mirror of
https://github.com/ikawrakow/ik_llama.cpp.git
synced 2026-03-03 02:20:01 +00:00
c1a0e15377
* Add mtmd: the beginning * Add mtmd: mtmd.cpp compiles * Add mtmd: clip initialization compiles * Add mtmd: clip.cpp compiles * Add mtmd: builds successfully * Add CPU implementation for GGML_OP_GLU * Add CUDA implementation for GGML_OP_GLU * Add CPU implementation for GGML_OP_CONV_2D and GGML_OP_CONV_2D_DW * Add CUDA implementation for GGML_OP_CONV_2D and GGML_OP_CONV_2D_DW * Add mtmd: refresh CPU rope * Add mtmd: refresh CUDA rope * Add mtmd: add Qwen2-VL * Add mtmd: Qwen2.5-VL text seems to work with this change * Add mtmd: fix swiglu * Add mtmd: use LOG_TEE so generated tokens show up in terminal * Add mtmd: do not attempt to load a GPU backend if none are available * GLU, not GPU * Fix typo * Fix new/free mismatch * LOG stuff * Add mtmd: this fixes gibberish on second image --------- Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com>
48 lines
2.1 KiB
Python
48 lines
2.1 KiB
Python
import argparse
|
|
import os
|
|
import torch
|
|
from transformers import AutoModel, AutoTokenizer
|
|
|
|
ap = argparse.ArgumentParser()
|
|
ap.add_argument("-m", "--model", help="Path to MiniCPM-V model")
|
|
args = ap.parse_args()
|
|
|
|
# find the model part that includes the the multimodal projector weights
|
|
model = AutoModel.from_pretrained(args.model, trust_remote_code=True, local_files_only=True, torch_dtype=torch.bfloat16)
|
|
checkpoint = model.state_dict()
|
|
|
|
# get a list of mm tensor names
|
|
mm_tensors = [k for k, v in checkpoint.items() if k.startswith("resampler")]
|
|
|
|
# store these tensors in a new dictionary and torch.save them
|
|
projector = {name: checkpoint[name].float() for name in mm_tensors}
|
|
if 'resampler.proj' in projector.keys() and hasattr(model.llm.config,'scale_emb') is True:
|
|
projector['resampler.proj'] = projector['resampler.proj'] / model.llm.config.scale_emb
|
|
torch.save(projector, f"{args.model}/minicpmv.projector")
|
|
|
|
clip_tensors = [k for k, v in checkpoint.items() if k.startswith("vpm")]
|
|
if len(clip_tensors) > 0:
|
|
clip = {name.replace("vpm.", ""): checkpoint[name].float() for name in clip_tensors}
|
|
torch.save(clip, f"{args.model}/minicpmv.clip")
|
|
|
|
# added tokens should be removed to be able to convert Mistral models
|
|
if os.path.exists(f"{args.model}/added_tokens.json"):
|
|
with open(f"{args.model}/added_tokens.json", "w") as f:
|
|
f.write("{}\n")
|
|
|
|
config = model.llm.config
|
|
config.auto_map = {
|
|
"AutoConfig": "configuration_minicpm.MiniCPMConfig",
|
|
"AutoModel": "modeling_minicpm.MiniCPMModel",
|
|
"AutoModelForCausalLM": "modeling_minicpm.MiniCPMForCausalLM",
|
|
"AutoModelForSeq2SeqLM": "modeling_minicpm.MiniCPMForCausalLM",
|
|
"AutoModelForSequenceClassification": "modeling_minicpm.MiniCPMForSequenceClassification"
|
|
}
|
|
model.llm.save_pretrained(f"{args.model}/model")
|
|
tok = AutoTokenizer.from_pretrained(args.model, trust_remote_code=True)
|
|
tok.save_pretrained(f"{args.model}/model")
|
|
|
|
print("Done!")
|
|
print(f"Now you can convert {args.model} to a regular LLaMA GGUF file.")
|
|
print(f"Also, use {args.model}/minicpmv.projector to prepare a minicpmv-encoder.gguf file.")
|