5d7d882ce6* Fix save_pretrained_merged for full-finetuned models save_pretrained_merged and push_to_hub_merged silently do nothing when the model is not a PeftModel (i.e. full finetuning without LoRA). merge_and_overwrite_lora returns None immediately for non-PeftModel, and unsloth_generic_save does not check the return value. Add a non-PeftModel branch in unsloth_generic_save that falls back to model.save_pretrained / model.push_to_hub. When save_method contains "16bit", cast weights to bfloat16 (or float16) via a state_dict copy to honor the user's intent without mutating the live model. The existing PeftModel (LoRA) code path is unchanged. * Forward create_pr and revision to tokenizer.push_to_hub The tokenizer push_to_hub call was missing create_pr and revision, which could cause the tokenizer to push to the wrong branch or bypass PR creation when the model push uses them. * Honor merged_16bit dtype contract for full-finetuned models Cast state_dict to bfloat16/float16 when save_method contains "16bit" to match the documented behavior of save_pretrained_merged. Also pass state_dict and save kwargs consistently to both save_pretrained and push_to_hub paths. * [pre-commit.ci] auto fixes from pre-commit.com hooks for more information, see https://pre-commit.ci * Address review feedback for PR #4755 - Simplify PeftModel isinstance check (PeftModelForCausalLM inherits from PeftModel) - Add is_main_process guard for distributed training - Forward variant to save_pretrained - Set tokenizer padding_side to "left" before saving (matches other save paths) * [pre-commit.ci] auto fixes from pre-commit.com hooks for more information, see https://pre-commit.ci --------- Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com> |
||
|---|---|---|
| .. | ||
| dataprep | ||
| kernels | ||
| models | ||
| optimizers | ||
| registry | ||
| utils | ||
| __init__.py | ||
| _auto_install.py | ||
| chat_templates.py | ||
| device_type.py | ||
| import_fixes.py | ||
| ollama_template_mappers.py | ||
| save.py | ||
| tokenizer_utils.py | ||
| trainer.py | ||