qlora w flash attention fixes (#333)
Browse files
src/axolotl/utils/models.py
CHANGED
|
@@ -407,6 +407,14 @@ def load_llama_adapter(model, cfg):
|
|
| 407 |
else:
|
| 408 |
model = get_peft_model(model, peft_config)
|
| 409 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 410 |
model.print_trainable_parameters()
|
| 411 |
|
| 412 |
return model, peft_config
|
|
|
|
| 407 |
else:
|
| 408 |
model = get_peft_model(model, peft_config)
|
| 409 |
|
| 410 |
+
if cfg.flash_attention:
|
| 411 |
+
for name, module in model.named_modules():
|
| 412 |
+
if "norm" in name:
|
| 413 |
+
module.to(torch.float16)
|
| 414 |
+
if "lm_head" in name or "embed_tokens" in name:
|
| 415 |
+
if hasattr(module, "weight"):
|
| 416 |
+
module.to(torch.float16)
|
| 417 |
+
|
| 418 |
model.print_trainable_parameters()
|
| 419 |
|
| 420 |
return model, peft_config
|