GLiNER2.5-Decide LiteRT: fp16-weight graphs s128/s256/s512 (+fp32 references), fp16/fp32 embedding tables, host contract, Python runtime, conversion scripts, Android sample, card
d9af00e verified | """Round 2 step 3: fp32 export at any contract shape (export_s128.py with the shape as an argument). | |
| Usage: export_shape.py --seq 256 (artifact exports/gliner25_decide_s{N}_fp32.tflite; an existing | |
| artifact is never overwritten). Sample inputs = the first oracle fixture that fits s{N}. | |
| Writes results/opscan_s{N}.json (op histogram, ranks, BATCH_MATMUL shapes, forbidden check) and | |
| results/export_s{N}.json (bytes, sha256, export seconds, op/forbidden summary). The forbidden check | |
| is round 1's `forbidden_check`, imported unchanged. | |
| """ | |
| import argparse | |
| import hashlib | |
| import json | |
| import time | |
| import traceback | |
| import torch | |
| from common import ROOT, dump_json, load_model | |
| from decide_graph import DecideClassifier, Shape, fixed_inputs | |
| from export_s128 import forbidden_check | |
| def main(): | |
| parser = argparse.ArgumentParser() | |
| parser.add_argument("--seq", type=int, required=True) | |
| args = parser.parse_args() | |
| seq = args.seq | |
| artifact = ROOT / "exports" / f"gliner25_decide_s{seq}_fp32.tflite" | |
| assert not artifact.exists(), "Never overwrite an exported artifact" | |
| for relative in (f"results/export_s{seq}.json", f"results/opscan_s{seq}.json"): | |
| assert not (ROOT / relative).exists(), f"Refusing to overwrite {relative}" | |
| shape = Shape(seq) | |
| oracle = json.loads((ROOT / "fixtures/oracle_fp32.json").read_text()) | |
| sample = next(f for f in oracle["fixtures"] if f["fits"][f"s{seq}"]) | |
| model = load_model() | |
| graph = DecideClassifier(model, shape).eval().requires_grad_(False) | |
| captured = torch.load(ROOT / f"fixtures/captured/{sample['id']}.pt", weights_only=False) | |
| inputs = fixed_inputs(model, captured, shape) | |
| entry = {"artifact": str(artifact.relative_to(ROOT)), "sample_fixture": sample["id"], | |
| "contract": shape.contract(), "status": "RUNNING", | |
| "converter": "litert-torch (venv lock) + converter_layout.register_rank4_matmul"} | |
| started = time.perf_counter() | |
| try: | |
| import litert_torch | |
| from converter_layout import register_rank4_matmul | |
| from litert_cpu import scan | |
| register_rank4_matmul() | |
| litert_torch.convert(graph, inputs).export(str(artifact)) | |
| elapsed = time.perf_counter() - started | |
| digest = hashlib.sha256(artifact.read_bytes()).hexdigest() | |
| opscan = scan(artifact) | |
| opscan.update(forbidden_check(artifact)) | |
| dump_json(f"results/opscan_s{seq}.json", opscan) | |
| shapes = {} | |
| for item in opscan["batch_matmuls"]: | |
| key = " x ".join(str(s) for s in item["input_shapes"]) | |
| shapes[key] = shapes.get(key, 0) + 1 | |
| entry.update(status="PASS", bytes=artifact.stat().st_size, sha256=digest, export_seconds=elapsed, | |
| operator_count=opscan["operator_count"], op_histogram=opscan["op_histogram"], | |
| max_tensor_rank=opscan["max_tensor_rank"], tensor_dtype_histogram=opscan["tensor_dtype_histogram"], | |
| all_batch_matmul_inputs_rank4=opscan["all_batch_matmul_inputs_rank4"], | |
| batch_matmul_input_shapes=shapes, forbidden_checked=opscan["checked"], | |
| forbidden_count=opscan["forbidden_count"], forbidden_found=opscan["forbidden_found"], | |
| signatures=opscan["signatures"], opscan=f"results/opscan_s{seq}.json") | |
| dump_json(f"results/export_s{seq}.json", entry) | |
| print(json.dumps({k: v for k, v in entry.items() if k not in ("contract", "signatures")}, indent=1), flush=True) | |
| except BaseException: | |
| log = ROOT / f"logs/export_{artifact.stem}.traceback.txt" | |
| log.write_text(traceback.format_exc()) | |
| entry.update(status="FAIL", export_seconds=time.perf_counter() - started, traceback=str(log.relative_to(ROOT))) | |
| dump_json(f"results/export_attempt_{artifact.stem}.json", entry) | |
| raise | |
| if __name__ == "__main__": | |
| main() | |