From: Xuan-Son Nguyen <redacted>
Date: Wed, 28 May 2025 08:05:54 +0000 (+0200)
Subject: convert : fix tensor naming conflict for llama 4 vision (#13836)
X-Git-Url: https://git.djapps.eu/?a=commitdiff_plain;h=26b79b6cb3e7840ff15729350e95907e19f9f480;p=pkg%2Fggml%2Fsources%2Fllama.cpp

convert : fix tensor naming conflict for llama 4 vision (#13836)

* convert : fix tensor naming conflict for llama 4 vision

* add comment
---

diff --git a/convert_hf_to_gguf.py b/convert_hf_to_gguf.py
index a015ecee..7f935d09 100755
--- a/convert_hf_to_gguf.py
+++ b/convert_hf_to_gguf.py
@@ -2169,6 +2169,9 @@ class Llama4VisionModel(MmprojModel):
             # process vision tensors
             if "positional_embedding_vlm" in name and ".weight" not in name:
                 name += ".weight"
+            if "multi_modal_projector.linear_1" in name:
+                # despite the name with number postfix, this is a single fully connected layer
+                return [(gguf.TENSOR_NAMES[gguf.MODEL_TENSOR.V_MMPROJ_FC], data_torch)]
             return [(self.map_tensor_name(name), data_torch)]
         return []
 
diff --git a/gguf-py/gguf/tensor_mapping.py b/gguf-py/gguf/tensor_mapping.py
index 000ffd00..48167dd6 100644
--- a/gguf-py/gguf/tensor_mapping.py
+++ b/gguf-py/gguf/tensor_mapping.py
@@ -902,7 +902,6 @@ class TensorNameMap:
 
         MODEL_TENSOR.V_MMPROJ_FC: (
             "model.connector.modality_projection.proj", # SmolVLM
-            "multi_modal_projector.linear_1", # llama 4
         ),
 
         MODEL_TENSOR.V_MMPROJ_MLP: (