llama3-8b-quantized

Uploaded model

• Developed by: tykiww
• License: apache-2.0
• Finetuned from model : unsloth/llama-3-8b-bnb-4bit

This llama model was trained 2x faster with Unsloth and Huggingface's TRL library.

<img src="https://raw.githubusercontent.com/unslothai/unsloth/main/images/unsloth%20made%20with%20love.png" width="200"/>

---

Setting up and testing own Endpoint Handler

Sources:

• https://www.philschmid.de/custom-inference-handler
• https://discuss.huggingface.co/t/model-wont-load-on-custom-inference-endpoint/91780
• https://huggingface.co/docs/inference-endpoints/guides/custom_handler

Setup Environment

Install necessary packages to set up and test endpoint handler.

# install git-lfs to interact with the repository
sudo apt-get update
sudo apt-get install git-lfs
# install transformers (not needed for inference since it is installed by default in the container)
pip install transformers[sklearn,sentencepiece,audio,vision]

Clone model weights of interest.

git lfs install
git clone https://huggingface.co/tykiww/llama3-8b-quantized

Login to huggingface

# setup cli with token
huggingface-cli login
git config --global credential.helper store

Confirm login in case you are unsure.

huggingface-cli whoami

Navigate to repo and create a handler.py file

cd llama3-8b-bnb-4bit-lora #&& touch handler.py

Create a requirements.txt file with the following items

huggingface_hub
unsloth[colab-new] @ git+https://github.com/unslothai/unsloth.git
xformers
trl<0.9.0 
peft==0.11.1
bitsandbytes
transformers==4.41.2 # must use /:

Must have a GPU compatible with Unsloth.