Merge remote-tracking branch 'origin/main' into upstream_merge_2025_04_21

Author: gshtras

.buildkite/test-template.j2

@@ -7,7 +7,7 @@ steps:
   - label: ":docker: build image"
     depends_on: ~
     commands:
-      - "docker build --build-arg max_jobs=16 --tag {{ docker_image_amd }} -f Dockerfile.rocm  --build-arg ARG_PYTORCH_ROCM_ARCH='gfx90a;gfx942' --target test --progress plain ."
+      - "docker build --build-arg max_jobs=16 --tag {{ docker_image_amd }} -f docker/Dockerfile.rocm  --build-arg ARG_PYTORCH_ROCM_ARCH='gfx90a;gfx942' --target test --progress plain ."
       - "docker push {{ docker_image_amd }}"
     key: "amd-build"
     env:
@@ -20,6 +20,7 @@ steps:
           limit: 5
     agents:
       queue: amd-cpu
+    soft_fail: false
 
 {% for step in steps %}
 {% if step.mirror_hardwares and "amd" in step.mirror_hardwares %}
@@ -37,10 +38,10 @@ steps:
       queue: amd_gpu
 {% endif%}
     commands: 
-      - bash .buildkite/run-amd-test.sh "cd {{ (step.working_dir or default_working_dir) | safe  }} ; {{ step.command  or (step.commands | join(" && ")) | safe }}"
+      - bash .buildkite/scripts/hardware_ci/run-amd-test.sh "cd {{ (step.working_dir or default_working_dir) | safe  }} ; {{ step.command  or (step.commands | join(" && ")) | safe }}"
     env:
       DOCKER_BUILDKIT: "1"
     priority: 100
-    soft_fail: true
+    soft_fail: false
 {% endif %}
 {% endfor %}

docker/Dockerfile.rocm_base

@@ -1,12 +1,11 @@
 ARG BASE_IMAGE=rocm/dev-ubuntu-22.04:6.4-complete
 ARG HIPBLASLT_BRANCH="0f5d6c6d"
+ARG HIPBLAS_COMMON_BRANCH="9b80ba8e"
 ARG LEGACY_HIPBLASLT_OPTION=
-ARG RCCL_BRANCH="648a58d"
-ARG RCCL_REPO="https://github.com/ROCm/rccl"
-ARG TRITON_BRANCH="5fe38ffd"
+ARG TRITON_BRANCH="981e987e"
 ARG TRITON_REPO="https://github.com/triton-lang/triton.git"
-ARG PYTORCH_BRANCH="13417947"
-ARG PYTORCH_VISION_BRANCH="v0.22.0-rc5"
+ARG PYTORCH_BRANCH="295f2ed4"
+ARG PYTORCH_VISION_BRANCH="v0.21.0"
 ARG PYTORCH_REPO="https://github.com/pytorch/pytorch.git"
 ARG PYTORCH_VISION_REPO="https://github.com/pytorch/vision.git"
 ARG FA_BRANCH="1a7f4dfa"
@@ -45,25 +44,26 @@ RUN pip install -U packaging 'cmake<4' ninja wheel setuptools pybind11 Cython
 
 FROM base AS build_hipblaslt
 ARG HIPBLASLT_BRANCH
+ARG HIPBLAS_COMMON_BRANCH
 # Set to "--legacy_hipblas_direct" for ROCm<=6.2
 ARG LEGACY_HIPBLASLT_OPTION
+RUN git clone https://github.com/ROCm/hipBLAS-common.git
+RUN apt-get remove -y hipblaslt && apt-get autoremove -y && apt-get autoclean -y
+RUN cd hipBLAS-common \
+    && git checkout ${HIPBLAS_COMMON_BRANCH} \
+    && mkdir build \
+    && cd build \
+    && cmake .. \
+    && make package \
+    && dpkg -i ./*.deb
 RUN git clone https://github.com/ROCm/hipBLASLt
 RUN cd hipBLASLt \
     && git checkout ${HIPBLASLT_BRANCH} \
     && apt-get install -y llvm-dev \
     && ./install.sh -dc --architecture ${PYTORCH_ROCM_ARCH} ${LEGACY_HIPBLASLT_OPTION} \
     && cd build/release \
     && make package
-RUN mkdir -p /app/install && cp /app/hipBLASLt/build/release/*.deb /app/install
-
-FROM base AS build_rccl
-ARG RCCL_BRANCH
-ARG RCCL_REPO
-RUN git clone ${RCCL_REPO}
-RUN cd rccl \
-    && git checkout ${RCCL_BRANCH} \
-    && ./install.sh -p --amdgpu_targets ${PYTORCH_ROCM_ARCH}
-RUN mkdir -p /app/install && cp /app/rccl/build/release/*.deb /app/install
+RUN mkdir -p /app/install && cp /app/hipBLASLt/build/release/*.deb /app/hipBLAS-common/build/*.deb /app/install
 
