Skill to check for prebuilt wheels during installation

.claude/skills/check-precompiled-wheels/SKILL.md

@@ -0,0 +1,73 @@
+---
+name: check-precompiled-wheels
+description: Check if pre-compiled wheels are available for the current commit before using VLLM_USE_PRECOMPILED=1.
+---
+
+# Checking Pre-Compiled Wheels Availability
+
+When installing vLLM with `VLLM_USE_PRECOMPILED=1`, wheels must already be built for your commit. This skill verifies availability before installation to avoid failures or unexpected compilation.
+
+## Quick Check
+
+```bash
+# Detect your system's CUDA version
+CUDA_VERSION=$(nvidia-smi --query-gpu=compute_cap --format=csv,noheader | head -1)
+# Convert to wheel variant (e.g., 12.1 → cu121, 11.8 → cu118)
+VARIANT="cu${CUDA_VERSION//./}"
+
+# Get your current commit
+COMMIT=$(git rev-parse HEAD)
+
+# Check if wheels are built (only worth checking if commit is recent)
+COMMIT_AGE=$(($(date +%s) - $(git log -1 --format=%ct)))
+if [ $COMMIT_AGE -lt 7200 ]; then  # Within 2 hours
+    curl -sI "https://wheels.vllm.ai/${COMMIT}/${VARIANT}/vllm/" | head -1
+    # If you get "200 OK", wheels are available. If "404 Not Found", they're not ready yet.
+fi
+```
+
+## Using in a Script
+
+```bash
+#!/bin/bash
+COMMIT=$(git rev-parse HEAD)
+COMMIT_AGE=$(($(date +%s) - $(git log -1 --format=%ct)))
+
+# Detect CUDA version from nvidia-smi
+if command -v nvidia-smi &> /dev/null; then
+    CUDA_VERSION=$(nvidia-smi --query-gpu=compute_cap --format=csv,noheader 2>/dev/null | head -1)
+    VARIANT="cu${CUDA_VERSION//./}"
+else
+    echo "⚠️  nvidia-smi not found, skipping wheel check"
+    uv pip install -e . --torch-backend=auto
+    exit 0
+fi
+
+# Only check wheels for recent commits (avoid unnecessary requests)
+if [ $COMMIT_AGE -lt 7200 ]; then  # 2 hours
+    if curl -sf "https://wheels.vllm.ai/${COMMIT}/${VARIANT}/vllm/" > /dev/null 2>&1; then
+        echo "✅ Wheels available for $VARIANT"
+        VLLM_USE_PRECOMPILED=1 uv pip install -e . --torch-backend=auto
+    else
+        echo "❌ Wheels not ready yet. Compiling..."
+        uv pip install -e . --torch-backend=auto
+    fi
+else
+    echo "⏭️  Commit is older than 2 hours, wheels likely built. Installing..."
+    VLLM_USE_PRECOMPILED=1 uv pip install -e . --torch-backend=auto
+fi
+```
+
+## Troubleshooting
+
+**nvidia-smi not found:** The script skips wheel checks on machines without NVIDIA GPUs. Install normally without `VLLM_USE_PRECOMPILED=1`.
+
+**Wheels take 5-15 minutes to build** after a commit is pushed. For commits older than 2 hours, wheels are assumed to be available and installation proceeds with `VLLM_USE_PRECOMPILED=1`. For newer commits, the script checks before installing.
+
+**Manually check a specific commit:**
+```bash
+COMMIT=$(git rev-parse HEAD)
+VARIANT="cu121"  # Replace with your CUDA variant
+curl -sI "https://wheels.vllm.ai/${COMMIT}/${VARIANT}/vllm/" | head -1
+# "200 OK" = available, "404 Not Found" = not ready yet
+```