vLLM: take the CPU wheel from the release's assets

The latest release is tagged v0.33.0rc3, but its wheels are named 0.33.0, so the URL built from the tag answered 404 and the CPU install stopped (2 runs). The wheel now comes from the release's asset list.

Author: MickLesk

ct/vllm.sh

@@ -50,9 +50,14 @@ function update_script() {
         --extra-index-url https://wheels.vllm.ai/rocm/
       ;;
     cpu)
-      $STD uv pip install --python /opt/vllm/bin/python \
-        "https://github.com/vllm-project/vllm/releases/download/v${VLLM_VERSION}/vllm-${VLLM_VERSION}+cpu-cp38-abi3-manylinux_2_39_x86_64.whl" \
-        --torch-backend cpu
+      github_api_call "https://api.github.com/repos/vllm-project/vllm/releases/tags/v${VLLM_VERSION}" /tmp/vllm-release.json
+      VLLM_WHEEL="$(jq -r '.assets[] | select(.name | endswith("+cpu-cp38-abi3-manylinux_2_39_x86_64.whl")) | .browser_download_url' /tmp/vllm-release.json | head -n1)"
+      rm -f /tmp/vllm-release.json
+      [[ -n "$VLLM_WHEEL" ]] || {
+        msg_error "vLLM ${VLLM_VERSION} has no CPU wheel for x86_64"
+        exit 1
+      }
+      $STD uv pip install --python /opt/vllm/bin/python "$VLLM_WHEEL" --torch-backend cpu
       ;;
     *)
       $STD uv pip install --python /opt/vllm/bin/python --upgrade vllm

install/vllm-install.sh

@@ -56,9 +56,14 @@ cpu)
     msg_error "vLLM publishes no prebuilt CPU wheel for $(uname -m)"
     exit 1
   fi
-  $STD uv pip install --python /opt/vllm/bin/python \
-    "https://github.com/vllm-project/vllm/releases/download/v${VLLM_VERSION}/vllm-${VLLM_VERSION}+cpu-cp38-abi3-manylinux_2_39_x86_64.whl" \
-    --torch-backend cpu
+  github_api_call "https://api.github.com/repos/vllm-project/vllm/releases/tags/v${VLLM_VERSION}" /tmp/vllm-release.json
+  VLLM_WHEEL="$(jq -r '.assets[] | select(.name | endswith("+cpu-cp38-abi3-manylinux_2_39_x86_64.whl")) | .browser_download_url' /tmp/vllm-release.json | head -n1)"
+  rm -f /tmp/vllm-release.json
+  [[ -n "$VLLM_WHEEL" ]] || {
+    msg_error "vLLM ${VLLM_VERSION} has no CPU wheel for x86_64"
+    exit 1
+  }
+  $STD uv pip install --python /opt/vllm/bin/python "$VLLM_WHEEL" --torch-backend cpu
   ;;
 *)
   msg_error "Unknown var_backend '${var_backend}' (expected auto, cuda, rocm or cpu)"