explodinggradients/Funtuner

Supervised instruction finetuning for LLM with HF trainer and Deepspeed

FunTuner

A no nonsense easy to configure model fine-tuning framework for GPT based models that can get the job done in a memory and time efficient manner.

☢️ Work in progress

Components

✅hydra configuration

✅Deepspeed support

✅8 bit training

✅LoRA using peft

✅Sequence bucketing

✅Inference

✅single
✅batch
❎stream

✅Supported Models

✅GPTNeoX - Redajajama, Pythia, etc
❎LLama
❎Falcon 

❎Flash attention

Train

• Using deepspeed

deepspeed funtuner/trainer.py

Inference

from funtuner.inference import Inference
model = Inference("shahules786/GPTNeo-125M-lora")
kwargs = {"temperature":0.1,
        "top_p":0.75,
        "top_k":5,
        "num_beams":2,
        "max_new_tokens":128,}

##single
output =model.generate("Which is a species of fish? Tope or Rope",**kwargs)

##batch
inputs = [["There was a tiger in the hidden"],["Which is a species of fish? Tope or Rope"]]
output = model.batch_generate(inputs,**kwargs)

Sampling

python funtuner/sampling.py --model_url shahules786/Redpajama-3B-CoT --dataset Dahoas/cot_gsm8k