merge v0.4.1

.buildkite/run-cpu-test.sh

+# This script build the CPU docker image and run the offline inference inside the container.
+# It serves a sanity check for compilation and basic model usage.
+set -ex
+
+# Try building the docker image
+docker build -t cpu-test -f Dockerfile.cpu .
+
+# Setup cleanup
+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 examples/offline_inference.py

.buildkite/run-neuron-test.sh

+# This script build the Neuron docker image and run the API server inside the container.
+# It serves a sanity check for compilation and basic model usage.
+set -e
+
+# Try building the docker image
+aws ecr get-login-password --region us-west-2 | docker login --username AWS --password-stdin 763104351884.dkr.ecr.us-west-2.amazonaws.com
+docker build -t neuron -f Dockerfile.neuron .
+
+# Setup cleanup
+remove_docker_container() { docker rm -f neuron || true; }
+trap remove_docker_container EXIT
+remove_docker_container
+
+# Run the image
+docker run --device=/dev/neuron0 --device=/dev/neuron1 --network host --name neuron neuron python3 -m vllm.entrypoints.api_server \
+       --model TinyLlama/TinyLlama-1.1B-Chat-v1.0 --max-num-seqs 8 --max-model-len 128 --block-size 128 --device neuron --tensor-parallel-size 2 &
+
+# Wait for the server to start
+wait_for_server_to_start() {
+    timeout=300
+    counter=0
+
+    while [ "$(curl -s -o /dev/null -w ''%{http_code}'' localhost:8000/health)" != "200" ]; do
+        sleep 1
+        counter=$((counter + 1))
+        if [ $counter -ge $timeout ]; then
+            echo "Timeout after $timeout seconds"
+            break
+        fi
+    done
+}
+wait_for_server_to_start
+
+# Test a simple prompt
+curl -X POST -H "Content-Type: application/json" \
+    localhost:8000/generate \
+    -d '{"prompt": "San Francisco is a"}'

.buildkite/test-pipeline.yaml

@@ -12,7 +12,11 @@ steps:
   command: pytest -v -s async_engine
 
 - label: Basic Correctness Test
-  command: pytest -v -s basic_correctness
+  commands:
+  - VLLM_ATTENTION_BACKEND=XFORMERS pytest -v -s basic_correctness/test_basic_correctness.py
+  - VLLM_ATTENTION_BACKEND=FLASH_ATTN pytest -v -s basic_correctness/test_basic_correctness.py
+  - VLLM_ATTENTION_BACKEND=XFORMERS pytest -v -s basic_correctness/test_chunked_prefill.py
+  - VLLM_ATTENTION_BACKEND=FLASH_ATTN pytest -v -s basic_correctness/test_chunked_prefill.py
 
 - label: Core Test
   command: pytest -v -s core
@@ -27,14 +31,20 @@ steps:
   num_gpus: 2 # only support 1 or 2 for now.
   commands:
   - pytest -v -s test_pynccl.py
+  - pytest -v -s test_pynccl_library.py
   - TEST_DIST_MODEL=facebook/opt-125m pytest -v -s test_basic_distributed_correctness.py
   - TEST_DIST_MODEL=meta-llama/Llama-2-7b-hf pytest -v -s test_basic_distributed_correctness.py
+  - TEST_DIST_MODEL=facebook/opt-125m pytest -v -s test_chunked_prefill_distributed.py
+  - TEST_DIST_MODEL=meta-llama/Llama-2-7b-hf pytest -v -s test_chunked_prefill_distributed.py
 
 - label: Engine Test
-  command: pytest -v -s engine tokenization test_sequence.py test_config.py
+  command: pytest -v -s engine tokenization test_sequence.py test_config.py test_logger.py
 
 - label: Entrypoints Test
-  command: pytest -v -s entrypoints
+  commands:
+  # these tests have to be separated, because each one will allocate all posible GPU memory
+  - pytest -v -s entrypoints --ignore=entrypoints/test_server_oot_registration.py
+  - pytest -v -s entrypoints/test_server_oot_registration.py
 
 - label: Examples Test
   working_dir: "/vllm-workspace/examples"
@@ -80,9 +90,15 @@ steps:
   command: pytest -v -s lora --shard-id=$$BUILDKITE_PARALLEL_JOB --num-shards=$$BUILDKITE_PARALLEL_JOB_COUNT
   parallelism: 4
 
