ensure flash-attn fixes happen in both adapter/lora modes, and use torch_dtype
Browse files
src/axolotl/utils/models.py
CHANGED
|
@@ -331,6 +331,14 @@ def load_model(
|
|
| 331 |
model, use_gradient_checkpointing=cfg.gradient_checkpointing
|
| 332 |
)
|
| 333 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 334 |
model, lora_config = load_adapter(model, cfg, adapter)
|
| 335 |
|
| 336 |
if cfg.ddp and not load_in_8bit:
|
|
@@ -407,14 +415,6 @@ def load_llama_adapter(model, cfg):
|
|
| 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
|
|
|
|
| 331 |
model, use_gradient_checkpointing=cfg.gradient_checkpointing
|
| 332 |
)
|
| 333 |
|
| 334 |
+
if cfg.flash_attention:
|
| 335 |
+
for name, module in model.named_modules():
|
| 336 |
+
if "norm" in name:
|
| 337 |
+
module.to(torch_dtype)
|
| 338 |
+
if "lm_head" in name or "embed_tokens" in name:
|
| 339 |
+
if hasattr(module, "weight"):
|
| 340 |
+
module.to(torch_dtype)
|
| 341 |
+
|
| 342 |
model, lora_config = load_adapter(model, cfg, adapter)
|
| 343 |
|
| 344 |
if cfg.ddp and not load_in_8bit:
|
|
|
|
| 415 |
else:
|
| 416 |
model = get_peft_model(model, peft_config)
|
| 417 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 418 |
model.print_trainable_parameters()
|
| 419 |
|
| 420 |
return model, peft_config
|