jev-training-data/convert_mmbert.py

39 lines
1.5 KiB
Python

#!/usr/bin/env python3
"""convert_mmbert.py — mmBERT bin → 完整 Laya DecisionModel checkpoint (strict 可加载)"""
import json, os, sys, shutil, torch
from safetensors.torch import save_file
base = sys.argv[1] if len(sys.argv) > 1 else "/tmp/base"
out = sys.argv[2] if len(sys.argv) > 2 else base
sys.path.insert(0, "/tmp/layasrc")
from laya.common import build_model
with open(os.path.join(base, "rl_agent_config.json")) as f:
cfg = json.load(f)
dm = build_model(cfg, encoder_dir=None) # 随机初始化的完整 DecisionModel
dm_state = dm.state_dict()
bin_sd = torch.load(os.path.join(base, "pytorch_model.bin"), map_location="cpu", weights_only=True)
matched, dropped = 0, 0
for k, v in bin_sd.items():
if k.startswith("cls."):
dropped += 1
continue
ek = "encoder." + k
if ek in dm_state:
dm_state[ek] = v.to(torch.float32)
matched += 1
else:
dropped += 1
save_file(dm_state, os.path.join(out, "model.safetensors"), metadata={"format": "pt"})
os.makedirs(os.path.join(out, "tokenizer"), exist_ok=True)
for fn in ["tokenizer.json", "tokenizer_config.json", "special_tokens_map.json"]:
src = os.path.join(base, fn)
if os.path.exists(src):
shutil.copy(src, os.path.join(out, "tokenizer", fn))
shutil.copy(os.path.join(base, "config.json"), os.path.join(out, "config.json"))
shutil.copy(os.path.join(base, "rl_agent_config.json"), os.path.join(out, "rl_agent_config.json"))
print(f"[convert] encoder匹配 {matched}, 丢弃 {dropped}, 总键 {len(dm_state)}", flush=True)