+- label: Tensorizer Test
+  command: apt-get install curl libsodium23 && pytest -v -s tensorizer_loader
+
 - label: Metrics Test
   command: pytest -v -s metrics
 
+- label: Quantization Test
+  command: pytest -v -s quantization
+
 - label: Benchmarks
   working_dir: "/vllm-workspace/.buildkite"
   commands:
@@ -90,7 +106,7 @@ steps:
   - bash run-benchmarks.sh
 
 - label: Documentation Build
-  working_dir: "/vllm-workspace/docs"
+  working_dir: "/vllm-workspace/test_docs/docs"
   no_gpu: True
   commands:
   - pip install -r requirements-docs.txt

.buildkite/test-template.j2

@@ -3,10 +3,6 @@
 {% set default_working_dir = "/vllm-workspace/tests" %}
 
 steps:
-  - label: "AMD Test"
-    agents:
-      queue: amd
-    command: bash .buildkite/run-amd-test.sh
 
   - label: ":docker: build image"
     commands:
@@ -20,6 +16,19 @@ steps:
           limit: 5
   - wait
 
+  - label: "AMD Test"
+    agents:
+      queue: amd
+    command: bash .buildkite/run-amd-test.sh
+
+  - label: "Neuron Test"
+    agents:
+      queue: neuron
+    command: bash .buildkite/run-neuron-test.sh
+
+  - label: "CPU Test"
+    command: bash .buildkite/run-cpu-test.sh
+
   {% for step in steps %}
   - label: "{{ step.label }}"
     agents:

.github/ISSUE_TEMPLATE/200-installation.yml

@@ -18,6 +18,7 @@ body:
       # For security purposes, please feel free to check the contents of collect_env.py before running it.
       python collect_env.py
       ```
+      It is suggested to download and execute the latest script, as vllm might frequently update the diagnosis information needed for accurately and quickly responding to issues.
     value: |
       ```text
       The output of `python collect_env.py`

.github/ISSUE_TEMPLATE/300-usage.yml

@@ -18,6 +18,7 @@ body:
       # For security purposes, please feel free to check the contents of collect_env.py before running it.
       python collect_env.py
       ```
+      It is suggested to download and execute the latest script, as vllm might frequently update the diagnosis information needed for accurately and quickly responding to issues.
     value: |
       ```text
       The output of `python collect_env.py`

.github/ISSUE_TEMPLATE/400-bug report.yml

@@ -18,6 +18,7 @@ body:
       # For security purposes, please feel free to check the contents of collect_env.py before running it.
       python collect_env.py
       ```
+      It is suggested to download and execute the latest script, as vllm might frequently update the diagnosis information needed for accurately and quickly responding to issues.
     value: |
       ```text
       The output of `python collect_env.py`
@@ -57,6 +58,8 @@ body:
       If the code is too long (hopefully, it isn't), feel free to put it in a public gist and link it in the issue: https://gist.github.com.
 
       Please also paste or describe the results you observe instead of the expected results. If you observe an error, please paste the error message including the **full** traceback of the exception. It may be relevant to wrap error messages in ```` ```triple quotes blocks``` ````.
+
+      If you experienced crashes or hangs, it would be helpful to run vllm with `export VLLM_TRACE_FUNCTION=1` . All the function calls in vllm will be recorded. Inspect these log files, and tell which function crashes or hangs.
     placeholder: |
       A clear and concise description of what the bug is.
 

.github/ISSUE_TEMPLATE/700-performance discussion.yml

@@ -39,6 +39,7 @@ body:
       # For security purposes, please feel free to check the contents of collect_env.py before running it.
       python collect_env.py
       ```
+      It is suggested to download and execute the latest script, as vllm might frequently update the diagnosis information needed for accurately and quickly responding to issues.
     value: |
       ```text
       The output of `python collect_env.py`

