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

Update README.md

This commit is contained in:
John Smith 2023-04-12 13:09:48 +08:00 committed by GitHub
parent e946f830d4
commit 17e6a1585f
No known key found for this signature in database
GPG Key ID: 4AEE18F83AFDEB23
1 changed files with 16 additions and 0 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

16
README.md
View File

@ -34,6 +34,8 @@ It's fast on a 3070 Ti mobile. Uses 5-6 GB of GPU RAM.
* Added Flash attention support. (Use --flash-attention)
* Added Triton backend to support model using groupsize and act-order. (Use --backend=triton)
* Added g_idx support in cuda backend (need recompile cuda kernel)
* Added xformers support
* Removed triton, flash-atten from requirements.txt for compatibility
# Requirements
gptq-for-llama <br>
@ -102,3 +104,17 @@ python server.py
It seems that we can apply a monkey patch for llama model. To use it, simply download the file from [MonkeyPatch](https://github.com/lm-sys/FastChat/blob/daa9c11080ceced2bd52c3e0027e4f64b1512683/fastchat/train/llama_flash_attn_monkey_patch.py). And also, flash-attention is needed, and currently do not support pytorch 2.0.
Just add --flash-attention to use it for finetuning.
# Xformers
* Install
```
pip install xformers
```
* Usage
```
from monkeypatch.llama_attn_hijack_xformers import hijack_llama_attention
hijack_llama_attention()
```