423A35C7/LLaMA-Factory
1
0
Fork 0
mirror of https://github.com/hiyouga/LLaMA-Factory.git synced 2025-08-04 20:52:59 +08:00
Code Issues Packages Projects Releases Wiki Activity

update neftune logic

Browse Source 
Former-commit-id: 7de7174ce3ea804b2ba58560193cda25cbd675ef
This commit is contained in:
hiyouga 2023-10-22 17:42:13 +08:00
parent 2af83198c7
commit daeff710eb
1 changed files with 18 additions and 17 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

15
src/llmtuner/tuner/core/utils.py
View File

@ -56,10 +56,10 @@ def prepare_model_for_training(
logger.info("Upcasting weights in layernorm in float32.") logger.info("Upcasting weights in layernorm in float32.")
if finetuning_args.neft_alpha > 1e-6: if finetuning_args.neft_alpha > 1e-6:
input_embed: torch.nn.Embedding = model.get_input_embeddings() input_embed = model.get_input_embeddings()
if isinstance(input_embed, torch.nn.Embedding):
def noisy_forward(self: torch.nn.Embedding, x: torch.Tensor) -> torch.Tensor: def noisy_forward(self: torch.nn.Embedding, x: torch.Tensor) -> torch.Tensor:
embeddings = torch.nn.Embedding.forward(self, x) embeddings = input_embed.__class__.forward(self, x)
if self.training: if self.training:
dims = self.num_embeddings * self.embedding_dim dims = self.num_embeddings * self.embedding_dim
mag_norm = finetuning_args.neft_alpha / (dims ** 0.5) mag_norm = finetuning_args.neft_alpha / (dims ** 0.5)
@ -68,6 +68,8 @@ def prepare_model_for_training(
input_embed.forward = MethodType(noisy_forward, input_embed) input_embed.forward = MethodType(noisy_forward, input_embed)
logger.info("Using noisy embedding with alpha={:.2f}".format(finetuning_args.neft_alpha)) logger.info("Using noisy embedding with alpha={:.2f}".format(finetuning_args.neft_alpha))
else:
logger.warning("Input embeddings are not normal nn.Embedding, cannot transform into noisy embedding.")
if use_gradient_checkpointing: if use_gradient_checkpointing:
if hasattr(model, "enable_input_require_grads"): if hasattr(model, "enable_input_require_grads"):
@ -82,11 +84,10 @@ def prepare_model_for_training(
logger.info("Gradient checkpointing enabled.") logger.info("Gradient checkpointing enabled.")
if finetuning_args.finetuning_type != "full" and hasattr(model, output_layer_name): if finetuning_args.finetuning_type != "full" and hasattr(model, output_layer_name):
output_layer: torch.nn.Linear = getattr(model, output_layer_name) output_layer = getattr(model, output_layer_name)
input_dtype = output_layer.weight.dtype if isinstance(output_layer, torch.nn.Linear):
def forward_in_fp32(self, x: torch.Tensor) -> torch.Tensor: def forward_in_fp32(self, x: torch.Tensor) -> torch.Tensor:
return torch.nn.Linear.forward(self, x.to(input_dtype)).to(torch.float32) return output_layer.__class__.forward(self, x.to(output_layer.weight.dtype)).to(torch.float32)
output_layer.forward = MethodType(forward_in_fp32, output_layer) output_layer.forward = MethodType(forward_in_fp32, output_layer)