I trained a model using the tutorial below
In this tutorial, we'll fine-tune Llama 3 on a dataset of patient-doctor conversations. After merging, converting, and quantizing the model, it will be ready for private local use via the Jan application.
Here the Kaggle Code
they mentioned its generate a adapter_model.safetensors
at 160+MB
size but when I run the same script its giving the 2.27GB
.
But when I use the unsloth
It can able to produce the model in the 160+MB
size. But I don’t want to use the Unsloth. Here the Unsloth Fine-tune code. (You can find in the reply)
After changing the version of trl
& peft
the issue is resolved
%pip install -U peft==0.11.1
%pip install -U "trl<0.9.0"
Fore more reference
opened 01:37AM - 30 Jan 24 UTC
closed 03:04PM - 10 Mar 24 UTC
For the same peft config and same based model previously my saved lora adapter f… iles seemed to have changed in size by orders of magnitude:
1. Previously: 79 MB `adapter_model.bin`
trl '0.7.3.dev0'
peft '0.5.0'
2. Currently: 1GB an adapter safetensors file
trl '0.7.10'
peft '0.7.1
Can anyone please help me understand the difference in behavior?
Please check the below Kaggle discussion.
When using the meta-llama/Meta-Llama-3-8B-Instruct
or meta-llama/Meta-Llama-3.1-8B-Instruct
both are giving the 2.27GB
of adapter_model.safetensors
.
Here the package I used.
%%capture
%pip install -U transformers
%pip install -U datasets
%pip install -U accelerate
%pip install -U peft==0.11.1
%pip install -U "trl<0.9.0"
%pip install -U bitsandbytes
%pip install -U wandb
system
Closed
July 27, 2024, 12:54am
4
This topic was automatically closed 12 hours after the last reply. New replies are no longer allowed.