.github/workflows/mypy.yaml

+name: mypy
+
+on:
+  # Trigger the workflow on push or pull request,
+  # but only for the main branch
+  push:
+    branches:
+      - main
+  pull_request:
+    branches:
+      - main
+
+jobs:
+  ruff:
+    runs-on: ubuntu-latest
+    strategy:
+      matrix:
+        python-version: ["3.8", "3.9", "3.10", "3.11"]
+    steps:
+    - uses: actions/checkout@v2
+    - name: Set up Python ${{ matrix.python-version }}
+      uses: actions/setup-python@v2
+      with:
+        python-version: ${{ matrix.python-version }}
+    - name: Install dependencies
+      run: |
+        python -m pip install --upgrade pip
+        pip install mypy==1.9.0
+        pip install types-setuptools
+        pip install types-PyYAML
+        pip install types-requests
+        pip install types-setuptools
+    - name: Mypy
+      run: |
+        mypy vllm/attention --config-file pyproject.toml
+        # TODO(sang): Fix nested dir
+        mypy vllm/core/*.py --follow-imports=skip --config-file pyproject.toml
+        mypy vllm/distributed --config-file pyproject.toml
+        mypy vllm/entrypoints --config-file pyproject.toml
+        mypy vllm/executor --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
+        mypy vllm/engine  --config-file pyproject.toml
+        mypy vllm/worker --config-file pyproject.toml
+        mypy vllm/spec_decode --config-file pyproject.toml
+        # TODO(sang): Fix nested dir
+        mypy vllm/model_executor/*.py  --config-file pyproject.toml
+        # TODO(sang): Fix nested dir
+        # mypy vllm/lora/*.py --config-file pyproject.toml
+

.github/workflows/publish.yml

@@ -49,13 +49,16 @@ jobs:
       matrix:
           os: ['ubuntu-20.04']
           python-version: ['3.8', '3.9', '3.10', '3.11']
-          pytorch-version: ['2.1.2']  # Must be the most recent version that meets requirements.txt.
+          pytorch-version: ['2.2.1']  # Must be the most recent version that meets requirements-cuda.txt.
           cuda-version: ['11.8', '12.1']
 
     steps:
       - name: Checkout
         uses: actions/checkout@v3
 
+      - name: Setup ccache
+        uses: hendrikmuhs/ccache-action@v1.2
+
       - name: Set up Linux Env
         if: ${{ runner.os == 'Linux' }}
         run: |

.github/workflows/ruff.yml

@@ -15,7 +15,7 @@ jobs:
     runs-on: ubuntu-latest
     strategy:
       matrix:
-        python-version: ["3.10"]
+        python-version: ["3.8", "3.9", "3.10", "3.11"]
     steps:
     - uses: actions/checkout@v2
     - name: Set up Python ${{ matrix.python-version }}

.github/workflows/scripts/build.sh

@@ -9,12 +9,13 @@ LD_LIBRARY_PATH=${cuda_home}/lib64:$LD_LIBRARY_PATH
 
 # Install requirements
 $python_executable -m pip install wheel packaging
-$python_executable -m pip install -r requirements.txt
+$python_executable -m pip install -r requirements-cuda.txt
 
 # Limit the number of parallel jobs to avoid OOM
 export MAX_JOBS=1
 # Make sure punica is built for the release (for LoRA)
 export VLLM_INSTALL_PUNICA_KERNELS=1
-
+# Make sure release wheels are built for the following architectures
+export TORCH_CUDA_ARCH_LIST="7.0 7.5 8.0 8.6 8.9 9.0+PTX"
 # Build
 $python_executable setup.py bdist_wheel --dist-dir=dist

.github/workflows/yapf.yml

@@ -14,7 +14,7 @@ jobs:
     runs-on: ubuntu-latest
     strategy:
       matrix:
-        python-version: ["3.10"]
+        python-version: ["3.8", "3.9", "3.10", "3.11"]
     steps:
     - uses: actions/checkout@v2
     - name: Set up Python ${{ matrix.python-version }}

.gitignore

@@ -70,6 +70,8 @@ instance/
 
 # Sphinx documentation
 docs/_build/
+docs/source/getting_started/examples/*.rst
+!**/*.template.rst
 
 # PyBuilder
 .pybuilder/
@@ -181,6 +183,7 @@ _build/
 # hip files generated by PyTorch
 *.hip
 *_hip*
+hip_compat.h
 
 # Benchmark dataset
 *.json

CMakeLists.txt

@@ -2,7 +2,10 @@ cmake_minimum_required(VERSION 3.21)
 
 project(vllm_extensions LANGUAGES CXX)
 
+option(VLLM_TARGET_DEVICE "Target device backend for vLLM" "cuda")
+
 message(STATUS "Build type: ${CMAKE_BUILD_TYPE}")
+message(STATUS "Target device: ${VLLM_TARGET_DEVICE}")
 
 include(${CMAKE_CURRENT_LIST_DIR}/cmake/utils.cmake)
 
