GPU Serve CPU Fallback Dogfood
Dogfood test for GH-261: CPU fallback when GPU serve fails per-request.
Starts apr serve --gpu on a small APR model, then verifies:
/healthreportsgpu_fallback: true/v1/completionsreturns a response (GPU or CPU-fallback)/v1/chat/completionsreturns a well-formed OpenAI-compatible response
Requires a small APR model (e.g. qwen2-0.5b-int8.apr).
Run
cargo run --release --example gpu_fallback_dogfood
Source
// Run this example:
// cargo run --example gpu_fallback_dogfood
//
// See the CLI reference and source code in crates/ for implementation details.