-7b-hf": [
+        ([4096, 12288], 1),
+        ([4096, 4096], 0),
+        ([4096, 22016], 1),
+        ([11008, 4096], 0),
+    ],
+    "meta-llama/Llama-2-13b-hf": [
+        ([5120, 15360], 1),
+        ([5120, 5120], 0),
+        ([5120, 27648], 1),
+        ([13824, 5120], 0),
+    ],
+    "meta-llama/Llama-2-70b-hf": [
+        ([8192, 10240], 1),
+        ([8192, 8192], 0),
+        ([8192, 57344], 1),
+        ([28672, 8192], 0),
+    ],
+}

benchmarks/kernels/benchmark_mixtral_moe.py

-import argparse
-import json
-import os
-import sys
-
-import torch
-import torch.nn.functional as F
-import triton
-from tqdm import tqdm
-
-from vllm.model_executor.layers.fused_moe import (fused_moe,
-                                                  get_config_file_name)
-
-os.environ['CUDA_VISIBLE_DEVICES'] = '0'
-
-
-def main(dtype: str):
-    method = fused_moe
-    for bs in [
-            1, 2, 4, 8, 16, 24, 32, 48, 64, 96, 128, 256, 512, 1024, 1536,
-            2048, 3072, 4096
-    ]:
-        run_grid(bs, method=method, dtype=dtype)
-
-
-def run_grid(bs, method, dtype: str):
-    d_model = 4096
-    num_total_experts = 8
-    top_k = 2
-    tp_size = 2
-    model_intermediate_size = 14336
-    num_layers = 32
-    num_calls = 100
-
-    num_warmup_trials = 1
-    num_trials = 1
-
-    configs = []
-
-    for block_size_n in [32, 64, 128, 256]:
-        for block_size_m in [16, 32, 64, 128, 256]:
-            for block_size_k in [64, 128, 256]:
-                for group_size_m in [1, 16, 32, 64]:
-                    for num_warps in [4, 8]:
-                        for num_stages in [2, 3, 4, 5]:
-                            configs.append({
-                                "BLOCK_SIZE_M": block_size_m,
-                                "BLOCK_SIZE_N": block_size_n,
-                                "BLOCK_SIZE_K": block_size_k,
-                                "GROUP_SIZE_M": group_size_m,
-                                "num_warps": num_warps,
-                                "num_stages": num_stages,
-                            })
-
-    best_config = None
-    best_time_us = 1e20
-
-    print(f'{tp_size=} {bs=}')
-
-    for config in tqdm(configs):
-        # warmup
-        try:
-            for _ in range(num_warmup_trials):
-                run_timing(
-                    num_calls=num_calls,
-                    bs=bs,
-                    d_model=d_model,
-                    num_total_experts=num_total_experts,
-                    top_k=top_k,
-                    tp_size=tp_size,
-                    model_intermediate_size=model_intermediate_size,
-                    method=method,
-                    config=config,
-                    dtype=dtype,
-                )
-        except triton.runtime.autotuner.OutOfResources:
-            continue
-
-        # trial
-        for _ in range(num_trials):
-            kernel_dur_ms = run_timing(
-                num_calls=num_calls,
-                bs=bs,
-                d_model=d_model,
-                num_total_experts=num_total_experts,
-                top_k=top_k,
-                tp_size=tp_size,
-                model_intermediate_size=model_intermediate_size,
-                method=method,
-                config=config,
-                dtype=dtype,
-            )
-
-            kernel_dur_us = 1000 * kernel_dur_ms
-            model_dur_ms = kernel_dur_ms * num_layers
-
-            if kernel_dur_us < best_time_us:
-                best_config = config
-                best_time_us = kernel_dur_us
-
-                tqdm.write(
-                    f'{kernel_dur_us=:.1f} {model_dur_ms=:.1f}'
-                    f' {bs=} {tp_size=} {top_k=} {num_total_experts=} '
-                    f'{d_model=} {model_intermediate_size=} {num_layers=}')
-
-    print("best_time_us", best_time_us)
-    print("best_config", best_config)
-
-    # holds Dict[str, Dict[str, int]]
-    filename = get_config_file_name(num_total_experts,
-                                    model_intermediate_size // tp_size,
