MUSA Benchmarks

MUSA benchmarks compare the existing torch path with optional native musa_aiter acceleration for the in-process vLLM MUSA connector.

Run on a Moore Threads MUSA host:

python -c "import torch, torch_musa; print(torch.musa.is_available(), torch.musa.device_count())"
pytest -q tests/v1/test_musa_native.py tests/v1/test_musa_connector.py -rs
python benchmarks/musa/bench_inprocess_transfer.py --memory-device musa --iters 20 --warmup-iters 5 --min-speedup 1.2

The benchmark passes when native mode is at least 1.2x faster than the torch fallback for the default shape. Use --memory-device musa when measuring native acceleration because the optional musa_aiter path consumes contiguous MUSA tensors directly. For review, attach the full benchmark command, MUSA driver version, torch_musa version, model-like shape, and speedup.