Support vLLM release wheels

Dockerfile.wheels

@@ -24,7 +24,7 @@ ENV UV_LINK_MODE=copy
 # Note: "devel" tools like cmake/gcc are NOT installed here to save space
 RUN apt update && apt upgrade -y \
     && apt install -y --allow-change-held-packages --no-install-recommends \
-    python3 python3-pip python3-dev vim curl git wget \
+    python3 python3-pip python3-dev vim curl git wget jq \
     libcudnn9-cuda-13 \
     libnccl-dev libnccl2 libibverbs1 libibverbs-dev rdma-core \
     && rm -rf /var/lib/apt/lists/* \
@@ -49,13 +49,20 @@ RUN --mount=type=cache,id=uv-cache,target=/root/.cache/uv \
 # Change THIS argument to force a fresh git clone and rebuild of vLLM
 # without re-installing the dependencies above.
 ARG CACHEBUST_VLLM=1
-ARG VLLM_WHEELS_URL=https://wheels.vllm.ai/nightly/cu130
+ARG WHEELS_FROM_GITHUB_RELEASE=0
 
-# Install nightly vLLM build from prebuilt wheels
+# Install vLLM
+# If INSTALL_FROM_GITHUB_RELEASE is 1, install from GitHub releases (specific for aarch64/cu130 as requested)
+# Otherwise, install from nightly wheels
 RUN --mount=type=cache,id=uv-cache,target=/root/.cache/uv \
-    uv pip install --system --break-system-packages -U vllm  \
-    --torch-backend=auto \
-    --extra-index-url $VLLM_WHEELS_URL
+    if [ "$WHEELS_FROM_GITHUB_RELEASE" = "1" ]; then \
+        export VLLM_VERSION=$(curl -s https://api.github.com/repos/vllm-project/vllm/releases/latest | jq -r .tag_name | sed 's/^v//') && \
+        uv pip install --system --break-system-packages -U https://github.com/vllm-project/vllm/releases/download/v${VLLM_VERSION}/vllm-${VLLM_VERSION}+cu130-cp38-abi3-manylinux_2_35_aarch64.whl --torch-backend=auto; \
+    else \
+        uv pip install --system --break-system-packages -U vllm \
+        --torch-backend=auto \
+        --extra-index-url https://wheels.vllm.ai/nightly/cu130; \
+    fi
 
 # Apply TEMPORARY PATCH for fastsafetensors loading in cluster setup - tracking https://github.com/foundation-model-stack/fastsafetensors/issues/36
 # Apply in site-packages