"""Stage 2: export the fixed d1 decision function to .aimodel (GPU layout, fp16 weights and compute). .venv/bin/python export_gpu.py # -> out/d1_gpu_fp16.aimodel """ import os, sys, time import torch from coreai_models.export.macos import export_to_coreai from d1_common import load_meta from d1_torch import D1Fixed out = sys.argv[1] if len(sys.argv) > 1 else "out/d1_gpu_fp16.aimodel" meta = load_meta() model = D1Fixed(meta, torch.float16).eval() _, gh, gw = meta["grid_thw"] ref = {"image": torch.zeros(1, 3, gh * 16, gw * 16, dtype=torch.float32)} t = time.time() program = export_to_coreai(model, ref, input_names=("image",), output_names=("logits",)) print("export_to_coreai %.1fs" % (time.time() - t), flush=True) t = time.time() program.optimize() print("optimize %.1fs" % (time.time() - t), flush=True) if os.path.exists(out): import shutil shutil.rmtree(out) from pathlib import Path program.save_asset(Path(out)) print("saved", out, flush=True)