@@ -16,7 +19,7 @@ set(PYTHON_SUPPORTED_VERSIONS "3.8" "3.9" "3.10" "3.11")
 set(CUDA_SUPPORTED_ARCHS "7.0;7.5;8.0;8.6;8.9;9.0")
 
 # Supported AMD GPU architectures.
-set(HIP_SUPPORTED_ARCHS "gfx908;gfx90a;gfx942;gfx1100;gfx906;gfx926;gfx928")
+set(HIP_SUPPORTED_ARCHS "gfx906;gfx908;gfx90a;gfx940;gfx941;gfx942;gfx1030;gfx1100;gfx926;gfx928")
 
 #
 # Supported/expected torch versions for CUDA/ROCm.
@@ -28,7 +31,7 @@ set(HIP_SUPPORTED_ARCHS "gfx908;gfx90a;gfx942;gfx1100;gfx906;gfx926;gfx928")
 # requirements.txt files and should be kept consistent.  The ROCm torch
 # versions are derived from Dockerfile.rocm
 #
-set(TORCH_SUPPORTED_VERSION_CUDA "2.1.2")
+set(TORCH_SUPPORTED_VERSION_CUDA "2.2.1")
 set(TORCH_SUPPORTED_VERSION_ROCM_5X "2.0.1")
 set(TORCH_SUPPORTED_VERSION_ROCM_6X "2.1.1")
 
@@ -76,6 +79,19 @@ find_package(Torch REQUIRED)
 find_library(torch_python_LIBRARY torch_python PATHS
   "${TORCH_INSTALL_PREFIX}/lib")
 
+#
+# Forward the non-CUDA device extensions to external CMake scripts.
+#
+if (NOT VLLM_TARGET_DEVICE STREQUAL "cuda" AND
+    NOT VLLM_TARGET_DEVICE STREQUAL "rocm")
+    if (VLLM_TARGET_DEVICE STREQUAL "cpu")
+        include(${CMAKE_CURRENT_LIST_DIR}/cmake/cpu_extension.cmake)
+    else()
+        message(FATAL_ERROR "Unsupported vLLM target device: ${VLLM_TARGET_DEVICE}")
+    endif()
+    return()
+endif()
+
 #
 # Set up GPU language and check the torch version and warn if it isn't
 # what is expected.
@@ -151,12 +167,14 @@ set(VLLM_EXT_SRC
   "csrc/layernorm_kernels.cu"
   "csrc/quantization/squeezellm/quant_cuda_kernel.cu"
   "csrc/quantization/gptq/q_gemm.cu"
+  "csrc/quantization/fp8/fp8_cuda_kernels.cu"
   "csrc/cuda_utils_kernels.cu"
   "csrc/moe_align_block_size_kernels.cu"
   "csrc/pybind.cpp")
 
 if(VLLM_GPU_LANG STREQUAL "CUDA")
   list(APPEND VLLM_EXT_SRC
+    "csrc/quantization/aqlm/gemm_kernels.cu"
     "csrc/quantization/awq/gemm_kernels.cu"
     "csrc/quantization/marlin/marlin_cuda_kernel.cu"
     "csrc/custom_all_reduce.cu")
@@ -194,23 +212,11 @@ define_gpu_extension_target(
 
 set(VLLM_PUNICA_EXT_SRC
   "csrc/punica/bgmv/bgmv_bf16_bf16_bf16.cu"
-  "csrc/punica/bgmv/bgmv_bf16_bf16_fp16.cu"
-  "csrc/punica/bgmv/bgmv_bf16_fp16_bf16.cu"
-  "csrc/punica/bgmv/bgmv_bf16_fp16_fp16.cu"
   "csrc/punica/bgmv/bgmv_bf16_fp32_bf16.cu"
-  "csrc/punica/bgmv/bgmv_bf16_fp32_fp16.cu"
-  "csrc/punica/bgmv/bgmv_fp16_bf16_bf16.cu"
-  "csrc/punica/bgmv/bgmv_fp16_bf16_fp16.cu"
-  "csrc/punica/bgmv/bgmv_fp16_fp16_bf16.cu"
   "csrc/punica/bgmv/bgmv_fp16_fp16_fp16.cu"
-  "csrc/punica/bgmv/bgmv_fp16_fp32_bf16.cu"
   "csrc/punica/bgmv/bgmv_fp16_fp32_fp16.cu"
   "csrc/punica/bgmv/bgmv_fp32_bf16_bf16.cu"
-  "csrc/punica/bgmv/bgmv_fp32_bf16_fp16.cu"
-  "csrc/punica/bgmv/bgmv_fp32_fp16_bf16.cu"
   "csrc/punica/bgmv/bgmv_fp32_fp16_fp16.cu"
-  "csrc/punica/bgmv/bgmv_fp32_fp32_bf16.cu"
-  "csrc/punica/bgmv/bgmv_fp32_fp32_fp16.cu"
   "csrc/punica/punica_ops.cc")
 
