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 5, 21:00 UTC
[ROCm][CI/Build] ROCm 7.2.1 release version; torch 2.10; triton 3.6 (#38252)
Signed-off-by: Gregory Shtrasberg <Gregory.Shtrasberg@amd.com>
Gregory Shtrasberg committed
Mar 27, 2026 at 18:03 UTC
731285c939836cbc79c5e8e34636c746552e21d8
2 files changed
+16
-7
docker/Dockerfile.rocm
+3
@@ -386,6 +386,9 @@ ENV MIOPEN_DEBUG_CONV_GEMM=0
386
# will not be imported by other tests
387
RUN mkdir src && mv vllm src/vllm
388
389
+# This is a workaround to ensure pytest exits with the correct status code in CI tests.
390
+RUN echo "import os\n\ndef pytest_sessionfinish(session, exitstatus):\n os._exit(int(exitstatus))" > /vllm-workspace/conftest.py
391
+
392
# -----------------------
393
# Final vLLM image
394
FROM base AS final
docker/Dockerfile.rocm_base
+13
-7
@@ -1,7 +1,7 @@
1
-ARG BASE_IMAGE=rocm/dev-ubuntu-22.04:7.0-complete
2
-ARG TRITON_BRANCH="57c693b6"
1
+ARG BASE_IMAGE=rocm/dev-ubuntu-22.04:7.2.1-complete
2
+ARG TRITON_BRANCH="ba5c1517"
3
ARG TRITON_REPO="https://github.com/ROCm/triton.git"
4
-ARG PYTORCH_BRANCH="89075173"
4
+ARG PYTORCH_BRANCH="8514f051" # release/2.10 as of 3/17
5
ARG PYTORCH_REPO="https://github.com/ROCm/pytorch.git"
6
ARG PYTORCH_VISION_BRANCH="v0.24.1"
7
ARG PYTORCH_VISION_REPO="https://github.com/pytorch/vision.git"
@@ -114,6 +114,8 @@ ARG TRITON_REPO
114
RUN git clone ${TRITON_REPO}
115
RUN cd triton \
116
&& git checkout ${TRITON_BRANCH} \
117
+ && git config --global user.email "you@example.com" && git config --global user.name "Your Name" \
118
+ && git cherry-pick 555d04f \
119
&& if [ ! -f setup.py ]; then cd python; fi \
120
&& python3 setup.py bdist_wheel --dist-dir=dist \
121
&& mkdir -p /app/install && cp dist/*.whl /app/install
@@ -142,10 +144,14 @@ ARG PYTORCH_VISION_REPO
144
ARG PYTORCH_AUDIO_REPO
145
ARG USE_SCCACHE
146
147
+RUN apt-get update && apt-get install -y pkg-config liblzma-dev
148
RUN git clone ${PYTORCH_REPO} pytorch
146
-RUN cd pytorch && git checkout ${PYTORCH_BRANCH} \
147
- && pip install -r requirements.txt && git submodule update --init --recursive \
148
- && python3 tools/amd_build/build_amd.py \
149
+RUN cd pytorch && git checkout ${PYTORCH_BRANCH}
150
+RUN cd pytorch \
151
+ && pip install -r requirements.txt && git submodule update --init --recursive
152
+RUN cd pytorch/third_party/kineto \
153
+ && git remote add rocm https://github.com/ROCm/kineto && git fetch rocm && git checkout 2d73be3
154
+RUN cd pytorch && python3 tools/amd_build/build_amd.py \
155
&& if [ "$USE_SCCACHE" = "1" ]; then \
156
export HIP_CLANG_PATH=/opt/sccache-wrappers \
157
&& export CMAKE_C_COMPILER_LAUNCHER=sccache \
@@ -239,7 +245,7 @@ RUN pip install pyyaml && cd aiter \
245
export HIP_CLANG_PATH=/opt/sccache-wrappers \
246
&& sccache --show-stats; \
247
fi \
242
- && GPU_ARCHS=${AITER_ROCM_ARCH} python3 setup.py bdist_wheel --dist-dir=dist \
248
+ && PREBUILD_KERNELS=1 GPU_ARCHS=${AITER_ROCM_ARCH} python3 setup.py bdist_wheel --dist-dir=dist \
249
&& if [ "$USE_SCCACHE" = "1" ]; then sccache --show-stats; fi \
250
&& ls /app/aiter/dist/*.whl
251
RUN mkdir -p /app/install && cp /app/aiter/dist/*.whl /app/install