This repository contains a Colab notebook that allows you to run Large Language Models (LLM) models with just one click.
in colab gpu T4 (15Gvram) you can use :
13b model gguf quantized upto Q5_K_M(context up to 4K)
[Q5_K_M u can get about 10token/s]11b model gguf quantized upto Q5_K_M(context up to 8K)
[Q5_K_M => 16~18 token/s] [Q8_0 => 11 token/s]7b model gguf quantized upto Q8_0(context up to 16K if the model support it)
[Q8_0 u can get upto 18~19token/s]13b model GPTQ quantized
11b model GPTQ quantized
7b model GPTQ quantized
[upto 17token/S]7b model exllama2 quantized (context up to 16k if the model support it)
[with version 8bt u can get upto 23token/S]
most mistral based models goes with 8k context length if u want to use longer context u need to make sure the models support longer context.
if u want to run model higher then 13B (such as 20B,4x7b..) on colab u may need to reduce the offloaded gpu models to split the ram usage between gpu vram and system ram. (slower but it works ๐)
if u dont have quantized version , you can use full precision
7b
modeles with gptq notebook but make sure to use flags--load-in-4bit
or--load-in-8bit
its slower then quantized versions but works well,so if u have quantized verions it will be better.
if u want a creative answers from ur model then increase the temp(0.9~1.25) and decrease the minp(0.05~0.1) , but when u need strict and accurate answers decrease the temp(0.5~0.7) and increase the min p (0.1~0.2)
-Run gguf LLM models in TextGen-webui :
-Run GPTQ and Exl2 LLM models in TextGen-webui :
To get started with the LLM Model Runner, follow these steps:
-
Open the Colab notebook in Google Colab by clicking on the "Open in Colab" button at the top of the notebook.
-
Choose The model that you want from the list .
3.Choose quantization type:
- Run the Cell and Visit the Generated link(
https://***.gradio.live
) and start your Conversation with your favorite model !
- no Requirement just open Colab in Gpu mode
All the necessary dependencies will be automatically installed when you run the Colab notebook.
All Models license belongs to thier owners please follow the link of the Model repo and read its license if u need.
- TheBloke for their Quantized LLM models.
- text-generation-webui for their Great Ui