The Metal backend requires macOS on Apple Silicon. On other platforms, or when checking a CPU-only build, run generation with:
cargo run --release --bin generate -- --backend cpu --prompt "Once upon a time"Model weights and tokenizer files are intentionally not committed. Download them before running generation or the end-to-end reference check:
python python/fetch_gpt2.pyThe files are placed under models/gpt2/ and require roughly 550 MB of disk space.
cargo test --test parity -- --nocapture validates Metal kernels against the pure-Rust CPU reference, so it must run on an Apple Silicon Mac. The regular library test suite remains useful for CPU-only environments:
cargo test --lib