Move logits.float() call

src/liger_kernel/transformers/model/llama.py

@@ -120,8 +120,9 @@ def lce_forward(
             logits = torch.cat(logits, dim=-1)
         else:
             logits = self.lm_head(hidden_states)
-        logits = logits.float()
         if labels is not None:
+            # Upcast to float if we need to compute the loss to avoid potential precision issues
+            logits = logits.float()
             # Shift so that tokens < n predict n
             shift_logits = logits[..., :-1, :].contiguous()
             shift_labels = labels[..., 1:].contiguous()

src/liger_kernel/transformers/model/mixtral.py

@@ -103,7 +103,6 @@ def lce_forward(
 
     hidden_states = outputs[0]
     logits = self.lm_head(hidden_states)
-    logits = logits.float()
 
     loss = None
     if self.training and (labels is not None):
@@ -116,6 +115,8 @@ def lce_forward(
         lce = LigerFusedLinearCrossEntropyLoss()
         loss = lce(self.lm_head.weight, shift_hidden_states, shift_labels)
     elif labels is not None:
+        # Upcast to float if we need to compute the loss to avoid potential precision issues
+        logits = logits.float()
         # Shift so that tokens < n predict n
         shift_logits = logits[..., :-1, :].contiguous()
         shift_labels = labels[..., 1:].contiguous()

src/liger_kernel/transformers/model/phi3.py

@@ -108,10 +108,11 @@ def lce_forward(
         loss = lce(self.lm_head.weight, shift_hidden_states, shift_labels)
     else:
         logits = self.lm_head(hidden_states)
-        logits = logits.float()
 
         loss = None
         if labels is not None:
+            # Upcast to float if we need to compute the loss to avoid potential precision issues
+            logits = logits.float()
             # Shift so that tokens < n predict n
             shift_logits = logits[..., :-1, :].contiguous()
             shift_labels = labels[..., 1:].contiguous()

src/liger_kernel/transformers/model/qwen2.py

@@ -109,8 +109,9 @@ def lce_forward(
 
     else:
         logits = self.lm_head(hidden_states)
-        logits = logits.float()
         if labels is not None:
+            # Upcast to float if we need to compute the loss to avoid potential precision issues
+            logits = logits.float()
             # Shift so that tokens < n predict n
             shift_logits = logits[..., :-1, :].contiguous()
             shift_labels = labels[..., 1:].contiguous()

src/liger_kernel/transformers/model/qwen2_vl.py

@@ -150,8 +150,9 @@ def lce_forward(
         loss = lce(self.lm_head.weight, shift_hidden_states, shift_labels)
     else:
         logits = self.lm_head(hidden_states)
-        logits = logits.float()
         if labels is not None:
+            # Upcast to float if we need to compute the loss to avoid potential precision issues
+            logits = logits.float()
             # Shift so that tokens < n predict n
             shift_logits = logits[..., :-1, :].contiguous()
             shift_labels = labels[..., 1:].contiguous()