Merge pull request #7 from bigcode-project/dev

Fix: Using tensor with dummy value rather than None for the temperatu…

Author: joaomonteirof

.gitignore

@@ -134,6 +134,6 @@ wandb-debug.log
 wandb/
 
 # hf
-cache-
-.lock
-.arrow
+*cache-*
+*.lock
+*.arrow

src/hf_trainer.py

@@ -119,7 +119,7 @@ def compute_loss(
             try:
                 temp_coef = temp_coef_fn.get_temp_coef()
             except AttributeError:
-                temp_coef = None
+                temp_coef = torch.Tensor([0.0]).to(contrastive_loss.device)
 
             return contrastive_loss, {
                 "source_embedding": normalized_source_embedding,