在使用lora对模型进行微调后启动模型报错

[itstamen]

System Info / 系統信息

使用GLM-4的finetune_demo对glm-4-9b-chat模型进行微调，由于我们机器是macbook apple m3，性能不佳，也没有GPU，因此对
因此将configs/lora.yaml的训练步数及保存点参数做了以下的调整：
save_steps:10 #原值为500 max_steps:30 #原值为3000
而后通过示例的如下命令进行微调：
python finetune.py data/AdvertiseGen/ THUDM/glm-4-9b-chat configs/lora.yaml
微后后产生了3个保存点：

而后使用微调后的模型：
python inference.py out/checkpoint-30
这时报了以下的错误：
KeyError: 'base_model.model.transformer.encoder.layers.6.input_layernorm.weight'
输出的错误信息为：
'NoneType' object has no attribute 'cadam32bit_grad_fp32'
Some parameters are on the meta device because they were offloaded to the disk.
╭────────────────────────────────────────────────────────── Traceback (most recent call last) ───────────────────────────────────────────────────────────╮
│ /Users/itstamen/Documents/workspace/ai/GLM-4/finetune_demo/inference.py:133 in main │
│ │
│ 130 │ │ "repetition_penalty": 1.2, │
│ 131 │ │ "eos_token_id": model.config.eos_token_id, │
│ 132 │ } │
│ ❱ 133 │ outputs = model.generate(**inputs, **generate_kwargs) │
│ 134 │ response = tokenizer.decode( │
│ 135 │ │ outputs[0][len(inputs["input_ids"][0]) :], skip_special_tokens=True │
│ 136 │ ).strip() │
│ │
│ /Users/itstamen/Documents/workspace/ai/GLM-4/.venv/lib/python3.10/site-packages/peft/peft_model.py:1838 in generate │
│ │
│ 1835 │ │ │ if not peft_config.is_prompt_learning: │
│ 1836 │ │ │ │ with self._enable_peft_forward_hooks(*args, **kwargs): │
│ 1837 │ │ │ │ │ kwargs = {k: v for k, v in kwargs.items() if k not in self.special_p │
│ ❱ 1838 │ │ │ │ │ outputs = self.base_model.generate(*args, **kwargs) │
│ 1839 │ │ │ else: │
│ 1840 │ │ │ │ outputs = self.base_model.generate(**kwargs) │
│ 1841 │ │ except: │
│ │
│ /Users/itstamen/Documents/workspace/ai/GLM-4/.venv/lib/python3.10/site-packages/torch/utils/_contextlib.py:116 in decorate_context │
│ │
│ 113 │ @functools.wraps(func) │
│ 114 │ def decorate_context(*args, **kwargs): │
│ 115 │ │ with ctx_factory(): │
│ ❱ 116 │ │ │ return func(*args, **kwargs) │
│ 117 │ │
│ 118 │ return decorate_context │
│ 119 │
│ │
│ /Users/itstamen/Documents/workspace/ai/GLM-4/.venv/lib/python3.10/site-packages/transformers/generation/utils.py:2255 in generate │
│ │
│ 2252 │ │ │ ) │
│ 2253 │ │ │ │
│ 2254 │ │ │ # 12. run sample (it degenerates to greedy search when `generation_config.do │
│ ❱ 2255 │ │ │ result = self._sample( │
│ 2256 │ │ │ │ input_ids, │
│ 2257 │ │ │ │ logits_processor=prepared_logits_processor, │
│ 2258 │ │ │ │ stopping_criteria=prepared_stopping_criteria, │
│ │
│ /Users/itstamen/Documents/workspace/ai/GLM-4/.venv/lib/python3.10/site-packages/transformers/generation/utils.py:3254 in _sample │
│ │
│ 3251 │ │ │ model_inputs.update({"output_hidden_states": output_hidden_states} if output │
│ 3252 │ │ │ │
│ 3253 │ │ │ if is_prefill: │
│ ❱ 3254 │ │ │ │ outputs = self(**model_inputs, return_dict=True) │
│ 3255 │ │ │ │ is_prefill = False │
│ 3256 │ │ │ else: │
│ 3257 │ │ │ │ outputs = model_forward(**model_inputs, return_dict=True) │
│ │
│ /Users/itstamen/Documents/workspace/ai/GLM-4/.venv/lib/python3.10/site-packages/torch/nn/modules/module.py:1736 in _wrapped_call_impl │
│ │
│ 1733 │ │ if self._compiled_call_impl is not None: │
│ 1734 │ │ │ return self._compiled_call_impl(*args, **kwargs) # type: ignore[misc] │
