Fix Qwen3VL attribute path: self.vlm.visual/.language_model -> self.vlm.model.visual/.language_model (transformers 4.57.6+ nests these under .model)

src/policies/LabVLA/modeling_labvla.py

@@ -314,8 +314,8 @@ class LabVLAModel(nn.Module):
 
     def set_requires_grad(self):
         if self.config.freeze_vision_encoder:
-            self.vlm.visual.eval()
-            for p in self.vlm.visual.parameters():
+            self.vlm.model.visual.eval()
+            for p in self.vlm.model.visual.parameters():
                 p.requires_grad = False
 
         if self.config.train_expert_only:
@@ -351,7 +351,7 @@ class LabVLAModel(nn.Module):
     def train(self, mode: bool = True):
         super().train(mode)
         if self.config.freeze_vision_encoder:
-            self.vlm.visual.eval()
+            self.vlm.model.visual.eval()
         if self.config.train_expert_only:
             self.vlm.eval()
         if self.config.train_vlm_only and self.dit_action_head is not None:
@@ -392,13 +392,13 @@ class LabVLAModel(nn.Module):
 
         # Disable checkpointing for vision encoder if not requested
         if not gc_visual_encoder:
-            for module in self.vlm.visual.modules():
+            for module in self.vlm.model.visual.modules():
                 if hasattr(module, 'gradient_checkpointing'):
                     module.gradient_checkpointing = False
 
         # Disable checkpointing for Language Model if not requested
         if not gc_language_model:
-            for module in self.vlm.language_model.modules():
+            for module in self.vlm.model.language_model.modules():
                 if hasattr(module, 'gradient_checkpointing'):
                     module.gradient_checkpointing = False
 
@@ -420,8 +420,8 @@ class LabVLAModel(nn.Module):
             total = sum(1 for m in root.modules() if hasattr(m, "gradient_checkpointing"))
             active = sum(1 for m in root.modules() if getattr(m, "gradient_checkpointing", False))
             return name, active, total
-        v = _count(self.vlm.visual, "visual")
-        l = _count(self.vlm.language_model, "language_model")
+        v = _count(self.vlm.model.visual, "visual")
+        l = _count(self.vlm.model.language_model, "language_model")
         dit_on = int(getattr(self.dit_action_head, "gradient_checkpointing", False)) if self.dit_action_head is not None else 0
         logger.info(
             f"GC layers active: visual={v[1]}/{v[2]}  language_model={l[1]}/{l[2]}  dit={dit_on}"
@@ -472,7 +472,7 @@ class LabVLAModel(nn.Module):
         D1 = pixel_values.shape[-1]
         pixel_values_flat = pixel_values.view(-1, D1)
         image_grid_thw_flat = image_grid_thw.view(-1, 3)
-        image_embs, _ = self.vlm.visual(pixel_values_flat, image_grid_thw_flat)
+        image_embs, _ = self.vlm.model.visual(pixel_values_flat, image_grid_thw_flat)
 
         embs = self.vlm.get_input_embeddings()(lang_tokens)
         B, L, D2 = embs.shape
@@ -518,7 +518,7 @@ class LabVLAModel(nn.Module):
 
         HF's ``create_causal_mask`` early-exits when given a 4D mask and
         uses it as-is (transformers/masking_utils.py). So routing this
-        mask through ``self.vlm.language_model(attention_mask=...)`` bypasses
+        mask through ``self.vlm.model.language_model(attention_mask=...)`` bypasses
         the auto-causal logic entirely.
 
         Returns:
@@ -633,7 +633,7 @@ class LabVLAModel(nn.Module):
 
         # Run VLM language model (use_cache=False must be passed explicitly; otherwise gradient checkpointing cannot save memory properly)
         layerwise = bool(getattr(self.config, "dit_layerwise_vlm_features", False))
-        vlm_output = self.vlm.language_model(
+        vlm_output = self.vlm.model.language_model(
             inputs_embeds=embs,
             attention_mask=lm_attention_mask,
             position_ids=position_ids,
@@ -761,7 +761,7 @@ class LabVLAModel(nn.Module):
         else:
             lm_attention_mask = full_attn
 
-        lm_out = self.vlm.language_model(
+        lm_out = self.vlm.model.language_model(
             inputs_embeds=full_embeds,
             attention_mask=lm_attention_mask,
             position_ids=full_pos_ids,
@@ -1209,14 +1209,14 @@ class LabVLAModel(nn.Module):
             # semantically equivalent for both branches.
             if self.config.train_expert_only:
                 with torch.no_grad():
-                    lm_out = self.vlm.language_model(
+                    lm_out = self.vlm.model.language_model(
                         inputs_embeds=full_embeds,
                         attention_mask=lm_attention_mask,
                         position_ids=full_pos_ids,
                         use_cache=False,
                     )
             else:
-                lm_out = self.vlm.language_model(
+                lm_out = self.vlm.model.language_model(
                     inputs_embeds=full_embeds,
                     attention_mask=lm_attention_mask,
                     position_ids=full_pos_ids,
@@ -1385,7 +1385,7 @@ class LabVLAModel(nn.Module):
         else:
             lm_attention_mask = full_attn
 
-        lm_out = self.vlm.language_model(
+        lm_out = self.vlm.model.language_model(
             inputs_embeds=full_embeds,
             attention_mask=lm_attention_mask,
             position_ids=full_pos_ids,
@@ -1614,7 +1614,7 @@ class LabVLAModel(nn.Module):
             lm_attention_mask = full_attn
 
         # (6) SINGLE LM forward
-        lm_out = self.vlm.language_model(
+        lm_out = self.vlm.model.language_model(
             inputs_embeds=full_embeds,
             attention_mask=lm_attention_mask,
             position_ids=full_pos_ids,