Merge branch 'main' into 3-node

.gitignore

@@ -1 +1,2 @@
-.env
+.env
+build-metadata.yaml

Dockerfile

@@ -250,7 +250,7 @@ RUN --mount=type=bind,from=base,source=/workspace/vllm/nccl/build/pkg/deb,target
     libxcb1 \
     && cd /workspace/nccl-pkg && apt install -y --no-install-recommends --allow-downgrades ./*.deb \
     && rm -rf /var/lib/apt/lists/* \
-    && pip install uv 
+    && pip install uv
 
 # Set final working directory
 WORKDIR $VLLM_BASE_DIR
@@ -265,7 +265,7 @@ ARG PRE_TRANSFORMERS=0
 # Install deps
 RUN --mount=type=cache,id=uv-cache,target=/root/.cache/uv \
      uv pip install torch torchvision torchaudio triton --index-url https://download.pytorch.org/whl/nightly/cu130 && \
-     uv pip install nvidia-nvshmem-cu13 "apache-tvm-ffi<0.2" 
+     uv pip install nvidia-nvshmem-cu13 "apache-tvm-ffi<0.2"
 
 # Install wheels from host ./wheels/ (bind-mounted from build context — no layer bloat)
 # With --tf5: override vLLM's transformers<5 constraint to get transformers>=5
@@ -294,4 +294,7 @@ RUN --mount=type=cache,id=uv-cache,target=/root/.cache/uv \
 
 # Fix NCCL
 RUN rm /usr/local/lib/python3.12/dist-packages/nvidia/nccl/lib/libnccl.so.2 && \
-    ln -s /usr/lib/aarch64-linux-gnu/libnccl.so.2 /usr/local/lib/python3.12/dist-packages/nvidia/nccl/lib/libnccl.so.2
+    ln -s /usr/lib/aarch64-linux-gnu/libnccl.so.2 /usr/local/lib/python3.12/dist-packages/nvidia/nccl/lib/libnccl.so.2
+    
+# Build metadata (generated by build-and-copy.sh)
+COPY build-metadata.yaml /workspace/build-metadata.yaml

Dockerfile.mxfp4

@@ -274,6 +274,9 @@ ENV PATH=$VLLM_BASE_DIR:$PATH
 RUN --mount=type=cache,id=uv-cache,target=/root/.cache/uv \
     uv pip install ray[default] fastsafetensors nvidia-nvshmem-cu13
 
+# Build metadata (generated by build-and-copy.sh)
+COPY build-metadata.yaml /workspace/build-metadata.yaml
+
 # If not compiling Triton
 # remove triton-kernels as they are not compatible with this vLLM version yet
 RUN uv pip uninstall triton-kernels

build-and-copy.sh

@@ -33,10 +33,42 @@ cleanup() {
         echo "Cleaning up temporary image $TMP_IMAGE"
         rm -f "$TMP_IMAGE"
     fi
+    rm -f ./build-metadata.yaml
 }
 
 trap cleanup EXIT
 
+generate_build_metadata() {
+    local dockerfile="$1"
+    local vllm_version="$2"
+    local vllm_commit="$3"
+    local flashinfer_commit="$4"
+    local vllm_ref="$5"
+    local pre_transformers="$6"
+    local exp_mxfp4="$7"
+    local vllm_prs="$8"
+
+    local base_image
+    base_image=$(grep -m1 '^FROM .* AS runner' "$dockerfile" | awk '{print $2}')
+
+    cat > ./build-metadata.yaml <<EOF
+build_date: $(date -u +"%Y-%m-%dT%H:%M:%SZ")
+build_script_commit: $(git rev-parse HEAD 2>/dev/null || echo "unknown")
+vllm_version: ${vllm_version:-unknown}
+vllm_commit: ${vllm_commit:-unknown}
+flashinfer_commit: ${flashinfer_commit:-unknown}
+gpu_arch: ${GPU_ARCH_LIST}
+base_image: ${base_image:-unknown}
+build_args:
+  vllm_ref: ${vllm_ref}
+  transformers_5: ${pre_transformers}
+  exp_mxfp4: ${exp_mxfp4}
+  vllm_prs: "${vllm_prs}"
+  build_jobs: ${BUILD_JOBS}
+EOF
+    echo "Generated build-metadata.yaml"
+}
+
 add_copy_hosts() {
     local token part
     for token in "$@"; do
@@ -365,6 +397,13 @@ RUNNER_BUILD_TIME=0
 if [ "$NO_BUILD" = false ]; then
     if [ "$EXP_MXFP4" = true ]; then
         echo "Building with experimental MXFP4 support..."
+
+        # Generate build metadata YAML for mxfp4 build
+        MXFP4_VLLM_SHA=$(grep -m1 '^ARG VLLM_SHA=' Dockerfile.mxfp4 | cut -d= -f2)
+        MXFP4_FLASHINFER_SHA=$(grep -m1 '^ARG FLASHINFER_SHA=' Dockerfile.mxfp4 | cut -d= -f2)
+        generate_build_metadata Dockerfile.mxfp4 "unknown" "$MXFP4_VLLM_SHA" "$MXFP4_FLASHINFER_SHA" \
+            "mxfp4-pinned" "false" "true" ""
+
         CMD=("docker" "build" "-t" "$IMAGE_TAG" "${COMMON_BUILD_FLAGS[@]}" "-f" "Dockerfile.mxfp4" ".")
         echo "Building image with command: ${CMD[*]}"
         BUILD_START=$(date +%s)
@@ -496,6 +535,15 @@ if [ "$NO_BUILD" = false ]; then
             exit 1
         fi
 
+        # Generate build metadata YAML
+        VLLM_VERSION=$(ls ./wheels/vllm-*.whl 2>/dev/null | head -1 | sed 's|.*/vllm-||;s|-.*||')
+        VLLM_COMMIT=""
+        [ -f "./wheels/.vllm-commit" ] && VLLM_COMMIT=$(cat ./wheels/.vllm-commit)
+        FLASHINFER_COMMIT=""
+        [ -f "./wheels/.flashinfer-commit" ] && FLASHINFER_COMMIT=$(cat ./wheels/.flashinfer-commit)
+        generate_build_metadata Dockerfile "$VLLM_VERSION" "$VLLM_COMMIT" "$FLASHINFER_COMMIT" \
+            "$VLLM_REF" "$PRE_TRANSFORMERS" "false" "$VLLM_PRS"
+
         RUNNER_CMD=("docker" "build"
             "-t" "$IMAGE_TAG"
             "${COMMON_BUILD_FLAGS[@]}")