带有embeding 同时训练的Lora 权重合并
from peftimportPeftModel from transformersimportAutoModelForCausalLM, AutoTokenizer base_model_path="/root/private_data/models/deepseek-ai/6epoch-merged"adapter_path="/root/private_data/output/sharedata/checkpoint-10500/"output_path="/root/private_data/models/deepseek-ai/merged_correct"# 加载基础模型model=AutoModelForCausalLM.from_pretrained(base_model_path,torch_dtype="bfloat16",device_map="cuda:0")# 加载 adaptermodel=PeftModel.from_pretrained(model, adapter_path)# 合并model=model.merge_and_unload()# 保存model.save_pretrained(output_path)tokenizer=AutoTokenizer.from_pretrained(base_model_path)tokenizer.save_pretrained(output_path)
错误的写法 合并后的权重的模型，再训练数的Loss 突然增加
### 合并llamafactory lora 到新的模型，### 注意 --export_dir这个路径，要新的路径，### 新训练的 checkpoint路径不要用原来checkpoint路径，否则会lora对不齐错误### 对应带有embedding 不合适 会丢失信息llamafactory-cliexport\--model_name_or_path/root/private_data/models/deepseek-ai/6epoch-merged\--adapter_name_or_path/root/private_data/output/binding_sft_short_part_004-005_lora64_epoch1/checkpoint-9000\--templatealpaca\--finetuning_typelora\--export_dir/root/private_data/models/deepseek-ai/6epoch-merged-lora64_data005\--export_size4\--export_deviceauto\--export_legacy_formatfalse
错误写法2
importtorchimportjsonimportosimportshutil from safetensors.torchimportload_file, save_file def merge_full_optimized(base_model_path: str, adapter_path: str, output_merged_path: str, device: str="cuda:0"): print(f"🚀 [START] 开始全量合并流程 (LoRA + Embedding)...")# 1. 加载 Adapter 权重并解析 LoRA 配置print(f"📥 加载 Adapter 权重...")adapter_weights=load_file(os.path.join(adapter_path,"adapter_model.safetensors"),device=device)with open(os.path.join(adapter_path,"adapter_config.json"),'r')as f: lora_config=json.load(f)r=lora_config["r"]alpha=lora_config["lora_alpha"]scaling=alpha / r print(f"✅ LoRA 配置读取成功: r={r}, alpha={alpha}, scaling={scaling}")# 2. 【修复】准备输出目录：先从基础模型复制配置文件os.makedirs(output_merged_path,exist_ok=True)print(f"📋 从基础模型复制配置文件...")forfileinos.listdir(base_model_path): src=os.path.join(base_model_path,file)ifos.path.isfile(src)and not any(file.endswith(ex)forexin['.safetensors','.bin','.pt']): shutil.copy2(src, os.path.join(output_merged_path,file))# adapter 目录的非权重文件一般不需要（adapter_config.json 不是合并模型的一部分）# 如果 adapter 目录有其他特殊文件，可以按需添加# 3. 核心合并逻辑：分片处理safetensors_files=sorted([fforfinos.listdir(base_model_path)iff.endswith('.safetensors')])weight_map={}new_num_tokens=Noneforst_fileinsafetensors_files: print(f"📦 处理分片: {st_file}...")src_path=os.path.join(base_model_path, st_file)dst_path=os.path.join(output_merged_path, st_file)state_dict=load_file(src_path,device=device)updated_keys=[]forkeyinlist(state_dict.keys()):# --- A. 处理 LoRA 合并 ---prefix="base_model.model."lora_a_key=f"{prefix}{key.replace('.weight', '.lora_A.weight')}"lora_b_key=f"{prefix}{key.replace('.weight', '.lora_B.weight')}"iflora_a_keyinadapter_weights and lora_b_keyinadapter_weights: W=state_dict[key].to(torch.bfloat16)A=adapter_weights[lora_a_key].to(torch.bfloat16)B=adapter_weights[lora_b_key].to(torch.bfloat16)state_dict[key]=W +(B @ A)* scaling updated_keys.append(key)# --- B. 处理 Embedding 和 LM_Head 替换 ---embed_save_key=f"{prefix}{key}.modules_to_save.default.weight"ifembed_save_keyinadapter_weights: state_dict[key]=adapter_weights[embed_save_key].to(state_dict[key].dtype)if"embed_tokens"inkey: new_num_tokens=state_dict[key].shape[0]print(f"💎 已替换模块: {key}")# 保存分片cpu_state_dict={k: v.contiguous().cpu()fork,vinstate_dict.items()}save_file(cpu_state_dict, dst_path)forkinstate_dict.keys(): weight_map[k]=st_file del state_dict, cpu_state_dict torch.cuda.empty_cache()# 4. 更新 config.jsonifnew_num_tokens: cfg_path=os.path.join(output_merged_path,"config.json")ifos.path.exists(cfg_path): with open(cfg_path,'r+')as f: config=json.load(f)config["vocab_size"]=new_num_tokens f.seek(0)json.dump(config, f,indent=2)f.truncate()print(f"📝 已更新 config.json: vocab_size = {new_num_tokens}")else: print(f"⚠️ 警告: 未找到 config.json，无法更新 vocab_size")# 5. 生成索引total_size=sum(os.path.getsize(os.path.join(output_merged_path, f))forfinsafetensors_files)index_path=os.path.join(output_merged_path,"model.safetensors.index.json")with open(index_path,'w')as f: json.dump({"metadata":{"total_size":total_size},"weight_map":weight_map}, f,indent=2)print(f"✨ [DONE] 全量合并成功！输出目录: {output_merged_path}")print(f"📁 请确认以下文件存在:")forfin["config.json","tokenizer.json","tokenizer_config.json","model.safetensors.index.json"]: p=os.path.join(output_merged_path, f)status="✅"ifos.path.exists(p)else"❌"print(f" {status} {f}")if__name__=="__main__":CONFIG={"base_model_path":"/root/private_data/models/deepseek-ai/6epoch-merged","adapter_path":"/root/private_data/output/sharedata/checkpoint-10500/","output_merged_path":"/root/private_data/models/deepseek-ai/6epoch-merged-sft","device":"cuda:0"}merge_full_optimized(**CONFIG)# 这是一个 Python 的语法技巧，叫做 "字典解包"（Dictionary Unpacking）。# 简单来说，CONFIG 的作用是将一个字典里的键值对，# "拆解"成一个个独立的关键字参数（Keyword Arguments）传给函数。# merge_full_optimized(**CONFIG)
