Minor Fix of Output Putting Rotary back to hf rotary

Signed-off-by: Dipankar Sarkar <dipankar@qti.qualcomm.com>

Author: qcdipankar

QEfficient/transformers/models/pytorch_transforms.py

@@ -175,7 +175,6 @@
     Qwen3VLTextDecoderLayer,
     Qwen3VLTextModel,
     Qwen3VLTextRMSNorm,
-    Qwen3VLTextRotaryEmbedding,
     Qwen3VLVisionAttention,
     Qwen3VLVisionModel,
 )
@@ -393,7 +392,7 @@
     QEffQwen3VLTextAttention,
     QEffQwen3VLTextDecoderLayer,
     QEffQwen3VLTextModel,
-    QEffQwen3VLTextRotaryEmbedding,
+    # QEffQwen3VLTextRotaryEmbedding,
     QEffQwen3VLVisionAttention,
     QEffQwen3VLVisionModel,
 )
@@ -588,17 +587,15 @@ class KVCacheTransform(ModuleMappingTransform):
         Qwen2_5_VLDecoderLayer: QEffQwen2_5_VLDecoderLayer,
         Qwen2_5_VisionTransformerPretrainedModel: QEffQwen2_5_VisionTransformerPretrainedModel,
         Qwen2_5_VLVisionAttention: QEffQwen2_5_VLVisionAttention,
-        
-
-        #Qwen3vl
+        # Qwen3vl
         Qwen3VLForConditionalGeneration: QEffQwen3VLForConditionalGeneration,
         Qwen3VLModel: QEffQwen3VLModel,
         Qwen3VLTextAttention: QEffQwen3VLTextAttention,
         Qwen3VLTextDecoderLayer: QEffQwen3VLTextDecoderLayer,
         Qwen3VLVisionAttention: QEffQwen3VLVisionAttention,
         Qwen3VLVisionModel: QEffQwen3VLVisionModel,
         Qwen3VLTextModel: QEffQwen3VLTextModel,
-        Qwen3VLTextRotaryEmbedding: QEffQwen3VLTextRotaryEmbedding,  # reusing decoder layer for rotary embedding as they are tightly coupled in forward pass
+        # Qwen3VLTextRotaryEmbedding: QEffQwen3VLTextRotaryEmbedding,  # reusing decoder layer for rotary embedding as they are tightly coupled in forward pass
         # Starcoder2
         Starcoder2Attention: QEffStarcoder2Attention,
         Starcoder2DecoderLayer: QEFFStarcoder2DecoderLayer,
@@ -856,4 +853,4 @@ def apply(cls, model: nn.Module, num_kv_blocks) -> Tuple[nn.Module, bool]:
                 transformed = True  # Set to True if at least one transformation occurs
             elif module.__class__.__name__.endswith("Attention") and type(module) not in cls._module_mapping:
                 warnings.warn(f"KV blocking is not yet supported for {type(module)}.")
-        return model, transformed
+        return model, transformed