`AttributeError: 'ModuleDict' object has no attribute 'get_input_embeddings'` when finetuning
#7
by
matorus
- opened
When trying to finetune using Mosaic's composer train.py (with a more or less default finetuning yaml) I get: AttributeError: 'ModuleDict' object has no attribute 'get_input_embeddings'
Any ideas how to circumvent the issue?
╭─────────────────────────────── Traceback (most recent call last) ────────────────────────────────╮
│ /home/matthias/src/replit-finetune/../llm-foundry/scripts/train/train.py:326 in <module> │
│ │
│ 323 │ │ yaml_cfg = om.load(f) │
│ 324 │ cli_cfg = om.from_cli(args_list) │
│ 325 │ cfg = om.merge(yaml_cfg, cli_cfg) │
│ ❱ 326 │ main(cfg) │
│ 327 │
│ │
│ /home/matthias/src/replit-finetune/../llm-foundry/scripts/train/train.py:215 in main │
│ │
│ 212 │ │ │ │ cfg.model, cfg.lora, tokenizer) │
│ 213 │ │ │ print_trainable_parameters(model) # should not be 100% │
│ 214 │ │ else: # standard model │
│ ❱ 215 │ │ │ model = build_composer_model(cfg.model, tokenizer) │
│ 216 │ cfg.n_params = sum(p.numel() for p in model.parameters()) │
│ 217 │ print(f'{cfg.n_params=:.2e}') │
│ 218 │
│ │
│ /home/matthias/src/replit-finetune/../llm-foundry/scripts/train/train.py:70 in │
│ build_composer_model │
│ │
│ 67 │ if model_cfg.name not in COMPOSER_MODEL_REGISTRY: │
│ 68 │ │ raise ValueError( │
│ 69 │ │ │ f'Not sure how to build model with name={model_cfg.name}') │
│ ❱ 70 │ return COMPOSER_MODEL_REGISTRY[model_cfg.name](model_cfg, tokenizer) │
│ 71 │
│ 72 │
│ 73 def build_composer_peft_model( │
│ │
│ /home/matthias/src/llm-foundry/llmfoundry/models/hf/hf_causal_lm.py:181 in __init__ │
│ │
│ 178 │ │ │ │ f'om_model_config must be either a DictConfig, PeftModel, or PreTrainedM │
│ 179 │ │ │ ) │
│ 180 │ │ │
│ ❱ 181 │ │ composer_model = super().__init__(model=model, │
│ 182 │ │ │ │ │ │ │ │ │ │ shift_labels=True, │
│ 183 │ │ │ │ │ │ │ │ │ │ tokenizer=tokenizer, │
│ 184 │ │ │ │ │ │ │ │ │ │ metrics=train_metrics, │
│ │
│ /home/matthias/src/llm-foundry/llmfoundry/models/hf/model_wrapper.py:65 in __init__ │
│ │
│ 62 │ │ │
│ 63 │ │ # Note: We need to add the FSDP related attributes to the model AFTER the super │
│ 64 │ │ # so that the (possible) embedding resizing doesn't destroy them │
│ ❱ 65 │ │ prepare_hf_model_for_fsdp(self.model, init_device) │
│ 66 │ │ │
│ 67 │ │ # This provides support for meta initialization when using FSDP │
│ 68 │ │ self.model.param_init_fn = lambda module: self.model._init_weights( │
│ │
│ /home/matthias/src/llm-foundry/llmfoundry/models/hf/hf_fsdp.py:118 in prepare_hf_model_for_fsdp │
│ │
│ 115 │ else: │
│ 116 │ │ # many common decoder-only model do not set the flag │
│ 117 │ │ # model.config.is_decoder, so we can't trust it │
│ ❱ 118 │ │ prepare_hf_causal_lm_model_for_fsdp(model, init_device) │
│ 119 │
│ 120 │
│ 121 def prepare_hf_causal_lm_model_for_fsdp(model: PreTrainedModel, │
│ │
│ /home/matthias/src/llm-foundry/llmfoundry/models/hf/hf_fsdp.py:136 in │
│ prepare_hf_causal_lm_model_for_fsdp │
│ │
│ 133 │ lm_head = model.get_output_embeddings() │
│ 134 │ # some models (OPT) implement .get_input_embeddings for the causal subclass │
│ 135 │ # but all of them implement it for the base model │
│ ❱ 136 │ tied_embeddings = causal_base_model.get_input_embeddings() # type: ignore │
│ 137 │ modules = { │
│ 138 │ │ 'base_model': causal_base_model, │
│ 139 │ │ 'model_block': model_block, │
│ │
│ /home/matthias/src/replit-finetune/env/lib/python3.10/site-packages/torch/nn/modules/module.py:1 │
│ 614 in __getattr__ │
│ │
│ 1611 │ │ │ modules = self.__dict__['_modules'] │
│ 1612 │ │ │ if name in modules: │
│ 1613 │ │ │ │ return modules[name] │
│ ❱ 1614 │ │ raise AttributeError("'{}' object has no attribute '{}'".format( │
│ 1615 │ │ │ type(self).__name__, name)) │
│ 1616 │ │
│ 1617 │ def __setattr__(self, name: str, value: Union[Tensor, 'Module']) -> None: │
╰──────────────────────────────────────────────────────────────────────────────────────────────────╯
AttributeError: 'ModuleDict' object has no attribute 'get_input_embeddings'