 FROM base AS build_triton
 ARG TRITON_BRANCH
@@ -119,15 +119,25 @@ RUN cd aiter \
 RUN pip install pyyaml && cd aiter && PREBUILD_KERNELS=1 GPU_ARCHS=gfx942 python3 setup.py bdist_wheel --dist-dir=dist && ls /app/aiter/dist/*.whl
 RUN mkdir -p /app/install && cp /app/aiter/dist/*.whl /app/install
 
+FROM base AS debs
+RUN mkdir /app/debs
+RUN --mount=type=bind,from=build_hipblaslt,src=/app/install/,target=/install \
+    cp /install/*.deb /app/debs
+RUN --mount=type=bind,from=build_triton,src=/app/install/,target=/install \
+    cp /install/*.whl /app/debs
+RUN --mount=type=bind,from=build_amdsmi,src=/app/install/,target=/install \
+    cp /install/*.whl /app/debs
+RUN --mount=type=bind,from=build_pytorch,src=/app/install/,target=/install \
+    cp /install/*.whl /app/debs
+RUN --mount=type=bind,from=build_aiter,src=/app/install/,target=/install \
+    cp /install/*.whl /app/debs
+
 FROM base AS final
 RUN --mount=type=bind,from=build_hipblaslt,src=/app/install/,target=/install \
     dpkg -i /install/*deb \
-    && sed -i 's/, hipblaslt-dev \(.*\), hipcub-dev/, hipcub-dev/g' /var/lib/dpkg/status \
-    && sed -i 's/, hipblaslt \(.*\), hipfft/, hipfft/g' /var/lib/dpkg/status
-RUN --mount=type=bind,from=build_rccl,src=/app/install/,target=/install \
-    dpkg -i /install/*deb \
-    && sed -i 's/, rccl-dev \(.*\), rocalution/, rocalution/g' /var/lib/dpkg/status \
-    && sed -i 's/, rccl \(.*\), rocalution/, rocalution/g' /var/lib/dpkg/status
+    && perl -p -i -e 's/, hipblas-common-dev \([^)]*?\), /, /g' /var/lib/dpkg/status \
+    && perl -p -i -e 's/, hipblaslt-dev \([^)]*?\), /, /g' /var/lib/dpkg/status \
+    && perl -p -i -e 's/, hipblaslt \([^)]*?\), /, /g' /var/lib/dpkg/status
 RUN --mount=type=bind,from=build_triton,src=/app/install/,target=/install \
     pip install /install/*.whl
 RUN --mount=type=bind,from=build_amdsmi,src=/app/install/,target=/install \
@@ -141,8 +151,6 @@ ARG BASE_IMAGE
 ARG HIPBLAS_COMMON_BRANCH
 ARG HIPBLASLT_BRANCH
 ARG LEGACY_HIPBLASLT_OPTION
-ARG RCCL_BRANCH
-ARG RCCL_REPO
 ARG TRITON_BRANCH
 ARG TRITON_REPO
 ARG PYTORCH_BRANCH
@@ -154,10 +162,9 @@ ARG FA_REPO
 ARG AITER_BRANCH
 ARG AITER_REPO
 RUN echo "BASE_IMAGE: ${BASE_IMAGE}" > /app/versions.txt \
+    && echo "HIPBLAS_COMMON_BRANCH: ${HIPBLAS_COMMON_BRANCH}" >> /app/versions.txt \
     && echo "HIPBLASLT_BRANCH: ${HIPBLASLT_BRANCH}" >> /app/versions.txt \
     && echo "LEGACY_HIPBLASLT_OPTION: ${LEGACY_HIPBLASLT_OPTION}" >> /app/versions.txt \
-    && echo "RCCL_BRANCH: ${RCCL_BRANCH}" >> /app/versions.txt \
-    && echo "RCCL_REPO: ${RCCL_REPO}" >> /app/versions.txt \
     && echo "TRITON_BRANCH: ${TRITON_BRANCH}" >> /app/versions.txt \
     && echo "TRITON_REPO: ${TRITON_REPO}" >> /app/versions.txt \
     && echo "PYTORCH_BRANCH: ${PYTORCH_BRANCH}" >> /app/versions.txt \

docs/dev-docker/README.md

@@ -26,6 +26,9 @@ Pull the most recent validated docker image with `docker pull rocm/vllm-dev:main
 - Detokenizer disablement
 - Torch.compile support
 
+## Known Issues and Workarounds
+- Mem fault encountered when running the model meta 405 fp8. To workaround this issue, set PYTORCH_TUNABLEOP_ENABLED=0
+
 ## Performance Results
 
 The data in the following tables is a reference point to help users validate observed performance. It should not be considered as the peak performance that can be delivered by AMD Instinct™ MI300X accelerator with vLLM. See the MLPerf section in this document for information about MLPerf 4.1 inference results. The performance numbers above were collected using the steps below.