convert: skip vision tensors in NVFP4 text-model path

The MmprojModel branch already skipped non-vision tensors. The inverse
case was missing: when processing the text body of a VLM (e.g. Qwopus3.6
Qwen3_5ForConditionalGeneration) that has NVFP4-quantized weights, the
visual.* tensors flow through _generate_nvfp4_tensors -> map_tensor_name
and crash with "Can not map tensor model.visual.blocks.0.attn.proj.weight".

Mirror the MmprojModel skip on text models so visual tensors are left
for the --mmproj pass.

Author: jarvis

convert_hf_to_gguf.py

@@ -654,6 +654,15 @@ def _generate_nvfp4_tensors(self):
                 or "vision_model" in name
             ):
                 continue
+            # Inverse: when processing text model, skip vision tensors that
+            # NVFP4 packing cannot route via the LLM tensor name map.
+            if not isinstance(self, MmprojModel) and (
+                name.startswith("model.visual")
+                or name.startswith("visual.")
+                or "vision_tower" in name
+                or "vision_model" in name
+            ):
+                continue
             scale_name = name.replace(".weight", ".weight_scale")
             scale2_name = name.replace(".weight", ".weight_scale_2")
             input_scale_name = name.replace(".weight", ".input_scale")