-                                    "float8" if dtype == "float8" else None)
-    print(f"writing config to file {filename}")
-    existing_content = {}
-    if os.path.exists(filename):
-        with open(filename, "r") as f:
-            existing_content = json.load(f)
-    existing_content[str(bs)] = best_config
-    with open(filename, "w") as f:
-        json.dump(existing_content, f, indent=4)
-        f.write("\n")
-
-
-def run_timing(num_calls: int, bs: int, d_model: int, num_total_experts: int,
-               top_k: int, tp_size: int, model_intermediate_size: int, method,
-               config, dtype: str) -> float:
-    shard_intermediate_size = model_intermediate_size // tp_size
-
-    hidden_states = torch.rand(
-        (bs, d_model),
-        device="cuda:0",
-        dtype=torch.float16,
-    )
-
-    w1 = torch.rand(
-        (num_total_experts, 2 * shard_intermediate_size, d_model),
-        device=hidden_states.device,
-        dtype=hidden_states.dtype,
-    )
-
-    w2 = torch.rand(
-        (num_total_experts, d_model, shard_intermediate_size),
-        device=hidden_states.device,
-        dtype=hidden_states.dtype,
-    )
-
-    w1_scale = None
-    w2_scale = None
-    a1_scale = None
-    a2_scale = None
-
-    if dtype == "float8":
-        w1 = w1.to(torch.float8_e4m3fn)
-        w2 = w2.to(torch.float8_e4m3fn)
-        w1_scale = torch.ones(num_total_experts,
-                              device=hidden_states.device,
-                              dtype=torch.float32)
-        w2_scale = torch.ones(num_total_experts,
-                              device=hidden_states.device,
-                              dtype=torch.float32)
-        a1_scale = torch.ones(1,
-                              device=hidden_states.device,
-                              dtype=torch.float32)
-        a2_scale = torch.ones(1,
-                              device=hidden_states.device,
-                              dtype=torch.float32)
-
-    gating_output = F.softmax(torch.rand(
-        (num_calls, bs, num_total_experts),
-        device=hidden_states.device,
-        dtype=torch.float32,
-    ),
-                              dim=-1)
-
-    start_event = torch.cuda.Event(enable_timing=True)
-    end_event = torch.cuda.Event(enable_timing=True)
-
-    start_event.record()
-    for i in range(num_calls):
-        hidden_states = method(
-            hidden_states=hidden_states,
-            w1=w1,
-            w2=w2,
-            w1_scale=w1_scale,
-            w2_scale=w2_scale,
-            a1_scale=a1_scale,
-            a2_scale=a2_scale,
-            gating_output=gating_output[i],
-            topk=2,
-            renormalize=True,
-            inplace=True,
-            override_config=config,
-            use_fp8=dtype == "float8",
-        )
-    end_event.record()
-    end_event.synchronize()
-
-    dur_ms = start_event.elapsed_time(end_event) / num_calls
-    return dur_ms
-
-
-if __name__ == "__main__":
-    parser = argparse.ArgumentParser(
-        prog='benchmark_mixtral_moe',
-        description='Benchmark and tune the fused_moe kernel',
-    )
-    parser.add_argument(
-        '--dtype',
-        type=str,
-        default='auto',
-        choices=['float8', 'float16'],
-        help='Data type used for fused_moe kernel computations',
-    )
-    args = parser.parse_args()
-    sys.exit(main(args.dtype))

benchmarks/kernels/benchmark_moe.py

+import argparse
+import time
+from datetime import datetime
+from typing import Any, Dict, List, Tuple
+
+import ray
+import torch
+import triton
+from ray.experimental.tqdm_ray import tqdm
+from transformers import AutoConfig
+
+from vllm.model_executor.layers.fused_moe.fused_moe import *
+
+
+def benchmark_config(
+    config: Dict[str, int],
+    num_tokens: int,
+    num_experts: int,
+    shard_intermediate_size: int,
+    hidden_size: int,
+    topk: int,
+    dtype: torch.dtype,
+    use_fp8: bool,
+    num_iters: int = 100,
+) -> float:
+    init_dtype = torch.float16 if use_fp8 else dtype
+    x = torch.randn(num_tokens, hidden_size, dtype=dtype)