│ 1735 │ │ else: │
│ ❱ 1736 │ │ │ return self._call_impl(*args, **kwargs) │
│ 1737 │ │
│ 1738 │ # torchrec tests the code consistency with the following code │
│ 1739 │ # fmt: off │
│ │
│ /Users/itstamen/Documents/workspace/ai/GLM-4/.venv/lib/python3.10/site-packages/torch/nn/modules/module.py:1747 in _call_impl │
│ │
│ 1744 │ │ if not (self._backward_hooks or self._backward_pre_hooks or self._forward_hooks │
│ 1745 │ │ │ │ or _global_backward_pre_hooks or _global_backward_hooks │
│ 1746 │ │ │ │ or _global_forward_hooks or _global_forward_pre_hooks): │
│ ❱ 1747 │ │ │ return forward_call(*args, **kwargs) │
│ 1748 │ │ │
│ 1749 │ │ result = None │
│ 1750 │ │ called_always_called_hooks = set() │
│ │
│ /Users/itstamen/Documents/workspace/ai/GLM-4/.venv/lib/python3.10/site-packages/accelerate/hooks.py:170 in new_forward │
│ │
│ 167 │ │ │ with torch.no_grad(): │
│ 168 │ │ │ │ output = module._old_forward(*args, **kwargs) │
│ 169 │ │ else: │
│ ❱ 170 │ │ │ output = module._old_forward(*args, **kwargs) │
│ 171 │ │ return module._hf_hook.post_forward(module, output) │
│ 172 │ │
│ 173 │ # Overriding a GraphModuleImpl forward freezes the forward call and later modificati │
│ │
│ /Users/itstamen/.cache/huggingface/modules/transformers_modules/glm-4-9b-chat/modeling_chatglm.py:994 in forward │
│ │
│ 991 │ │ use_cache = use_cache if use_cache is not None else self.config.use_cache │
│ 992 │ │ return_dict = return_dict if return_dict is not None else self.config.use_return │
│ 993 │ │ │
│ ❱ 994 │ │ transformer_outputs = self.transformer( │
│ 995 │ │ │ input_ids=input_ids, │
│ 996 │ │ │ position_ids=position_ids, │
│ 997 │ │ │ attention_mask=attention_mask, │
│ │
│ /Users/itstamen/Documents/workspace/ai/GLM-4/.venv/lib/python3.10/site-packages/torch/nn/modules/module.py:1736 in _wrapped_call_impl │
│ │
│ 1733 │ │ if self._compiled_call_impl is not None: │
│ 1734 │ │ │ return self._compiled_call_impl(*args, **kwargs) # type: ignore[misc] │
│ 1735 │ │ else: │
│ ❱ 1736 │ │ │ return self._call_impl(*args, **kwargs) │
│ 1737 │ │
│ 1738 │ # torchrec tests the code consistency with the following code │
│ 1739 │ # fmt: off │
│ │
│ /Users/itstamen/Documents/workspace/ai/GLM-4/.venv/lib/python3.10/site-packages/torch/nn/modules/module.py:1747 in _call_impl │
│ │
│ 1744 │ │ if not (self._backward_hooks or self._backward_pre_hooks or self._forward_hooks │
│ 1745 │ │ │ │ or _global_backward_pre_hooks or _global_backward_hooks │
│ 1746 │ │ │ │ or _global_forward_hooks or global_forward_pre_hooks): │
│ ❱ 1747 │ │ │ return forward_call(*args, **kwargs) │
│ 1748 │ │ │
│ 1749 │ │ result = None │
│ 1750 │ │ called_always_called_hooks = set() │
│ │
│ /Users/itstamen/.cache/huggingface/modules/transformers_modules/glm-4-9b-chat/modeling_chatglm.py:892 in forward │
│ │
│ 889 │ │ │ rotary_pos_emb = rotary_pos_emb[None, :seq_length] │
│ 890 │ │ │
│ 891 │ │ # Run encoder. │
│ ❱ 892 │ │ hidden_states, presents, all_hidden_states, all_self_attentions = self.encoder( │
│ 893 │ │ │ inputs_embeds, full_attention_mask, rotary_pos_emb=rotary_pos_emb, │
│ 894 │ │ │ kv_caches=past_key_values, use_cache=use_cache, output_hidden_states=output │
│ 895 │ │ ) │
│ │
