the output only hava adapter_config.json,adapter_model.bin

[zxyscz]

i use lora to fintune model , the final output do not have checkpoint ,only hava adapter_config.json,adapter_model.bin

[shibing624]

it is right， only two files.

[pacman100]

Hello, that is the expected output as the final checkpoints with PEFT methods are tiny. You can load them via PeftModel.from_pretrained(base_model, peft_model_name_or_path)

[vincentmin]

Hi @pacman100 , could you explain why the code is structured such that you must provide the base_model? It seems to me that the base_model is already present in the adapter_config.json and thus we should be able to call PeftModel.from_pretrained(peft_model_name_or_path) and the base_model should be loaded internally. Ideally, we can even call AutoModel.from_pretrained(peft_model_name_or_path) and the user is oblivious as to whether the underlying weights are coming from a standard or PEFT model.

[clxyder]

Is there a way where one can save the base model and the adapter merged into one checkpoint?

I think I found my answer:

https://github.com/tloen/alpaca-lora/blob/main/export_state_dict_checkpoint.py

[pacman100]

@clxyder, with the latest main branch, you can simply do model = model.merge_and_unload() to get the base model with lora weights merged into it.

[clxyder]

@clxyder, with the latest main branch, you can simply do model = model.merge_and_unload() to get the base model with lora weights merged into it.

Thank you for the answer! Is there a way I can update or set the state_dict without reloading it again?

[gschoeni]

What would it take to support GPT2 models in merge_and_unload?

Getting the error: "GPT2 models are not supported for merging LORA layers"

[Opdoop]

@clxyder, with the latest main branch, you can simply do model = model.merge_and_unload() to get the base model with lora weights merged into it.

@pacman100 @younesbelkada Hi, thanks for this cool library! I'm new to peft. The model.merge_and_unload() method looks like magic to me. Could you give a quick introduction about model.merge_and_unload()？

In my view, LoRA adds new trainable parameters/layers and inserts these layers into the base model, that is the LoRA model has additional structures on top of the base model. And we can save the merge_and_unload model and reload it with base_model.from_pretrained(unloaded_model_path) interface. But where are the additional layers and parameters?

[younesbelkada]

Hi @Opdoop ,

Thanks for raising up this! In a nutshell this is a diagram that explains how the merging works under the hood:

So during training, you have these two independent modules (A & B) that are trainable. In that scenario, the output hidden states h can be computed as:

h = (Wx + BAx) + b

During training as you only want to update A & B you can't simplify the mathematical expression and run the computation in a single matrix multiplication.

However once A & B have been trained, you can "merge" these weights by simply adding them to W as follows`:

W_merged = (W + BA) 

since

W_merged x + b = (Wx + BAx) + b

The merged model is totally equivalent as the un-merged model, but this time you only need a single weight, W_merged

Let us know if anything else is unclear

[Opdoop]

@younesbelkada Cool! A big thanks to you! This explanation solved my confusion in a very simple and concrete way. Thanks for this beautiful diagram and math explanation! 🌹🌹🌹

[github-actions[bot]]

This issue has been automatically marked as stale because it has not had recent activity. If you think this still needs to be addressed please comment on this thread.