+    w1 = torch.randn(num_experts,
+                     shard_intermediate_size,
+                     hidden_size,
+                     dtype=init_dtype)
+    w2 = torch.randn(num_experts,
+                     hidden_size,
+                     shard_intermediate_size // 2,
+                     dtype=init_dtype)
+    gating_output = torch.randn(num_iters,
+                                num_tokens,
+                                num_experts,
+                                dtype=torch.float32)
+
+    w1_scale = None
+    w2_scale = None
+    a1_scale = None
+    a2_scale = None
+    if use_fp8:
+        w1_scale = torch.randn(num_experts, dtype=torch.float32)
+        w2_scale = torch.randn(num_experts, dtype=torch.float32)
+        a1_scale = torch.randn(1, dtype=torch.float32)
+        a2_scale = torch.randn(1, dtype=torch.float32)
+
+        w1 = w1.to(torch.float8_e4m3fn)
+        w2 = w2.to(torch.float8_e4m3fn)
+
+    input_gating = torch.empty(num_tokens, num_experts, dtype=torch.float32)
+
+    def prepare(i: int):
+        input_gating.copy_(gating_output[i])
+
+    def run():
+        fused_moe(
+            x,
+            w1,
+            w2,
+            input_gating,
+            topk,
+            renormalize=True,
+            inplace=True,
+            override_config=config,
+            use_fp8=use_fp8,
+            w1_scale=w1_scale,
+            w2_scale=w2_scale,
+            a1_scale=a1_scale,
+            a2_scale=a2_scale,
+        )
+
+    # JIT compilation & warmup
+    run()
+    torch.cuda.synchronize()
+
+    # Capture 10 invocations with CUDA graph
+    graph = torch.cuda.CUDAGraph()
+    with torch.cuda.graph(graph):
+        for _ in range(10):
+            run()
+    torch.cuda.synchronize()
+
+    # Warmup
+    for _ in range(5):
+        graph.replay()
+    torch.cuda.synchronize()
+
+    start_event = torch.cuda.Event(enable_timing=True)
+    end_event = torch.cuda.Event(enable_timing=True)
+
+    latencies = []
+    for i in range(num_iters):
+        prepare(i)
+        torch.cuda.synchronize()
+
+        start_event.record()
+        graph.replay()
+        end_event.record()
+        end_event.synchronize()
+        latencies.append(start_event.elapsed_time(end_event))
+    avg = sum(latencies) / (num_iters * 10) * 1000  # us
+    graph.reset()
+    return avg
+
+
+def get_configs_compute_bound() -> List[Dict[str, int]]:
+    # Reduced search space for faster tuning.
+    # TODO(woosuk): Increase the search space and use a performance model to
+    # prune the search space.
+    configs = []
+    for num_stages in [2, 3, 4, 5]:
+        for block_m in [16, 32, 64, 128, 256]:
+            for block_k in [64, 128, 256]:
+                for block_n in [32, 64, 128, 256]:
+                    for num_warps in [4, 8]:
+                        for group_size in [1, 16, 32, 64]:
+                            configs.append({
+                                "BLOCK_SIZE_M": block_m,
+                                "BLOCK_SIZE_N": block_n,
+                                "BLOCK_SIZE_K": block_k,
+                                "GROUP_SIZE_M": group_size,
+                                "num_warps": num_warps,
+                                "num_stages": num_stages,
+                            })
+    return configs
+
+
+@ray.remote(num_gpus=1)
+class BenchmarkWorker:
+
+    def __init__(self, seed: int) -> None:
+        torch.set_default_device("cuda")
+        torch.cuda.manual_seed_all(seed)
+        self.seed = seed
+
+    def benchmark(
+        self,
+        num_tokens: int,
+        num_experts: int,
+        shard_intermediate_size: int,
+        hidden_size: int,
+        topk: int,
+        dtype: torch.dtype,
+        use_fp8: bool,
+    ) -> Tuple[Dict[str, int], float]:
+        torch.cuda.manual_seed_all(self.seed)
+
+        dtype_str = "float8" if use_fp8 else None
+        # NOTE(woosuk): The current naming convention uses w2.shape[2], which
+        # is the intermediate size after silu_and_mul.
