@tduncan / Nvidia-Optimization-Stack-VLLM Mirror · read-only / commits / f758aed0e8
Fork 0

A high-throughput and memory-efficient inference and serving engine for LLMs

Mirrored from https://github.com/tduncan-lum/Nvidia-Optimization-Stack---VLLM.git

Synced Nov 16, 23:15 UTC

[Bugfix][CI/Build][AMD][ROCm]Fixed the cmake build bug which generate garbage on certain devices (#5641)

Hongxia Yang committed Jun 19, 2024 at 02:21 UTC f758aed0e851687e919a4ee09ab872ee2c8fe159
2 files changed +12 -10
Dockerfile.rocm
+8 -9
@@ -7,9 +7,8 @@ ARG BASE_IMAGE="rocm/pytorch:rocm6.0_ubuntu20.04_py3.9_pytorch_2.1.1"
7
8 RUN echo "Base image is $BASE_IMAGE"
9
10 -# BASE_IMAGE for ROCm_5.7: "rocm/pytorch:rocm5.7_ubuntu22.04_py3.10_pytorch_2.0.1"
11 -# BASE_IMAGE for ROCm_6.0: "rocm/pytorch:rocm6.0_ubuntu20.04_py3.9_pytorch_2.1.1"
12 -
10 +ARG ROCm_5_7_BASE="rocm/pytorch:rocm5.7_ubuntu22.04_py3.10_pytorch_2.0.1" \
11 + ROCm_6_0_BASE="rocm/pytorch:rocm6.0_ubuntu20.04_py3.9_pytorch_2.1.1"
12
13 ARG FA_GFX_ARCHS="gfx90a;gfx942"
14 RUN echo "FA_GFX_ARCHS is $FA_GFX_ARCHS"
@@ -68,7 +67,7 @@ RUN if [ "$BUILD_FA" = "1" ]; then \
67 && git checkout ${FA_BRANCH} \
68 && git submodule update --init \
69 && export GPU_ARCHS=${FA_GFX_ARCHS} \
71 - && if [ "$BASE_IMAGE" = "rocm/pytorch:rocm5.7_ubuntu22.04_py3.10_pytorch_2.0.1" ]; then \
70 + && if [ "$BASE_IMAGE" = "$ROCm_5_7_BASE" ]; then \
71 patch /opt/conda/envs/py_3.10/lib/python3.10/site-packages/torch/utils/hipify/hipify_python.py hipify_patch.patch; fi \
72 && python3 setup.py install \
73 && cd ..; \
@@ -76,7 +75,7 @@ RUN if [ "$BUILD_FA" = "1" ]; then \
75
76 # Error related to odd state for numpy 1.20.3 where there is no METADATA etc, but an extra LICENSES_bundled.txt.
77 # Manually removed it so that later steps of numpy upgrade can continue
79 -RUN if [ "$BASE_IMAGE" = "rocm/pytorch:rocm6.0_ubuntu20.04_py3.9_pytorch_2.1.1" ]; then \
78 +RUN if [ "$BASE_IMAGE" = "$ROCm_6_0_BASE" ]; then \
79 rm -rf /opt/conda/envs/py_3.9/lib/python3.9/site-packages/numpy-1.20.3.dist-info/; fi
80
81 # build triton
@@ -107,11 +106,11 @@ ENV CCACHE_DIR=/root/.cache/ccache
106 RUN --mount=type=cache,target=/root/.cache/ccache \
107 --mount=type=cache,target=/root/.cache/pip \
108 pip install -U -r requirements-rocm.txt \
110 - && patch /opt/rocm/include/hip/amd_detail/amd_hip_bf16.h ./rocm_patch/rocm_bf16.patch \
109 + && if [ "$BASE_IMAGE" = "$ROCm_6_0_BASE" ]; then \
110 + patch /opt/rocm/include/hip/amd_detail/amd_hip_bf16.h ./rocm_patch/rocm_bf16.patch; fi \
111 && python3 setup.py install \
112 - && cp build/lib.linux-x86_64-cpython-39/vllm/_C.abi3.so vllm/ \
113 - && cp build/lib.linux-x86_64-cpython-39/vllm/_punica_C.abi3.so vllm/ \
114 - && cp build/lib.linux-x86_64-cpython-39/vllm/_moe_C.abi3.so vllm/ \
112 + && export VLLM_PYTHON_VERSION=$(python -c "import sys; print(str(sys.version_info.major) + str(sys.version_info.minor))") \
113 + && cp build/lib.linux-x86_64-cpython-${VLLM_PYTHON_VERSION}/vllm/*.so vllm/ \
114 && cd ..
115
116
cmake/utils.cmake
+4 -1
@@ -155,8 +155,11 @@ macro(override_gpu_arches GPU_ARCHES GPU_LANG GPU_SUPPORTED_ARCHES)
155 # Find the intersection of the supported + detected architectures to
156 # set the module architecture flags.
157 #
158 +
159 + set(VLLM_ROCM_SUPPORTED_ARCHS "gfx908;gfx90a;gfx942;gfx1100")
160 +
161 set(${GPU_ARCHES})
159 - foreach (_ARCH ${CMAKE_HIP_ARCHITECTURES})
162 + foreach (_ARCH ${VLLM_ROCM_SUPPORTED_ARCHS})
163 if (_ARCH IN_LIST _GPU_SUPPORTED_ARCHES_LIST)
164 list(APPEND ${GPU_ARCHES} ${_ARCH})
165 endif()