fix: haru's sft being incompatible with the ColossalAI fine-tune script

training/harubaru-convogpt/sft.py

@@ -4,9 +4,9 @@ import accelerate
 import tqdm
 import time
 import argparse
-import wandb
+# import wandb
 
-from dataset import TokenizedDataset, FeedbackDataset, SFTDataset
+from .dataset import TokenizedDataset, FeedbackDataset, SFTDataset
 
 from transformers import AutoModelForCausalLM, AutoTokenizer
 from transformers.modeling_outputs import CausalLMOutput
@@ -33,6 +33,7 @@ def sft_forward(
     except AttributeError:
         return_dict = True
 
+    '''
     outputs = self.transformer(
         input_ids,
         attention_mask=attention_mask,
@@ -48,6 +49,21 @@ def sft_forward(
     sequence_output = outputs[0]
 
     logits = self.lm_head(sequence_output)
+    '''
+    outputs = self(
+        input_ids,
+        attention_mask=attention_mask,
+        use_cache=False,
+        # token_type_ids=token_type_ids,
+        # position_ids=position_ids,
+        # head_mask=head_mask,
+        # inputs_embeds=inputs_embeds,
+        # output_attentions=output_attentions,
+        # output_hidden_states=output_hidden_states,
+        # return_dict=return_dict,
+    )
+
+    logits = outputs["logits"]
 
     answer_logits = logits[:, start_positions[0]:end_positions[0]+1]
     answer_input_ids = input_ids[:, start_positions[0]:end_positions[0]+1]