llama

1
import argparse
2
import os
3
import torch
4
from transformers import AutoModel, AutoTokenizer
5

6
ap = argparse.ArgumentParser()
7
ap.add_argument("-m", "--model", help="Path to MiniCPM-V model")
8
args = ap.parse_args()
9

10
# find the model part that includes the the multimodal projector weights
11
model = AutoModel.from_pretrained(args.model, trust_remote_code=True, local_files_only=True)
12
checkpoint = model.state_dict()
13

14
# get a list of mm tensor names
15
mm_tensors = [k for k, v in checkpoint.items() if k.startswith("resampler")]
16

17
# store these tensors in a new dictionary and torch.save them
18
projector = {name: checkpoint[name].float() for name in mm_tensors}
19
torch.save(projector, f"{args.model}/minicpmv.projector")
20

21
clip_tensors = [k for k, v in checkpoint.items() if k.startswith("vpm")]
22
if len(clip_tensors) > 0:
23
    clip = {name.replace("vpm.", ""): checkpoint[name].float() for name in clip_tensors}
24
    torch.save(clip, f"{args.model}/minicpmv.clip")
25

26
    # added tokens should be removed to be able to convert Mistral models
27
    if os.path.exists(f"{args.model}/added_tokens.json"):
28
        with open(f"{args.model}/added_tokens.json", "w") as f:
29
            f.write("{}\n")
30

31
config = model.llm.config
32
config.auto_map = {
33
    "AutoConfig": "configuration_minicpm.MiniCPMConfig",
34
    "AutoModel": "modeling_minicpm.MiniCPMModel",
35
    "AutoModelForCausalLM": "modeling_minicpm.MiniCPMForCausalLM",
36
    "AutoModelForSeq2SeqLM": "modeling_minicpm.MiniCPMForCausalLM",
37
    "AutoModelForSequenceClassification": "modeling_minicpm.MiniCPMForSequenceClassification"
38
}
39
model.llm.save_pretrained(f"{args.model}/model")
40
tok = AutoTokenizer.from_pretrained(args.model, trust_remote_code=True)
41
tok.save_pretrained(f"{args.model}/model")
42

43
print("Done!")
44
print(f"Now you can convert {args.model} to a regular LLaMA GGUF file.")
45
print(f"Also, use {args.model}/minicpmv.projector to prepare a minicpmv-encoder.gguf file.")
46