Open JhonDan1999 opened 5 months ago
Hmm Gemma has tied embeddings - do you know if other tied embeddings and lm_head models have the same issue?
Thank you for your prompt response. I just experimented with the Gemma and "unsloth/llama-2-7b-bnb-4bit" models, and I'm encountering this issue specifically with Gemma.
Yep! It seems like the tied embeddings issue - apologies it was a super busy so didn't have time on this - apologies again!!
Was there ever a resolution to this issue? I'm now running into the same with the 9b variant of Gemma 2.
I ran into the same error while trying to save the merged model to HF, but adding safe_serialization=False can get around this issue. i.e.
model.push_to_hub_merged("hf_account/model_name", tokenizer, save_method = "merged_4bit_forced", token="hf_token"), safe_serialization=False)
I'm encountering a RuntimeError when attempting to save checkpoints during fine-tuning the "unsloth/gemma-2b-it-bnb-4bit" model. Below is a breakdown of my setup and the error encountered.
Model: unsloth/gemma-2b-it-bnb-4bit Relevant Training Configuration:
The training process halts with the following RuntimeError: