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  \
+    if [ "$WHEELS_FROM_GITHUB_RELEASE" = "1" ]; then \
-    --torch-backend=auto \
+        export VLLM_VERSION=$(curl -s https://api.github.com/repos/vllm-project/vllm/releases/latest | jq -r .tag_name | sed 's/^v//') && \
-    --extra-index-url $VLLM_WHEELS_URL
+        uv pip install --system --break-system-packages -U https://github.com/vllm-project/vllm/releases/download/v${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

README.md

@@ -1,8 +1,10 @@
 
-# vLLM Ray Cluster Node Docker for DGX Spark
+# vLLM Docker Optimized for DGX Spark (single or multi-node)
 
 This repository contains the Docker configuration and startup scripts to run a multi-node vLLM inference cluster using Ray. It supports InfiniBand/RDMA (NCCL) and custom environment configuration for high-performance setups.
 
+While it was primarily developed to support multi-node inference, it works just as well on a single node setups.
+
 ## Table of Contents
 
 - [DISCLAIMER](#disclaimer)
@@ -20,7 +22,7 @@ This repository contains the Docker configuration and startup scripts to run a m
 
 This repository is not affiliated with NVIDIA or their subsidiaries. This is a community effort aimed to help DGX Spark users to set up and run the most recent versions of vLLM on Spark cluster or single nodes. 
 
-The Dockerfile builds from the main branch of VLLM, so depending on when you run the build process, it may not be in fully functioning state. You can target a specific vLLM release by setting `--vllm-ref` parameter.
+The Dockerfile builds from the main branch of VLLM, so depending on when you run the build process, it may not be in fully functioning state. You can target a specific vLLM release by setting `--vllm-ref` parameter or use `--use-wheels release` to install pre-built release wheels.
 
 ## CHANGELOG
 
@@ -44,6 +46,11 @@ Don't do it every time you rebuild, because it will slow down compilation times.
 
 For periodic maintenance, I recommend using a filter: `docker builder prune --filter until=72h`
 
+### 2025-12-21
+
+Pre-built wheels now support release versions. Use with `--use-wheels release`.
+Using nightly wheels or building from source is recommended for better performance.
+
 ### 2025-12-20
 
 - Limited ccache to 50G when building from source to reduce build cache size.
@@ -52,7 +59,7 @@ For periodic maintenance, I recommend using a filter: `docker builder prune --fi
   - Allows building the container using pre-built vLLM wheels instead of compiling from source.
   - Reduced build time and container size.
   - `mode` is optional and defaults to `nightly`.
-  - Supported modes: `nightly` (release wheels are broken with CUDA 13 currently).
+  - Supported modes: `nightly` (release wheels are broken with CUDA 13 currently). UPDATE: `release` also works now.
 ### 2025-12-19
 
 Updated `build-and-copy.sh` to support copying to multiple hosts (thanks @ericlewis for the contribution).

build-and-copy.sh

@@ -155,11 +155,7 @@ if [ "$NO_BUILD" = false ]; then
         echo "Using pre-built vLLM wheels (mode: $USE_WHEELS_MODE)"
         CMD+=("-f" "Dockerfile.wheels")
         if [ "$USE_WHEELS_MODE" = "release" ]; then
-             echo "Release wheels are currently broken with CUDA 13, use nightly instead."
+             CMD+=("--build-arg" "WHEELS_FROM_GITHUB_RELEASE=1")
-             exit 1
-             CMD+=("--build-arg" "VLLM_WHEELS_URL=https://wheels.vllm.ai/cu130")
-        else
-             CMD+=("--build-arg" "VLLM_WHEELS_URL=https://wheels.vllm.ai/nightly/cu130")
         fi
     else
         echo "Building vLLM from source"