granite4_vision: replace unbound forward call with self.language_model()

The Granite4VisionTextModel.forward(self.language_model, ...) pattern was
added to bypass nn.Module.__call__ overhead, but compare_implementations
shows no measurable difference (1.05x total vs card, within noise).

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
Signed-off-by: artemspector <artems@il.ibm.com>

Author: artemspector

src/transformers/models/granite4_vision/modeling_granite4_vision.py

@@ -1045,10 +1045,7 @@ def forward(
                     inputs_embeds = inputs_embeds.masked_fill(vision_mask, 0.0)
                 deepstack_features[llm_layer_idx] = concat_features
 
-        # Bypass nn.Module.__call__ overhead by calling the unbound forward directly.
-        # nn.Module.__call__ has non-trivial per-call overhead that accumulates across 40 layers × N steps.
-        outputs = Granite4VisionTextModel.forward(
-            self.language_model,
+        outputs = self.language_model(
             input_ids=None,
             inputs_embeds=inputs_embeds,
             attention_mask=attention_mask,

src/transformers/models/granite4_vision/modular_granite4_vision.py

@@ -704,10 +704,7 @@ def forward(
                     inputs_embeds = inputs_embeds.masked_fill(vision_mask, 0.0)
                 deepstack_features[llm_layer_idx] = concat_features
 
-        # Bypass nn.Module.__call__ overhead by calling the unbound forward directly.
-        # nn.Module.__call__ has non-trivial per-call overhead that accumulates across 40 layers × N steps.
-        outputs = Granite4VisionTextModel.forward(
-            self.language_model,
+        outputs = self.language_model(
             input_ids=None,
             inputs_embeds=inputs_embeds,
             attention_mask=attention_mask,