🧠 两小时从 0 训出你的 64M 大模型。Train your own 64M LLM from scratch in 2 hours.
import torch
from torch import optim, nn
# 定义Lora网络结构
class LoRA(nn.Module):
def __init__(self, in_features, out_features, rank, lora_alpha=None):
super().__init__()
self.rank = rank # LoRA的秩(rank),控制低秩矩阵的大小
scaling = (rank if lora_alpha is None else lora_alpha) / rank
self.register_buffer("scaling", torch.tensor(scaling, dtype=torch.float32))
self.A = nn.Linear(in_features, rank, bias=False) # 低秩矩阵A
self.B = nn.Linear(rank, out_features, bias=False) # 低秩矩阵B
# 矩阵A高斯初始化
self.A.weight.data.normal_(mean=0.0, std=0.02)
# 矩阵B全0初始化
self.B.weight.data.zero_()
def forward(self, x):
return self.B(self.A(x)) * self.scaling
def apply_lora(model, rank=16, lora_alpha=None):
for name, module in model.named_modules():
if isinstance(module, nn.Linear) and module.in_features == module.out_features:
lora = LoRA(module.in_features, module.out_features, rank=rank, lora_alpha=lora_alpha).to(model.device)
setattr(module, "lora", lora)
original_forward = module.forward
# 显式绑定
def forward_with_lora(x, layer1=original_forward, layer2=lora):
return layer1(x) + layer2(x)
module.forward = forward_with_lora
def load_lora(model, path):
state_dict = torch.load(path, map_location=model.device)
state_dict = {(k[7:] if k.startswith('module.') else k): v for k, v in state_dict.items()}
for name, module in model.named_modules():
if hasattr(module, 'lora'):
lora_state = {k.replace(f'{name}.lora.', ''): v for k, v in state_dict.items() if f'{name}.lora.' in k}
module.lora.load_state_dict(lora_state, strict=False)
def save_lora(model, path):
raw_model = getattr(model, '_orig_mod', model)
state_dict = {}
for name, module in raw_model.named_modules():
if hasattr(module, 'lora'):
clean_name = name[7:] if name.startswith("module.") else name
lora_state = {f'{clean_name}.lora.{k}': v.cpu().half() if k.endswith('.weight') else v.cpu().float()
for k, v in module.lora.state_dict().items()}
state_dict.update(lora_state)
torch.save(state_dict, path)
def merge_lora(model, lora_path, save_path):
load_lora(model, lora_path)
raw_model = getattr(model, '_orig_mod', model)
state_dict = {k: v.cpu().half() for k, v in raw_model.state_dict().items() if '.lora.' not in k}
for name, module in raw_model.named_modules():
if isinstance(module, nn.Linear) and '.lora.' not in name:
state_dict[f'{name}.weight'] = module.weight.data.clone().cpu().float()
if hasattr(module, 'lora'):
delta = (module.lora.B.weight.data.float() @ module.lora.A.weight.data.float()) * module.lora.scaling.float()
state_dict[f'{name}.weight'] += delta.cpu()
state_dict[f'{name}.weight'] = state_dict[f'{name}.weight'].half()
torch.save(state_dict, save_path)