rename var and reformat
Browse files
src/axolotl/utils/models.py
CHANGED
@@ -355,7 +355,7 @@ def load_model(
|
|
355 |
if hasattr(module, "weight"):
|
356 |
module.to(torch.float32)
|
357 |
|
358 |
-
|
359 |
if not cfg.gptq and (
|
360 |
(cfg.adapter == "lora" and load_in_8bit)
|
361 |
or (cfg.adapter == "qlora" and cfg.load_in_4bit)
|
@@ -364,13 +364,11 @@ def load_model(
|
|
364 |
model = prepare_model_for_kbit_training(
|
365 |
model, use_gradient_checkpointing=cfg.gradient_checkpointing
|
366 |
)
|
367 |
-
|
368 |
|
369 |
# LlamaRMSNorm layers are in fp32 after kbit_training or full finetune, so we need to
|
370 |
# convert them back to fp16/bf16 for flash-attn compatibility.
|
371 |
-
if
|
372 |
-
cfg.flash_attention and cfg.is_llama_derived_model
|
373 |
-
):
|
374 |
for name, module in model.named_modules():
|
375 |
if "norm" in name:
|
376 |
module.to(cfg.torch_dtype)
|
|
|
355 |
if hasattr(module, "weight"):
|
356 |
module.to(torch.float32)
|
357 |
|
358 |
+
needs_fa2_dtype = not cfg.adapter
|
359 |
if not cfg.gptq and (
|
360 |
(cfg.adapter == "lora" and load_in_8bit)
|
361 |
or (cfg.adapter == "qlora" and cfg.load_in_4bit)
|
|
|
364 |
model = prepare_model_for_kbit_training(
|
365 |
model, use_gradient_checkpointing=cfg.gradient_checkpointing
|
366 |
)
|
367 |
+
needs_fa2_dtype = True
|
368 |
|
369 |
# LlamaRMSNorm layers are in fp32 after kbit_training or full finetune, so we need to
|
370 |
# convert them back to fp16/bf16 for flash-attn compatibility.
|
371 |
+
if needs_fa2_dtype and (cfg.flash_attention and cfg.is_llama_derived_model):
|
|
|
|
|
372 |
for name, module in model.named_modules():
|
373 |
if "norm" in name:
|
374 |
module.to(cfg.torch_dtype)
|