 #

CONTRIBUTING.md

@@ -21,7 +21,6 @@ Express your support on Twitter if vLLM aids you, or simply offer your appreciat
 ### Build from source
 
 ```bash
-pip install -r requirements.txt
 pip install -e .  # This may take several minutes.
 ```
 
@@ -30,6 +29,8 @@ pip install -e .  # This may take several minutes.
 ```bash
 pip install -r requirements-dev.txt
 
+# linting and formatting
+bash format.sh
 # Static type checking
 mypy
 # Unit tests

Dockerfile

@@ -2,6 +2,7 @@
 # to run the OpenAI compatible server.
 
 #################### BASE BUILD IMAGE ####################
+# prepare basic build environment
 FROM nvidia/cuda:12.1.0-devel-ubuntu22.04 AS dev
 
 RUN apt-get update -y \
@@ -16,18 +17,26 @@ RUN ldconfig /usr/local/cuda-12.1/compat/
 WORKDIR /workspace
 
 # install build and runtime dependencies
-COPY requirements.txt requirements.txt
+COPY requirements-common.txt requirements-common.txt
+COPY requirements-cuda.txt requirements-cuda.txt
 RUN --mount=type=cache,target=/root/.cache/pip \
-    pip install -r requirements.txt
+    pip install -r requirements-cuda.txt
 
 # install development dependencies
 COPY requirements-dev.txt requirements-dev.txt
 RUN --mount=type=cache,target=/root/.cache/pip \
     pip install -r requirements-dev.txt
+
+# cuda arch list used by torch
+# can be useful for both `dev` and `test`
+# explicitly set the list to avoid issues with torch 2.2
+# see https://github.com/pytorch/pytorch/pull/123243
+ARG torch_cuda_arch_list='7.0 7.5 8.0 8.6 8.9 9.0+PTX'
+ENV TORCH_CUDA_ARCH_LIST=${torch_cuda_arch_list}
 #################### BASE BUILD IMAGE ####################
 
 
-#################### EXTENSION BUILD IMAGE ####################
+#################### WHEEL BUILD IMAGE ####################
 FROM dev AS build
 
 # install build dependencies
@@ -38,18 +47,16 @@ RUN --mount=type=cache,target=/root/.cache/pip \
 # install compiler cache to speed up compilation leveraging local or remote caching
 RUN apt-get update -y && apt-get install -y ccache
 
-# copy input files
+# files and directories related to build wheels
 COPY csrc csrc
 COPY setup.py setup.py
 COPY cmake cmake
 COPY CMakeLists.txt CMakeLists.txt
-COPY requirements.txt requirements.txt
+COPY requirements-common.txt requirements-common.txt
+COPY requirements-cuda.txt requirements-cuda.txt
 COPY pyproject.toml pyproject.toml
-COPY vllm/__init__.py vllm/__init__.py
+COPY vllm vllm
 
-# cuda arch list used by torch
-ARG torch_cuda_arch_list='7.0 7.5 8.0 8.6 8.9 9.0+PTX'
-ENV TORCH_CUDA_ARCH_LIST=${torch_cuda_arch_list}
 # max jobs used by Ninja to build extensions
 ARG max_jobs=2
 ENV MAX_JOBS=${max_jobs}
@@ -61,7 +68,15 @@ ENV VLLM_INSTALL_PUNICA_KERNELS=1
 
 ENV CCACHE_DIR=/root/.cache/ccache
 RUN --mount=type=cache,target=/root/.cache/ccache \
-    python3 setup.py build_ext --inplace
+    --mount=type=cache,target=/root/.cache/pip \
+    python3 setup.py bdist_wheel --dist-dir=dist
+
+# the `vllm_nccl` package must be installed from source distribution
+# pip is too smart to store a wheel in the cache, and other CI jobs
+# will directly use the wheel from the cache, which is not what we want.
+# we need to remove it manually
+RUN --mount=type=cache,target=/root/.cache/pip \
+    pip cache remove vllm_nccl*
 #################### EXTENSION Build IMAGE ####################
 
 #################### FLASH_ATTENTION Build IMAGE ####################
@@ -81,57 +96,59 @@ RUN pip --verbose wheel flash-attn==${FLASH_ATTN_VERSION} \
 
