clip : fix confused naming ffn_up and ffn_down (ggml-org#13290)

* clip :  fix confused naming ffn_up and ffn_down

* rm ffn_i/o/g naming

* rename n_embd, n_ff

* small fix

* no check n_ff

Author: ngxson

convert_hf_to_gguf.py

@@ -1778,6 +1778,12 @@ class LlamaModel(TextModel):
     model_arch = gguf.MODEL_ARCH.LLAMA
     undo_permute = True
 
+    def __init__(self, *args, **kwargs):
+        super().__init__(*args, **kwargs)
+        # fix for SmolVLM2, missing `num_attention_heads` in config.json
+        if self.hf_arch == "VLlama3ForCausalLM":
+            self.hparams["num_attention_heads"] = self.hparams.get("num_attention_heads", 32)
+
     def set_vocab(self):
         try:
             self._set_vocab_sentencepiece()

gguf-py/gguf/tensor_mapping.py

@@ -977,15 +977,12 @@ class TensorNameMap:
             "visual.blocks.{bid}.norm2", # qwen2vl
         ),
 
-        # some namings are messed up because the original llava code swapped fc1 and fc2
-        # we have no better way to fix it, just be careful
-        # new models like pixtral use the correct naming
         MODEL_TENSOR.V_ENC_FFN_UP: (
             "vision_tower.vision_model.encoder.layers.{bid}.mlp.fc1",
             "vpm.encoder.layers.{bid}.mlp.fc1",
-            "model.vision_model.encoder.layers.{bid}.mlp.fc2", # SmolVLM, gemma3 (note: name is swapped)
+            "model.vision_model.encoder.layers.{bid}.mlp.fc1", # SmolVLM, gemma3
             "vision_tower.transformer.layers.{bid}.feed_forward.up_proj", # pixtral
-            "visual.blocks.{bid}.mlp.fc2", # qwen2vl
+            "visual.blocks.{bid}.mlp.fc1", # qwen2vl
             "visual.blocks.{bid}.mlp.up_proj", # qwen2.5vl
         ),
 
@@ -997,9 +994,9 @@ class TensorNameMap:
         MODEL_TENSOR.V_ENC_FFN_DOWN: (
             "vision_tower.vision_model.encoder.layers.{bid}.mlp.fc2",
             "vpm.encoder.layers.{bid}.mlp.fc2",
-            "model.vision_model.encoder.layers.{bid}.mlp.fc1", # SmolVLM, gemma3 (note: name is swapped)
+            "model.vision_model.encoder.layers.{bid}.mlp.fc2", # SmolVLM, gemma3
             "vision_tower.transformer.layers.{bid}.feed_forward.down_proj", # pixtral
-            "visual.blocks.{bid}.mlp.fc1", # qwen2vl
+            "visual.blocks.{bid}.mlp.fc2", # qwen2vl
             "visual.blocks.{bid}.mlp.down_proj", # qwen2.5vl
         ),