Remove adapter migration section and script
Browse files- convert_adapter_weights.py +0 -107
convert_adapter_weights.py
DELETED
|
@@ -1,107 +0,0 @@
|
|
| 1 |
-
from __future__ import annotations
|
| 2 |
-
|
| 3 |
-
import argparse
|
| 4 |
-
import json
|
| 5 |
-
import shutil
|
| 6 |
-
from pathlib import Path
|
| 7 |
-
|
| 8 |
-
from safetensors.torch import load_file, save_file
|
| 9 |
-
|
| 10 |
-
|
| 11 |
-
TEXT_PREFIX_IN = "base_model.model.model.text_model."
|
| 12 |
-
TEXT_PREFIX_OUT = "model.text_model."
|
| 13 |
-
PROJ_PREFIX_IN = "base_model.model.custom_text_proj."
|
| 14 |
-
PROJ_PREFIX_OUT = "model.custom_text_proj."
|
| 15 |
-
|
| 16 |
-
|
| 17 |
-
def parse_args() -> argparse.Namespace:
|
| 18 |
-
parser = argparse.ArgumentParser(
|
| 19 |
-
description="Convert the legacy paultltc/colmodernvbert LoRA adapter into the official ModernVBERT/colmodernvbert format."
|
| 20 |
-
)
|
| 21 |
-
parser.add_argument("input_dir", type=Path, help="Legacy adapter directory")
|
| 22 |
-
parser.add_argument("output_dir", type=Path, help="Converted adapter directory")
|
| 23 |
-
return parser.parse_args()
|
| 24 |
-
|
| 25 |
-
|
| 26 |
-
def ensure_new_dir(path: Path) -> None:
|
| 27 |
-
if path.exists():
|
| 28 |
-
raise FileExistsError(f"{path} already exists; refusing to overwrite it")
|
| 29 |
-
path.mkdir(parents=True)
|
| 30 |
-
|
| 31 |
-
|
| 32 |
-
def copy_support_files(src: Path, dst: Path) -> None:
|
| 33 |
-
excluded = {"adapter_model.safetensors", "adapter_config.json", "BUILD_INFO.json"}
|
| 34 |
-
for item in src.iterdir():
|
| 35 |
-
if item.name in excluded:
|
| 36 |
-
continue
|
| 37 |
-
target = dst / item.name
|
| 38 |
-
if item.is_dir():
|
| 39 |
-
shutil.copytree(item, target)
|
| 40 |
-
else:
|
| 41 |
-
shutil.copy2(item, target)
|
| 42 |
-
|
| 43 |
-
|
| 44 |
-
def convert_adapter_weights(src_dir: Path, dst_dir: Path) -> dict[str, int]:
|
| 45 |
-
src_weights = load_file(str(src_dir / "adapter_model.safetensors"))
|
| 46 |
-
out_weights = {}
|
| 47 |
-
converted_text = 0
|
| 48 |
-
converted_proj = 0
|
| 49 |
-
untouched = []
|
| 50 |
-
|
| 51 |
-
for key, value in src_weights.items():
|
| 52 |
-
if key.startswith(TEXT_PREFIX_IN):
|
| 53 |
-
out_key = TEXT_PREFIX_OUT + key[len(TEXT_PREFIX_IN) :]
|
| 54 |
-
out_weights[out_key] = value
|
| 55 |
-
converted_text += 1
|
| 56 |
-
elif key.startswith(PROJ_PREFIX_IN):
|
| 57 |
-
out_key = PROJ_PREFIX_OUT + key[len(PROJ_PREFIX_IN) :]
|
| 58 |
-
out_weights[out_key] = value
|
| 59 |
-
converted_proj += 1
|
| 60 |
-
else:
|
| 61 |
-
untouched.append(key)
|
| 62 |
-
|
| 63 |
-
if untouched:
|
| 64 |
-
sample = ", ".join(sorted(untouched)[:5])
|
| 65 |
-
raise ValueError(f"Found unexpected legacy adapter keys that were not converted: {sample}")
|
| 66 |
-
|
| 67 |
-
save_file(out_weights, str(dst_dir / "adapter_model.safetensors"))
|
| 68 |
-
return {
|
| 69 |
-
"source_tensor_count": len(src_weights),
|
| 70 |
-
"output_tensor_count": len(out_weights),
|
| 71 |
-
"converted_text_tensors": converted_text,
|
| 72 |
-
"converted_projection_tensors": converted_proj,
|
| 73 |
-
}
|
| 74 |
-
|
| 75 |
-
|
| 76 |
-
def convert_adapter_config(src_dir: Path, dst_dir: Path) -> None:
|
| 77 |
-
config = json.loads((src_dir / "adapter_config.json").read_text())
|
| 78 |
-
config["target_modules"] = "(.*(model.text_model).*(Wo|Wqkv|Wi).*$|.*(custom_text_proj).*$)"
|
| 79 |
-
(dst_dir / "adapter_config.json").write_text(json.dumps(config, indent=2) + "\n")
|
| 80 |
-
|
| 81 |
-
|
| 82 |
-
def main() -> None:
|
| 83 |
-
args = parse_args()
|
| 84 |
-
ensure_new_dir(args.output_dir)
|
| 85 |
-
copy_support_files(args.input_dir, args.output_dir)
|
| 86 |
-
|
| 87 |
-
weight_info = convert_adapter_weights(args.input_dir, args.output_dir)
|
| 88 |
-
convert_adapter_config(args.input_dir, args.output_dir)
|
| 89 |
-
|
| 90 |
-
build_info = {
|
| 91 |
-
"description": "Legacy paultltc/colmodernvbert adapter converted to the official ModernVBERT/colmodernvbert key layout.",
|
| 92 |
-
"input_dir": str(args.input_dir),
|
| 93 |
-
"output_dir": str(args.output_dir),
|
| 94 |
-
**weight_info,
|
| 95 |
-
"key_mapping": {
|
| 96 |
-
TEXT_PREFIX_IN: TEXT_PREFIX_OUT,
|
| 97 |
-
PROJ_PREFIX_IN: PROJ_PREFIX_OUT,
|
| 98 |
-
},
|
| 99 |
-
}
|
| 100 |
-
(args.output_dir / "BUILD_INFO.json").write_text(json.dumps(build_info, indent=2) + "\n")
|
| 101 |
-
|
| 102 |
-
print(f"Wrote {args.output_dir}")
|
| 103 |
-
print(f"Converted {weight_info['output_tensor_count']} adapter tensors")
|
| 104 |
-
|
| 105 |
-
|
| 106 |
-
if __name__ == "__main__":
|
| 107 |
-
main()
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|