default use xentropy and use only last logit in sampling

This commit is contained in:
Yu Li
2023-09-14 21:31:48 -05:00
parent 2c4eec6aa8
commit 6327b9fb1a
2 changed files with 2 additions and 1 deletions
+1
View File
@@ -325,6 +325,7 @@ class SampleGenerateCallback(transformers.TrainerCallback):
input_ids=input_ids,
max_new_tokens=15,
do_sample = False,
only_last_logit=True, # to save memory
)
+1 -1
View File
@@ -782,7 +782,7 @@ class LlamaForCausalLM(LlamaPreTrainedModel):
output_hidden_states: Optional[bool] = None,
return_dict: Optional[bool] = None,
only_last_logit: Optional[bool] = None,
xentropy: Optional[bool] = None,
xentropy: Optional[bool] = True,
is_padded_inputs: Optional[bool] = None,
) -> Union[Tuple, CausalLMOutputWithPast]:
r"""