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

This commit is contained in:
John Smith 2023-04-22 16:35:37 +08:00
parent eb442494d1 51bf103269
commit 4e42965c0d
1 changed files with 21 additions and 0 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

21
README.md
View File

@ -42,6 +42,13 @@ It's fast on a 3070 Ti mobile. Uses 5-6 GB of GPU RAM.
* Removed triton, flash-atten from requirements.txt for compatibility * Removed triton, flash-atten from requirements.txt for compatibility
* Removed bitsandbytes from requirements * Removed bitsandbytes from requirements
* Added pip installable branch based on winglian's PR * Added pip installable branch based on winglian's PR
* Added cuda backend quant attention and fused mlp from GPTQ_For_Llama.
* Added lora patch for GPTQ_For_Llama triton backend.
```
from monkeypatch.gptq_for_llala_lora_monkey_patch import inject_lora_layers
inject_lora_layers(model, lora_path, device, dtype)
```
# Requirements # Requirements
gptq-for-llama <br> gptq-for-llama <br>
@ -133,3 +140,17 @@ pip install xformers
from monkeypatch.llama_attn_hijack_xformers import hijack_llama_attention from monkeypatch.llama_attn_hijack_xformers import hijack_llama_attention
hijack_llama_attention() hijack_llama_attention()
``` ```
# Quant Attention and MLP Patch
Note: Currently does not support peft lora, but can use inject_lora_layers to load simple lora with only q_proj and v_proj.<br>
Usage:
```
from model_attn_mlp_patch import make_quant_attn, make_fused_mlp, inject_lora_layers
make_quant_attn(model)
make_fused_mlp(model)
# Lora
inject_lora_layers(model, lora_path)
```