+        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)
+        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)
+        return config, kernel_time
+
+    def tune(
+        self,
+        num_tokens: int,
+        num_experts: int,
+        shard_intermediate_size: int,
+        hidden_size: int,
+        topk: int,
+        dtype: torch.dtype,
+        use_fp8: bool,
+        search_space: List[Dict[str, int]],
+    ) -> Dict[str, int]:
+        best_config = None
+        best_time = float("inf")
+        for config in tqdm(search_space):
+            try:
+                kernel_time = benchmark_config(config,
+                                               num_tokens,
+                                               num_experts,
+                                               shard_intermediate_size,
+                                               hidden_size,
+                                               topk,
+                                               dtype,
+                                               use_fp8,
+                                               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={num_tokens}")
+        return best_config
+
+
+def sort_config(config: Dict[str, int]) -> Dict[str, int]:
+    return {
+        "BLOCK_SIZE_M": config["BLOCK_SIZE_M"],
+        "BLOCK_SIZE_N": config["BLOCK_SIZE_N"],
+        "BLOCK_SIZE_K": config["BLOCK_SIZE_K"],
+        "GROUP_SIZE_M": config["GROUP_SIZE_M"],
+        "num_warps": config["num_warps"],
+        "num_stages": config["num_stages"],
+    }
+
+
+def save_configs(
+    configs: Dict[int, Dict[str, int]],
+    num_experts: int,
+    shard_intermediate_size: int,
+    hidden_size: int,
+    topk: int,
+    dtype: torch.dtype,
+    use_fp8: bool,
+) -> None:
+    dtype_str = "float8" if use_fp8 else None
+    # 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)
+    print(f"Writing best config to {filename}...")
+    with open(filename, "w") as f:
+        json.dump(configs, f, indent=4)
+        f.write("\n")
+
+
+def main(args: argparse.Namespace):
+    print(args)
+
+    config = AutoConfig.from_pretrained(args.model)
+    if config.architectures[0] == "DbrxForCausalLM":
+        E = config.ffn_config.moe_num_experts
+        topk = config.ffn_config.moe_top_k
+        intermediate_size = config.ffn_config.ffn_hidden_size
+        shard_intermediate_size = 2 * intermediate_size // args.tp_size
+    else:
+        # Default: Mixtral.
+        E = config.num_local_experts
+        topk = config.num_experts_per_tok
+        intermediate_size = config.intermediate_size
+        shard_intermediate_size = 2 * intermediate_size // args.tp_size
+
+    hidden_size = config.hidden_size
+    dtype = config.torch_dtype
+    use_fp8 = args.dtype == "fp8"
+
+    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]
+
+    ray.init()
+    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]:
+        outputs = []
+        worker_idx = 0
+        for input_args in inputs:
+            worker = workers[worker_idx]
+            worker_method = getattr(worker, method)
+            output = worker_method.remote(*input_args)
+            outputs.append(output)
+            worker_idx = (worker_idx + 1) % num_gpus
+        return ray.get(outputs)
+
+    if args.tune:
+        search_space = get_configs_compute_bound()
+        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, search_space)
+                     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)
+        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)
+                               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}")
+            print(f"Kernel time: {kernel_time:.2f} us")
+
+
+if __name__ == "__main__":
+    parser = argparse.ArgumentParser()
+    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("--dtype",
+                        type=str,
+                        choices=["auto", "fp8"],
+                        default="auto")
+    parser.add_argument("--seed", type=int, default=0)
+    parser.add_argument("--batch-size", type=int, required=False)
+    parser.add_argument("--tune", action="store_true")
+    args = parser.parse_args()
+
+    main(args)