Checkpoint loading issue: ValueError

[andyhandsom6]

Hi! Thanks for the amazing work your team has done. I have a question when evaluating the fine-tuned model.
When loading the model checkpoint(automatically saved after finetuning on image stage), there's a ValueError:

Traceback (most recent call last):
File "run.py", line 138, in
main()
File "run.py", line 77, in main
run_exp(**vars(args))
File "run.py", line 114, in run_exp
evaluate_agent(config, dataset_split)
File "/home/lianqi/LongVU/VLN-CE/vlnce_baselines/nonlearning_agents.py", line 281, in evaluate_agent
agent = LMVidAgnet_special_token_seperator_multi_step_simplify_four_view(save_path)
File "/home/lianqi/LongVU/VLN-CE/vlnce_baselines/nonlearning_agents.py", line 2294, in init
self.tokenizer, self.model, self.image_processor, self.context_len = load_pretrained_model(self.args.model_path, None, self.model_name, devic
e_map=None)
File "/home/lianqi/LongVU/VLN-CE/longvu/builder.py", line 231, in load_pretrained_model
model.resize_token_embeddings(len(tokenizer))
File "/home/lianqi/.conda/envs/longvu_inf/lib/python3.8/site-packages/transformers/modeling_utils.py", line 1938, in resize_token_embeddings
model_embeds = self._resize_token_embeddings(new_num_tokens, pad_to_multiple_of)
File "/home/lianqi/.conda/envs/longvu_inf/lib/python3.8/site-packages/transformers/modeling_utils.py", line 1956, in _resize_token_embeddings
new_embeddings = self._get_resized_embeddings(old_embeddings, new_num_tokens, pad_to_multiple_of)
File "/home/lianqi/.conda/envs/longvu_inf/lib/python3.8/site-packages/transformers/modeling_utils.py", line 2052, in _get_resized_embeddings
old_num_tokens, old_embedding_dim = old_embeddings.weight.size()
ValueError: not enough values to unpack (expected 2, got 1)

I see similar issue raised by @mfarre before: #11 , but it still has the same problem after I switch to torchrun.
Any suggestions will be appreciated. Thanks in advance.

[andyhandsom6]

Solved.
#21 (comment)