prepare_checkpoint.py

#!/usr/bin/env python
"""Prepare the laya checkpoint for a low-memory run (run this once, before label_anger.py).

  1. `models/laya-ml/multilingual`  - the published `convaiinnovations/laya` repo, but only the
     `multilingual/*` subfolder (643 MB).  laya's own loader accepts a local directory, so nothing
     else is downloaded and no Hub call happens at inference time.
  2. `models/enc-bf16/model.safetensors` - the `encoder.*` tensors of that checkpoint, prefix
     stripped, in one file next to the encoder config.  laya's `build_model` builds the encoder from
     `encoder/config.json` with `from_config`, i.e. it allocates a full random fp32 copy of the
     encoder (1.29 GB) that `Agent.__init__` then overwrites from `model.safetensors`.  Loading the
     same tensors through `from_pretrained(..., low_cpu_mem_usage=True)` instead materialises each
     tensor exactly once, in bf16 - see laya_opt.py.  Verified: the two paths end up with identical
     weights (laya's own strict `load_state_dict` overwrites every key), and logits match laya's
     stock `Agent.predict` to 4 decimal places (dev_probe.py).

Usage: python prepare_checkpoint.py [--force]
"""
import argparse
import json
import os


SRC_REPO = "convaiinnovations/laya"
CKPT = "models/laya-ml/multilingual"
ENC = "models/enc-bf16"


def main():
    ap = argparse.ArgumentParser()
    ap.add_argument("--force", action="store_true")
    args = ap.parse_args()
    import torch
    from safetensors import safe_open
    from safetensors.torch import save_file

    if args.force or not os.path.exists(os.path.join(CKPT, "model.safetensors")):
        from huggingface_hub import snapshot_download

        os.makedirs(CKPT.rsplit("/", 1)[0], exist_ok=True)
        snapshot_download(SRC_REPO, allow_patterns=["multilingual/*"], local_dir="models/laya-ml")
        print("[prep] checkpoint at", CKPT)
    else:
        print("[prep] checkpoint already present")

    os.makedirs(ENC, exist_ok=True)
    cfg = json.load(open(os.path.join(CKPT, "encoder", "config.json")))
    json.dump(cfg, open(os.path.join(ENC, "config.json"), "w"), indent=2)
    out_path = os.path.join(ENC, "model.safetensors")
    if args.force or not os.path.exists(out_path):
        enc = {}
        with safe_open(os.path.join(CKPT, "model.safetensors"), framework="pt") as f:
            for k in f.keys():
                if k.startswith("encoder."):
                    enc[k[len("encoder."):]] = f.get_tensor(k)   # already bf16 on disk
        save_file(enc, out_path, metadata={"format": "pt"})
        n = sum(t.numel() for t in enc.values())
        print("[prep] encoder copy: %d tensors, %.1fM params, %.0f MB -> %s"
              % (len(enc), n / 1e6, os.path.getsize(out_path) / 1e6, out_path))
    else:
        print("[prep] encoder copy already present")

    # sanity: every key laya's DecisionModel expects must exist in the checkpoint
    from laya.common import build_model  # noqa: F401  (import check only)

    with safe_open(os.path.join(CKPT, "model.safetensors"), framework="pt") as f:
        keys = set(f.keys())
    assert {"encoder.embeddings.tok_embeddings.weight", "scorer.0.weight", "type_emb.weight",
            "act_head.0.weight", "temperature"} <= keys, sorted(keys)[:5]
    print("[prep] ok - checkpoint exposes %d tensors incl. all laya head keys" % len(keys))


if __name__ == "__main__":
    main()