modified: src/flux/modules/conditioner.py

src/flux/modules/conditioner.py

@@ -96,11 +96,32 @@ class HFEmbedder(nn.Module):
 
         print(f"Input IDs device: {input_ids.device}")
         print(f"Attention Mask device: {attention_mask.device}")
+
+        # --- FIX FOR CLIP POSITION IDs ---
+        # Prepare arguments for the model call
+        model_kwargs = {
+            "input_ids": input_ids,
+            "attention_mask": attention_mask,
+            "output_hidden_states": False,
+        }
+
+        # If it's a CLIP model, explicitly generate and pass position_ids
+        if self.is_clip:
+            # Generate position_ids: [0, 1, 2, ..., max_length-1] for each item in the batch
+            # Shape: (batch_size, max_length)
+            position_ids = torch.arange(self.max_length, dtype=torch.long, device=input_ids.device).expand(input_ids.size(0), -1)
+            print(f"Generated CLIP position_ids: shape={position_ids.shape}, range=[{position_ids.min().item()}, {position_ids.max().item()}]")
+            # Check if generated position_ids are within the model's limit
+            max_pos_emb = getattr(self.hf_module.config, 'max_position_embeddings', -1)
+            if max_pos_emb > 0 and position_ids.max() >= max_pos_emb:
+                 raise ValueError(f"Generated position_ids max ({position_ids.max().item()}) >= model's max_position_embeddings ({max_pos_emb})")
+
         try:
             outputs = self.hf_module(
                 input_ids=input_ids,
                 attention_mask=attention_mask,
                 output_hidden_states=False,
+                position_ids=position_ids
             )
         except IndexError as e:
             # 捕获并提供更详细的错误上下文