| |
| """Merge sharded lambda/low-rank results into one calibration artifact. |
| |
| Shards are disjoint by layer, so the merge is a dict update; the check is that all 312 target |
| layers are present, since a silently short artifact would convert without complaint and only |
| show up as a bad video. |
| """ |
| import argparse, glob |
| from pathlib import Path |
| import torch |
|
|
| ap = argparse.ArgumentParser() |
| ap.add_argument("--shards", required=True) |
| ap.add_argument("--out", required=True) |
| ap.add_argument("--span", required=True) |
| ap.add_argument("--expect", type=int, default=312) |
| a = ap.parse_args() |
|
|
| layers = {} |
| files = sorted(Path(a.shards).glob("slr_*.pt")) |
| for f in files: |
| layers.update(torch.load(f, map_location="cpu", weights_only=False)["layers"]) |
| assert len(layers) == a.expect, f"got {len(layers)} layers from {len(files)} shards, want {a.expect}" |
| torch.save({"rank": 32, "num_grids": 20, "span": a.span, |
| "base_model": "MiniMaxAI/MiniMax-H3@bfc8ed0353f5a9733be73e6b2c98ec0948195b86", |
| "layers": layers}, a.out) |
| print(f"merged {len(files)} shards, {len(layers)} layers -> {a.out}") |
|
|
