"""Stage 2: export the BC1S (Neural Engine layout) d1 function to .aimodel, fp16. .venv/bin/python export_ane.py [out/d1_ane_prompt_fp16.aimodel] """ import os, shutil, sys, time from pathlib import Path import torch from coreai_models.export.macos import export_to_coreai from d1_common import embeddings, load_meta, prompt_inputs from d1_torch import D1Fixed from d1_ane import D1ANE out = Path(sys.argv[1] if len(sys.argv) > 1 else "out/d1_ane_prompt_fp16.aimodel") meta = load_meta() D1ANE.MLP_SPLIT = int(os.environ.get("MLP_SPLIT", D1ANE.MLP_SPLIT)) model = D1ANE(D1Fixed(meta, torch.float16)).eval() _, gh, gw = meta["grid_thw"] ref = {"image": torch.zeros(1, 3, gh * 16, gw * 16, dtype=torch.float32), **{k: torch.from_numpy(v) for k, v in prompt_inputs(meta, embeddings()).items()}} # the questions are inputs t = time.time() program = export_to_coreai(model, ref, input_names=("image", "text", "head", "attend", "slots"), output_names=("logits",)) print("export_to_coreai %.1fs" % (time.time() - t), flush=True) program.optimize() if out.exists(): shutil.rmtree(out) program.save_asset(out) print("saved", out, flush=True)