metadata
license: llama3.2
language:
- en
base_model:
- meta-llama/Llama-3.2-1B
pipeline_tag: text-generation
library_name: transformers
class LayerSkipSFTTrainer(SFTTrainer):
def __init__(self, *args, **kwargs):
super().__init__(*args, **kwargs)
self.early_exit_layer = 0 # initialize with 0
self.always_last_layer = True
def compute_loss(self, model, inputs, return_outputs=False, num_items_in_batch=None):
self.early_exit_layer = (self.early_exit_layer % (model.config.num_hidden_layers - 1)) + 1 # rotates between [1, num_hidden_layers-1]
labels = inputs.pop("labels")
outputs = model(**inputs, output_hidden_states=True)
hidden_state = outputs["hidden_states"][self.early_exit_layer]
logits = model.lm_head(hidden_state)
loss = model.loss_function(logits=logits, labels=labels, vocab_size=model.vocab_size)
if self.always_last_layer:
loss = loss + model.loss_function(logits=outputs["logits"], labels=labels, vocab_size=model.vocab_size)
return loss