Skip to content

Latest commit

 

History

History
143 lines (91 loc) · 4.04 KB

convert_lit_models.md

File metadata and controls

143 lines (91 loc) · 4.04 KB
Raw

Converting LitGPT weights to Hugging Face Transformers

LitGPT weights need to be converted to a format that Hugging Face understands with a conversion script before our scripts can run.

We provide a helpful command to convert models LitGPT models back to their equivalent Hugging Face Transformers format:

litgpt convert_from_litgpt checkpoint_dir converted_dir

These paths are just placeholders, you will need to customize them based on which finetuning or pretraining command you ran and its configuration.

Loading converted LitGPT checkpoints into transformers

For example,

cp checkpoints/repo_id/config.json converted/config.json

Then, you can load the checkpoint file in a Python session as follows:

import torch
from transformers import AutoModel


state_dict = torch.load("output_dir/model.pth")
model = AutoModel.from_pretrained(
    "output_dir/", local_files_only=True, state_dict=state_dict
)

Alternatively, you can also load the model without copying the config.json file as follows:

model = AutoModel.from_pretrained("online_repo_id", state_dict=state_dict)

Merging LoRA weights

Please note that if you want to convert a model that has been finetuned using an adapter like LoRA, these weights should be merged to the checkpoint prior to converting.

litgpt merge_lora path/to/lora/checkpoint_dir


A finetuning and conversion tutorial

This section contains a reproducible example for finetuning a LitGPT model and converting it back into a HF transformer model.

  1. Download a model of interest:

For convenience, we first specify an environment variable (optional) to avoid copy and pasting the whole path:

export repo_id=TinyLlama/TinyLlama-1.1B-intermediate-step-1431k-3T

Instead of using TinyLlama, you can replace the repo_id target with any other model repository specifier that is currently supported by LitGPT. You can get a list of supported repository specifier by running litgpt/scripts/download.py without any additional arguments.

Then, we download the model we specified via $repo_id above:

litgpt download $repo_id
  1. Finetune the model:
export finetuned_dir=out/lit-finetuned-model

litgpt finetune_lora $repo_id \
   --out_dir $finetuned_dir \
   --train.epochs 1 \
   --data Alpaca
  1. Merge LoRA weights:

Note that this step only applies if the model was finetuned with lora.py above and not when full.py was used for finetuning.

litgpt merge_lora $finetuned_dir/final
  1. Convert the finetuning model back into a HF format:
litgpt convert_from_litgpt $finetuned_dir/final/ out/hf-tinyllama/converted
  1. Load the model into a transformers model:
import torch
from transformers import AutoModel

state_dict = torch.load('out/hf-tinyllama/converted/model.pth')
model = AutoModel.from_pretrained("TinyLlama/TinyLlama-1.1B-intermediate-step-1431k-3T", state_dict=state_dict)

 

Using the LM Evaluation Harness

To evaluate LitGPT models, use the integrated evaluation utilities based on Eleuther AI's LM Evaluation Harness. For more information, please see the evaluation documentation.

Alternatively, if you wish to use converted LitGPT models with the LM Evaluation Harness from Eleuther AI's GitHub repository, you can use the following steps.

  1. Follow the instructions above to load the model into a Hugging Face transformers model.

  2. Create a model.safetensor file:

model.save_pretrained("out/hf-tinyllama/converted/")
  1. Copy the tokenizer files into the model-containing directory:
cp checkpoints/$repo_id/tokenizer* out/hf-tinyllama/converted
  1. Run the evaluation harness, for example:
lm_eval --model hf \
    --model_args pretrained=out/hf-tinyllama/converted \
    --tasks "hellaswag,gsm8k,truthfulqa_mc2,mmlu,winogrande,arc_challenge" \
    --device "cuda:0" \
    --batch_size 4