 #################### FLASH_ATTENTION Build IMAGE ####################
 
-#################### TEST IMAGE ####################
-# image to run unit testing suite
-FROM dev AS test
-
-# copy pytorch extensions separately to avoid having to rebuild
-# when python code changes
+#################### vLLM installation IMAGE ####################
+# image with vLLM installed
+FROM nvidia/cuda:12.1.0-base-ubuntu22.04 AS vllm-base
 WORKDIR /vllm-workspace
-# ADD is used to preserve directory structure
-ADD . /vllm-workspace/
-COPY --from=build /workspace/vllm/*.so /vllm-workspace/vllm/
-# Install flash attention (from pre-built wheel)
+
+RUN apt-get update -y \
+    && apt-get install -y python3-pip git vim
+
+# Workaround for https://github.com/openai/triton/issues/2507 and
+# https://github.com/pytorch/pytorch/issues/107960 -- hopefully
+# this won't be needed for future versions of this docker image
+# or future versions of triton.
+RUN ldconfig /usr/local/cuda-12.1/compat/
+
+# install vllm wheel first, so that torch etc will be installed
+RUN --mount=type=bind,from=build,src=/workspace/dist,target=/vllm-workspace/dist \
+    --mount=type=cache,target=/root/.cache/pip \
+    pip install dist/*.whl --verbose
+
 RUN --mount=type=bind,from=flash-attn-builder,src=/usr/src/flash-attention-v2,target=/usr/src/flash-attention-v2 \
+    --mount=type=cache,target=/root/.cache/pip \
     pip install /usr/src/flash-attention-v2/*.whl --no-cache-dir
-# ignore build dependencies installation because we are using pre-complied extensions
-RUN rm pyproject.toml
-RUN --mount=type=cache,target=/root/.cache/pip VLLM_USE_PRECOMPILED=1 pip install . --verbose
-#################### TEST IMAGE ####################
+#################### vLLM installation IMAGE ####################
 
 
-#################### RUNTIME BASE IMAGE ####################
-# We used base cuda image because pytorch installs its own cuda libraries.
-# However pynccl depends on cuda libraries so we had to switch to the runtime image
-# In the future it would be nice to get a container with pytorch and cuda without duplicating cuda
-FROM nvidia/cuda:12.1.0-runtime-ubuntu22.04 AS vllm-base
+#################### TEST IMAGE ####################
+# image to run unit testing suite
+# note that this uses vllm installed by `pip`
+FROM vllm-base AS test
 
-# libnccl required for ray
-RUN apt-get update -y \
-    && apt-get install -y python3-pip
+ADD . /vllm-workspace/
 
-WORKDIR /workspace
-COPY requirements.txt requirements.txt
+# install development dependencies (for testing)
 RUN --mount=type=cache,target=/root/.cache/pip \
-    pip install -r requirements.txt
-
-# Install flash attention (from pre-built wheel)
-RUN --mount=type=bind,from=flash-attn-builder,src=/usr/src/flash-attention-v2,target=/usr/src/flash-attention-v2 \
-    pip install /usr/src/flash-attention-v2/*.whl --no-cache-dir
+    pip install -r requirements-dev.txt
 
-#################### RUNTIME BASE IMAGE ####################
+# doc requires source code
+# we hide them inside `test_docs/` , so that this source code
+# will not be imported by other tests
+RUN mkdir test_docs
+RUN mv docs test_docs/
+RUN mv vllm test_docs/
 
+#################### TEST IMAGE ####################
 
 #################### OPENAI API SERVER ####################
 # openai api server alternative
 FROM vllm-base AS vllm-openai
+
 # install additional dependencies for openai api server
 RUN --mount=type=cache,target=/root/.cache/pip \
