ilotoki_thu
/
alpaca_lora_4bit
2
0
Fork 0
Code Issues Pull Requests Projects Releases Wiki Activity

Merge branch 'main' of github.com:johnsmith0031/alpaca_lora_4bit 

merged
This commit is contained in:
John Smith 2023-03-28 21:14:35 +08:00
parent 211af574b6 ac07457473
commit 8a6c8661df
1 changed files with 2 additions and 15 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

17
README.md
View File

@ -1,30 +1,17 @@
# Alpaca Lora 4bit # Alpaca Lora 4bit
Made some adjust for the code in peft and gptq for llama, and make it possible for lora finetuning with a 4 bits base model. The same adjustment can be made for 2, 3 and 8 bits. Made some adjust for the code in peft and gptq for llama, and make it possible for lora finetuning with a 4 bits base model. The same adjustment can be made for 2, 3 and 8 bits.
<br>
* Install Manual by s4rduk4r: https://github.com/s4rduk4r/alpaca_lora_4bit_readme/blob/main/README.md (**NOTE:** don't use the install script, use the requirements.txt instead.)
<br>
* Install Manual by s4rduk4r: https://github.com/s4rduk4r/alpaca_lora_4bit_readme/blob/main/README.md (**NOTE:** don't use the install script, use the requirements.txt instead.)
* Also Remember to create a venv if you do not want the packages be overwritten. * Also Remember to create a venv if you do not want the packages be overwritten.
<br>
# Update Logs # Update Logs
* Resolved numerically unstable issue * Resolved numerically unstable issue
<br>
* Reconstruct fp16 matrix from 4bit data and call torch.matmul largely increased the inference speed. * Reconstruct fp16 matrix from 4bit data and call torch.matmul largely increased the inference speed.
<br>
* Added install script for windows and linux. * Added install script for windows and linux.
<br>
* Added Gradient Checkpointing. Now It can finetune 30b model 4bit on a single GPU with 24G VRAM with Gradient Checkpointing enabled. (finetune.py updated) (but would reduce training speed, so if having enough VRAM this option is not needed) * Added Gradient Checkpointing. Now It can finetune 30b model 4bit on a single GPU with 24G VRAM with Gradient Checkpointing enabled. (finetune.py updated) (but would reduce training speed, so if having enough VRAM this option is not needed)
<br>
* Added install manual by s4rduk4r * Added install manual by s4rduk4r
<br>
* Added pip install support by sterlind, preparing to merge changes upstream * Added pip install support by sterlind, preparing to merge changes upstream
<br> * Add V2 model support (with groupsize, both inference + finetune)
# Requirements # Requirements
gptq-for-llama: https://github.com/qwopqwop200/GPTQ-for-LLaMa<br> gptq-for-llama: https://github.com/qwopqwop200/GPTQ-for-LLaMa<br>