mtmd: add n_embd_head (#26342)

Co-authored-by: Daniel Han <unslothai@gmail.com>

gguf-py/gguf/constants.py

@@ -353,6 +353,7 @@ class Keys:
         class Attention:
             HEAD_COUNT      = "clip.vision.attention.head_count"
             HEAD_COUNT_KV   = "clip.vision.attention.head_count_kv" # used by mimovl (GQA)
+            HEAD_DIM        = "clip.vision.attention.head_dim" # set when qkv width != n_embd
             LAYERNORM_EPS   = "clip.vision.attention.layer_norm_epsilon"
 
         class Projector:

gguf-py/gguf/gguf_writer.py

@@ -1226,6 +1226,9 @@ class GGUFWriter:
     def add_vision_head_count_kv(self, value: int) -> None:
         self.add_uint32(Keys.ClipVision.Attention.HEAD_COUNT_KV, value)
 
+    def add_vision_head_dim(self, value: int) -> None:
+        self.add_uint32(Keys.ClipVision.Attention.HEAD_DIM, value)
+
     def add_vision_attention_layernorm_eps(self, value: float) -> None:
         self.add_float32(Keys.ClipVision.Attention.LAYERNORM_EPS, value)