fix: Working Arm Build Dockerfile for Vllm_v1 (#1844)

container/Dockerfile.vllm

@@ -119,12 +119,13 @@ WORKDIR /workspace
 
 ### NIXL SETUP ###
 # Copy nixl source, and use commit hash as cache hint
+# TEMP: disable gds backend for arm64
 COPY --from=nixl_base /opt/nixl /opt/nixl
 COPY --from=nixl_base /opt/nixl/commit.txt /opt/nixl/commit.txt
 RUN if [ "$ARCH" = "arm64" ]; then \
         cd /opt/nixl && \
         mkdir build && \
-        meson setup build/ --buildtype=release --prefix=/usr/local/nixl -Dgds_path=/usr/local/cuda/targets/sbsa-linux && \
+        meson setup build/ --buildtype=release --prefix=/usr/local/nixl -Ddisable_gds_backend=true -Dgds_path=/usr/local/cuda/targets/sbsa-linux && \
         cd build/ && \
         ninja && \
         ninja install; \
@@ -163,8 +164,10 @@ ENV PATH="${VIRTUAL_ENV}/bin:${PATH}"
 
 # Install NIXL Python module
 # TODO: Move gds_path selection based on arch into NIXL build
+# TEMP: disable gds backend for arm64
 RUN if [ "$ARCH" = "arm64" ]; then \
         cd /opt/nixl && uv build . --out-dir /workspace/wheels/nixl \
+        --config-settings=setup-args="-Ddisable_gds_backend=true" \
         --config-settings=setup-args="-Dgds_path=/usr/local/cuda/targets/sbsa-linux"; \
     else \
         cd /opt/nixl && uv build . --out-dir /workspace/wheels/nixl; \
@@ -177,22 +180,43 @@ RUN uv pip install /workspace/wheels/nixl/*.whl
 # Install vllm - keep this early in Dockerfile to avoid
 # rebuilds from unrelated source code changes
 ARG VLLM_REF="059d4cd"
+ARG MAX_JOBS=16
+ENV MAX_JOBS=$MAX_JOBS
 ENV CUDA_HOME=/usr/local/cuda
 RUN --mount=type=bind,source=./container/deps/,target=/tmp/deps \
     --mount=type=cache,target=/root/.cache/uv \
-    uv pip install pip cuda-python && \
-    mkdir /opt/vllm && \
-    cd /opt/vllm && \
-    git clone https://github.com/vllm-project/vllm.git && \
-    cd vllm && \
-    git checkout $VLLM_REF && \
-    VLLM_USE_PRECOMPILED=1 uv pip install -e . && \
-    cd tools/ep_kernels && \
-    bash install_python_libraries.sh && \
-    cd ep_kernels_workspace && \
-    git clone --recursive https://github.com/deepseek-ai/DeepGEMM.git && \
-    cd DeepGEMM && \
-    python setup.py install
+    if [ "$ARCH" = "arm64" ]; then \
+        uv pip install pip cuda-python && \
+        mkdir /opt/vllm && \
+        cd /opt/vllm && \
+        git clone https://github.com/vllm-project/vllm.git && \
+        cd vllm && \
+        git checkout $VLLM_REF && \
+        uv pip install --pre torch torchvision torchaudio --index-url https://download.pytorch.org/whl/nightly/cu128 && \
+        python use_existing_torch.py && \
+        uv pip install -r requirements/build.txt && \
+        MAX_JOBS=${MAX_JOBS} uv pip install --no-build-isolation -e . -v && \
+        cd tools/ep_kernels && \
+        bash install_python_libraries.sh && \
+        cd ep_kernels_workspace && \
+        git clone --recursive https://github.com/deepseek-ai/DeepGEMM.git && \
+        cd DeepGEMM && \
+        python setup.py install; \
+    else \
+        uv pip install pip cuda-python && \
+        mkdir /opt/vllm && \
+        cd /opt/vllm && \
+        git clone https://github.com/vllm-project/vllm.git && \
+        cd vllm && \
+        git checkout $VLLM_REF && \
+        VLLM_USE_PRECOMPILED=1 uv pip install -e . && \
+        cd tools/ep_kernels && \
+        bash install_python_libraries.sh && \
+        cd ep_kernels_workspace && \
+        git clone --recursive https://github.com/deepseek-ai/DeepGEMM.git && \
+        cd DeepGEMM && \
+        python setup.py install; \
+    fi
 
 # Common dependencies
 RUN --mount=type=bind,source=./container/deps/requirements.txt,target=/tmp/requirements.txt \

container/build.sh

@@ -389,6 +389,8 @@ ARCH="amd64"
 if [[ "$PLATFORM" == *"linux/arm64"* ]]; then
     ARCH="arm64"
     BUILD_ARGS+=" --build-arg ARCH=arm64 --build-arg ARCH_ALT=aarch64 "
+    # TEMP: Pin to nixl 0.3.1 for arm build, since 0.4.0 fails
+    NIXL_COMMIT=3503658e71143b56f9d5b1b440d84a94b9c41af8
 fi
 
 # Update DOCKERFILE if framework is VLLM