Support precomputed multimodal features for Qwen-VL and Gemma3 models. (#6136)

Co-authored-by: Yury Sulsky <ysulsky@tesla.com>

python/sglang/srt/models/qwen2_5_vl.py

@@ -497,6 +497,12 @@ class Qwen2_5_VLForConditionalGeneration(nn.Module):
         return pattern.pad_input_tokens(input_ids, mm_inputs)
 
     def get_image_feature(self, items: List[MultimodalDataItem]) -> torch.Tensor:
+        if any(item.precomputed_features is not None for item in items):
+            if not all(item.precomputed_features is not None for item in items):
+                raise NotImplementedError(
+                    "MM inputs where only some items are precomputed."
+                )
+            return torch.concat([item.precomputed_features for item in items])
         # in qwen-vl, last dim is the same
         pixel_values = torch.cat([item.pixel_values for item in items], dim=0).type(
             self.visual.dtype