     pip install accelerate hf_transfer modelscope
 
-COPY --from=build /workspace/vllm/*.so /workspace/vllm/
-COPY vllm vllm
-
 ENV VLLM_USAGE_SOURCE production-docker-image
 
 ENTRYPOINT ["python3", "-m", "vllm.entrypoints.openai.api_server"]

Dockerfile.cpu

+# This vLLM Dockerfile is used to construct image that can build and run vLLM on x86 CPU platform.
+
+FROM ubuntu:22.04
+
+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
+
+COPY ./ /workspace/vllm
+
+WORKDIR /workspace/vllm
+
+RUN pip install -v -r requirements-cpu.txt --extra-index-url https://download.pytorch.org/whl/cpu
+
+RUN VLLM_TARGET_DEVICE=cpu python3 setup.py install
+
+CMD ["/bin/bash"]

Dockerfile.neuron

+# default base image
+ARG BASE_IMAGE="763104351884.dkr.ecr.us-west-2.amazonaws.com/pytorch-inference-neuronx:2.1.1-neuronx-py310-sdk2.17.0-ubuntu20.04"
+
+FROM $BASE_IMAGE
+
+RUN echo "Base image is $BASE_IMAGE"
+
+# Install some basic utilities
+RUN apt-get update && apt-get install python3 python3-pip -y
+
+### Mount Point ###
+# When launching the container, mount the code directory to /app
+ARG APP_MOUNT=/app
+VOLUME [ ${APP_MOUNT} ]
+WORKDIR ${APP_MOUNT}
+
+RUN python3 -m pip install --upgrade pip
+RUN python3 -m pip install --no-cache-dir fastapi ninja tokenizers pandas
+RUN python3 -m pip install sentencepiece transformers==4.36.2 -U
+RUN python3 -m pip install transformers-neuronx --extra-index-url=https://pip.repos.neuron.amazonaws.com -U
+RUN python3 -m pip install --pre neuronx-cc==2.12.* --extra-index-url=https://pip.repos.neuron.amazonaws.com -U
+
+COPY ./vllm /app/vllm/vllm
+COPY ./setup.py /app/vllm/setup.py
+COPY ./requirements-common.txt /app/vllm/requirements-common.txt
+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
+RUN cd /app/vllm \
+    && pip install -e . \
+    && cd ..
+
+CMD ["/bin/bash"]

Dockerfile.rocm

@@ -14,7 +14,7 @@ RUN echo "Base image is $BASE_IMAGE"
 ARG FA_GFX_ARCHS="gfx90a;gfx942"
 RUN echo "FA_GFX_ARCHS is $FA_GFX_ARCHS"
 
-ARG FA_BRANCH="3d2b6f5"
+ARG FA_BRANCH="ae7928c"
 RUN echo "FA_BRANCH is $FA_BRANCH"
 
 # whether to build flash-attention
@@ -23,6 +23,9 @@ RUN echo "FA_BRANCH is $FA_BRANCH"
 # In that case, we need to use the python reference attention implementation in vllm
 ARG BUILD_FA="1"
 
+# whether to build triton on rocm
+ARG BUILD_TRITON="1"
+
 # Install some basic utilities
 RUN apt-get update && apt-get install python3 python3-pip -y
 
@@ -75,16 +78,24 @@ RUN if [ "$BUILD_FA" = "1" ]; then \
 RUN if [ "$BASE_IMAGE" = "rocm/pytorch:rocm6.0_ubuntu20.04_py3.9_pytorch_2.1.1" ]; then \
     rm -rf /opt/conda/envs/py_3.9/lib/python3.9/site-packages/numpy-1.20.3.dist-info/; fi
 
+# build triton
+RUN if [ "$BUILD_TRITON" = "1" ]; then \
+    mkdir -p libs \
+    && cd libs \
+    && pip uninstall -y triton \
+    && git clone https://github.com/ROCm/triton.git \
+    && cd triton/python \
+    && pip3 install . \
+    && cd ../..; \
+    fi
+
 COPY ./ /app/vllm
 
-RUN python3 -m pip install --upgrade pip
-RUN python3 -m pip install xformers==0.0.23 --no-deps
+RUN python3 -m pip install --upgrade pip numba
 
 RUN cd /app \
     && cd vllm \
     && pip install -U -r requirements-rocm.txt \
-    && if [ "$BUILD_FA" = "1" ]; then \
-       bash patch_xformers.rocm.sh; fi \
     && patch /opt/rocm/include/hip/amd_detail/amd_hip_bf16.h /app/vllm/rocm_patch/rocm_bf16.patch \
     && python3 setup.py install \
     && cd ..