│ /Users/itstamen/Documents/workspace/ai/GLM-4/.venv/lib/python3.10/site-packages/torch/nn/modules/module.py:1736 in _wrapped_call_impl │
│ │
│ 1733 │ │ if self._compiled_call_impl is not None: │
│ 1734 │ │ │ return self._compiled_call_impl(*args, **kwargs) # type: ignore[misc] │
│ 1735 │ │ else: │
│ ❱ 1736 │ │ │ return self._call_impl(*args, **kwargs) │
│ 1737 │ │
│ 1738 │ # torchrec tests the code consistency with the following code │
│ 1739 │ # fmt: off │
│ │
│ /Users/itstamen/Documents/workspace/ai/GLM-4/.venv/lib/python3.10/site-packages/torch/nn/modules/module.py:1747 in _call_impl │
│ │
│ 1744 │ │ if not (self._backward_hooks or self._backward_pre_hooks or self._forward_hooks │
│ 1745 │ │ │ │ or _global_backward_pre_hooks or _global_backward_hooks │
│ 1746 │ │ │ │ or _global_forward_hooks or _global_forward_pre_hooks): │
│ ❱ 1747 │ │ │ return forward_call(*args, **kwargs) │
│ 1748 │ │ │
│ 1749 │ │ result = None │
│ 1750 │ │ called_always_called_hooks = set() │
│ │
│ /Users/itstamen/.cache/huggingface/modules/transformers_modules/glm-4-9b-chat/modeling_chatglm.py:722 in forward │
│ │
│ 719 │ │ │ │ │ use_reentrant=False │
│ 720 │ │ │ │ ) │
│ 721 │ │ │ else: │
│ ❱ 722 │ │ │ │ layer_ret = layer( │
│ 723 │ │ │ │ │ hidden_states, │
│ 724 │ │ │ │ │ attention_mask, │
│ 725 │ │ │ │ │ rotary_pos_emb, │
│ │
│ /Users/itstamen/Documents/workspace/ai/GLM-4/.venv/lib/python3.10/site-packages/torch/nn/modules/module.py:1736 in _wrapped_call_impl │
│ │
│ 1733 │ │ if self._compiled_call_impl is not None: │
│ 1734 │ │ │ return self._compiled_call_impl(*args, **kwargs) # type: ignore[misc] │
│ 1735 │ │ else: │
│ ❱ 1736 │ │ │ return self._call_impl(*args, **kwargs) │
│ 1737 │ │
│ 1738 │ # torchrec tests the code consistency with the following code │
│ 1739 │ # fmt: off │
│ │
│ /Users/itstamen/Documents/workspace/ai/GLM-4/.venv/lib/python3.10/site-packages/torch/nn/modules/module.py:1747 in _call_impl │
│ │
│ 1744 │ │ if not (self._backward_hooks or self._backward_pre_hooks or self._forward_hooks │
│ 1745 │ │ │ │ or _global_backward_pre_hooks or _global_backward_hooks │
│ 1746 │ │ │ │ or _global_forward_hooks or _global_forward_pre_hooks): │
│ ❱ 1747 │ │ │ return forward_call(*args, **kwargs) │
│ 1748 │ │ │
│ 1749 │ │ result = None │
│ 1750 │ │ called_always_called_hooks = set() │
│ │
│ /Users/itstamen/Documents/workspace/ai/GLM-4/.venv/lib/python3.10/site-packages/accelerate/hooks.py:170 in new_forward │
│ │
│ 167 │ │ │ with torch.no_grad(): │
│ 168 │ │ │ │ output = module._old_forward(*args, **kwargs) │
│ 169 │ │ else: │
│ ❱ 170 │ │ │ output = module._old_forward(*args, **kwargs) │
│ 171 │ │ return module._hf_hook.post_forward(module, output) │
│ 172 │ │
│ 173 │ # Overriding a GraphModuleImpl forward freezes the forward call and later modificati │
│ │
│ /Users/itstamen/.cache/huggingface/modules/transformers_modules/glm-4-9b-chat/modeling_chatglm.py:623 in forward │
│ │
│ 620 │ │ # hidden_states: [s, b, h] │
│ 621 │ │ │
│ 622 │ │ # Layer norm at the beginning of the transformer layer. │
│ ❱ 623 │ │ layernorm_output = self.input_layernorm(hidden_states) │
│ 624 │ │ # Self attention. │
│ 625 │ │ attention_output, kv_cache = self.self_attention( │
│ 626 │ │ │ layernorm_output, │
│ │
