NPU TorchAir Inference & Migration (主流程)

Overview

Bring an arbitrary HuggingFace model up on Ascend NPU torchair graph mode, verify it numerically against CPU/NPU-eager, and measure its speedup. The bundled scripts are model-agnostic (config-driven inputs, recursive output comparison), so the same flow applies to the next model with only a changed --model_name_or_path.

Bundled assets:

• scripts/torch_air_infer.py — minimal single-backend inference + latency stats.
• scripts/benchmark.py — 3-backend (torchair / npu_eager / cpu) accuracy + perf → JSON + Markdown.
• scripts/run_benchmark.sh — generic driver: prepare model → sanity infer → benchmark.

The one rule you cannot break

torch_npu MUST be imported before torchair, or graph mode does not engage.