tocmo0nlord/axolotl
1
0
Fork 0
Code Issues Pull Requests Actions 2 Packages Projects Releases Wiki Activity

qlora w flash attention fixes (#333)

Browse Source
This commit is contained in:
Wing Lian
2023-08-01 23:26:16 -04:00
committed by GitHub
parent db2a3586f3
commit 77085ea24e
1 changed files with 8 additions and 0 deletions
Download Patch File Download Diff File Expand all files Collapse all files

8
src/axolotl/utils/models.py
View File

@@ -407,6 +407,14 @@ def load_llama_adapter(model, cfg):
else: else:
model = get_peft_model(model, peft_config) model = get_peft_model(model, peft_config)
if cfg.flash_attention:
for name, module in model.named_modules():
if "norm" in name:
module.to(torch.float16)
if "lm_head" in name or "embed_tokens" in name:
if hasattr(module, "weight"):
module.to(torch.float16)
model.print_trainable_parameters() model.print_trainable_parameters()
return model, peft_config return model, peft_config