convert_safetensors.py

import collections
import numpy
import os
import torch
from safetensors.torch import serialize_file, load_file

import argparse

parser = argparse.ArgumentParser()
parser.add_argument("--input", type=str, help="Path to input pth model")
parser.add_argument(
    "--output",
    type=str,
    default="./converted.st",
    help="Path to output safetensors model",
)
args = parser.parse_args()


def rename_key(rename, name):
    for k, v in rename.items():
        if k in name:
            name = name.replace(k, v)
    return name


def convert_file(pt_filename: str, sf_filename: str, rename={}, transpose_names=[]):
    loaded: collections.OrderedDict = torch.load(pt_filename, map_location="cpu")
    if "state_dict" in loaded:
        loaded = loaded["state_dict"]

    kk = list(loaded.keys())
    version = 4
    for x in kk:
        if "ln_x" in x:
            version = max(5, version)
        if "gate.weight" in x:
            version = max(5.1, version)
        if int(version) == 5 and "att.time_decay" in x:
            if len(loaded[x].shape) > 1:
                if loaded[x].shape[1] > 1:
                    version = max(5.2, version)
        if "time_maa" in x:
            version = max(6, version)

    print(f"Model detected: v{version:.1f}")

    if version == 5.1:
        _, n_emb = loaded["emb.weight"].shape
        for k in kk:
            if "time_decay" in k or "time_faaaa" in k:
                # print(k, mm[k].shape)
                loaded[k] = (
                    loaded[k].unsqueeze(1).repeat(1, n_emb // loaded[k].shape[0])
                )

    with torch.no_grad():
        for k in kk:
            new_k = rename_key(rename, k).lower()
            v = loaded[k].half()
            del loaded[k]
            for transpose_name in transpose_names:
                if transpose_name in new_k:
                    dims = len(v.shape)
                    v = v.transpose(dims - 2, dims - 1)
            print(f"{new_k}\t{v.shape}\t{v.dtype}")
            loaded[new_k] = {
                "dtype": str(v.dtype).split(".")[-1],
                "shape": v.shape,
                "data": v.numpy().tobytes(),
            }

    dirname = os.path.dirname(sf_filename)
    os.makedirs(dirname, exist_ok=True)
    serialize_file(loaded, sf_filename, metadata={"format": "pt"})
    # reloaded = load_file(sf_filename)
    # for k in loaded:
    #     pt_tensor = torch.Tensor(
    #         numpy.frombuffer(
    #             bytearray(loaded[k]["data"]),
    #             dtype=getattr(numpy, loaded[k]["dtype"]),
    #         ).reshape(loaded[k]["shape"])
    #     )
    #     sf_tensor = reloaded[k]
    #     if not torch.equal(pt_tensor, sf_tensor):
    #         raise RuntimeError(f"The output tensors do not match for key {k}")


if __name__ == "__main__":
    convert_file(args.input, args.output,
                 rename={"time_faaaa": "time_first", "time_maa": "time_mix",
                         "lora_A": "lora.0", "lora_B": "lora.1"},
                 transpose_names=["time_mix_w1", "time_mix_w2", "time_decay_w1", "time_decay_w2", "lora.0"])
    print(f"Saved to {args.output}")