LLaMA-Factory/scripts/loftq_init.py

83 lines
3.1 KiB
Python
Raw Normal View History

2023-12-14 13:53:56 +00:00
# coding=utf-8
# Initializes LoRA weights with LoRA-fine-tuning-aware Quantization (LoftQ)
# Usage: python loftq_init.py --model_name_or_path path_to_model --save_dir output_dir
# Inspired by: https://github.com/huggingface/peft/blob/main/examples/loftq_finetuning/quantize_save_load.py
import os
2024-01-20 12:15:56 +00:00
from typing import TYPE_CHECKING, Optional
2023-12-14 13:53:56 +00:00
import fire
import torch
import torch.nn as nn
from peft import LoftQConfig, LoraConfig, TaskType, get_peft_model
2024-01-20 12:15:56 +00:00
from transformers import AutoModelForCausalLM, AutoTokenizer
2023-12-14 13:53:56 +00:00
2024-01-20 11:58:04 +00:00
if TYPE_CHECKING:
from transformers import PreTrainedModel
2023-12-14 13:53:56 +00:00
class Shell(nn.Module):
def __init__(self, weight: torch.Tensor, bias: Optional[torch.Tensor] = None):
super().__init__()
self.weight = nn.Parameter(weight, requires_grad=False)
if bias is not None:
self.bias = nn.Parameter(bias, requires_grad=False)
def unwrap_model(model: nn.Module, pattern=".base_layer") -> None:
2024-02-14 18:27:36 +00:00
for name in {k.split(pattern)[0] for k, _ in model.named_modules() if pattern in k}:
2023-12-14 13:53:56 +00:00
parent_name = ".".join(name.split(".")[:-1])
child_name = name.split(".")[-1]
parent_module = model.get_submodule(parent_name)
child_module = getattr(parent_module, child_name)
base_layer = getattr(child_module, "base_layer")
weight = getattr(base_layer, "weight", None)
bias = getattr(base_layer, "bias", None)
setattr(parent_module, child_name, Shell(weight, bias))
2024-01-20 12:15:56 +00:00
2023-12-14 13:53:56 +00:00
print("Model unwrapped.")
def quantize_loftq(
model_name_or_path: str,
save_dir: str,
loftq_bits: Optional[int] = 4,
loftq_iter: Optional[int] = 1,
lora_alpha: Optional[int] = None,
lora_rank: Optional[int] = 16,
2024-01-20 11:58:04 +00:00
lora_target: Optional[str] = "q_proj,v_proj",
save_safetensors: Optional[bool] = False,
2023-12-14 13:53:56 +00:00
):
tokenizer = AutoTokenizer.from_pretrained(model_name_or_path, trust_remote_code=True)
model = AutoModelForCausalLM.from_pretrained(model_name_or_path, trust_remote_code=True, torch_dtype="auto")
loftq_config = LoftQConfig(loftq_bits=loftq_bits, loftq_iter=loftq_iter)
lora_config = LoraConfig(
task_type=TaskType.CAUSAL_LM,
inference_mode=True,
r=lora_rank,
lora_alpha=lora_alpha if lora_alpha is not None else lora_rank * 2,
lora_dropout=0.1,
target_modules=[name.strip() for name in lora_target.split(",")],
init_lora_weights="loftq",
2024-01-20 12:15:56 +00:00
loftq_config=loftq_config,
2023-12-14 13:53:56 +00:00
)
# Init LoftQ model
lora_model = get_peft_model(model, lora_config)
2024-01-20 11:58:04 +00:00
base_model: "PreTrainedModel" = lora_model.get_base_model()
2023-12-14 13:53:56 +00:00
# Save LoftQ model
setattr(lora_model.base_model.peft_config["default"], "base_model_name_or_path", save_dir)
setattr(lora_model.base_model.peft_config["default"], "init_lora_weights", True)
2024-01-20 11:58:04 +00:00
lora_model.save_pretrained(os.path.join(save_dir, "adapters"), safe_serialization=save_safetensors)
2023-12-14 13:53:56 +00:00
# Save base model
unwrap_model(base_model)
2024-01-20 11:58:04 +00:00
base_model.save_pretrained(save_dir, safe_serialization=save_safetensors)
2023-12-14 13:53:56 +00:00
tokenizer.save_pretrained(save_dir)
if __name__ == "__main__":
fire.Fire(quantize_loftq)