│ /Users/itstamen/Documents/workspace/ai/GLM-4/.venv/lib/python3.10/site-packages/torch/nn/modules/module.py:1736 in _wrapped_call_impl │
│ │
│ 1733 │ │ if self._compiled_call_impl is not None: │
│ 1734 │ │ │ return self._compiled_call_impl(*args, **kwargs) # type: ignore[misc] │
│ 1735 │ │ else: │
│ ❱ 1736 │ │ │ return self._call_impl(*args, **kwargs) │
│ 1737 │ │
│ 1738 │ # torchrec tests the code consistency with the following code │
│ 1739 │ # fmt: off │
│ │
│ /Users/itstamen/Documents/workspace/ai/GLM-4/.venv/lib/python3.10/site-packages/torch/nn/modules/module.py:1747 in _call_impl │
│ │
│ 1744 │ │ if not (self._backward_hooks or self._backward_pre_hooks or self._forward_hooks │
│ 1745 │ │ │ │ or _global_backward_pre_hooks or _global_backward_hooks │
│ 1746 │ │ │ │ or _global_forward_hooks or _global_forward_pre_hooks): │
│ ❱ 1747 │ │ │ return forward_call(*args, **kwargs) │
│ 1748 │ │ │
│ 1749 │ │ result = None │
│ 1750 │ │ called_always_called_hooks = set() │
│ │
│ /Users/itstamen/Documents/workspace/ai/GLM-4/.venv/lib/python3.10/site-packages/accelerate/hooks.py:165 in new_forward │
│ │
│ 162 │ module._hf_hook = hook │
│ 163 │ │
│ 164 │ def new_forward(module, *args, **kwargs): │
│ ❱ 165 │ │ args, kwargs = module._hf_hook.pre_forward(module, *args, **kwargs) │
│ 166 │ │ if module._hf_hook.no_grad: │
│ 167 │ │ │ with torch.no_grad(): │
│ 168 │ │ │ │ output = module._old_forward(*args, **kwargs) │
│ │
│ /Users/itstamen/Documents/workspace/ai/GLM-4/.venv/lib/python3.10/site-packages/accelerate/hooks.py:336 in pre_forward │
│ │
│ 333 │ │ │ │ remove_non_persistent=True, │
│ 334 │ │ │ ): │
│ 335 │ │ │ │ fp16_statistics = None │
│ ❱ 336 │ │ │ │ value = self.weights_map[name] │
│ 337 │ │ │ │ if "weight" in name and name.replace("weight", "SCB") in self.weights_ma │
│ 338 │ │ │ │ │ if value.dtype == torch.int8: │
│ 339 │ │ │ │ │ │ fp16_statistics = self.weights_map[name.replace("weight", "SCB") │
│ │
│ /Users/itstamen/Documents/workspace/ai/GLM-4/.venv/lib/python3.10/site-packages/accelerate/utils/offload.py:118 in getitem │
│ │
│ 115 │ │ self.prefix = prefix │
│ 116 │ │
│ 117 │ def getitem(self, key): │
│ ❱ 118 │ │ return self.dataset[f"{self.prefix}{key}"] │
│ 119 │ │
│ 120 │ def iter(self): │
│ 121 │ │ return iter([key for key in self.dataset if key.startswith(self.prefix)]) │
│ │
│ /Users/itstamen/Documents/workspace/ai/GLM-4/.venv/lib/python3.10/site-packages/accelerate/utils/offload.py:165 in getitem │
│ │
│ 162 │ │ # State dict gets priority │
│ 163 │ │ if key in self.state_dict: │
│ 164 │ │ │ return self.state_dict[key] │
│ ❱ 165 │ │ weight_info = self.index[key] │
│ 166 │ │ if weight_info.get("safetensors_file") is not None: │
│ 167 │ │ │ device = "cpu" if self.device is None else self.device │
│ 168 │ │ │ tensor = None │
╰────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────╯

请问是什么原因造成的，有什么解决思路和方法，谢谢！

Who can help? / 谁可以帮助到您？

@zRzRzRzRzRzRzR @wwewwt @Sengxian

Information / 问题信息

• The official example scripts / 官方的示例脚本
• My own modified scripts / 我自己修改的脚本和任务

Reproduction / 复现过程

python finetune.py data/AdvertiseGen/ THUDM/glm-4-9b-chat configs/lora.yaml
微后后产生了3个保存点：

而后使用微调后的模型：
python inference.py out/checkpoint-30

Expected behavior / 期待表现

希望可以正常启动微调后的模型。

[zhipuch]

是直接使用m3芯片跑吗？我们没有对它进行测试