move V_ENC_EMBD_PATCH handling to Qwen2VLVisionModel

Author: ngxson

convert_hf_to_gguf.py

@@ -2653,8 +2653,8 @@ def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iter
                 del c1, c2, kh, kw  # unused
                 assert kt == 2, "Current implmentation only support temporal_patch_size of 2"
                 return [
-                    (self.map_tensor_name(name), data_torch[:, :, 0, ...]),
-                    (self.map_tensor_name(name + '.1'), data_torch[:, :, 1, ...]),
+                    (gguf.TENSOR_NAMES[gguf.MODEL_TENSOR.V_ENC_EMBD_PATCH] + ".weight"  , data_torch[:, :, 0, ...]),
+                    (gguf.TENSOR_NAMES[gguf.MODEL_TENSOR.V_ENC_EMBD_PATCH] + ".weight.1", data_torch[:, :, 1, ...]),
                 ]
             else:
                 return [(self.map_tensor_name(name), data_torch)]

gguf-py/gguf/constants.py

@@ -481,7 +481,6 @@ class MODEL_TENSOR(IntEnum):
     V_MMPROJ_PEG         = auto()
     V_ENC_EMBD_CLS       = auto()
     V_ENC_EMBD_PATCH     = auto()
-    V_ENC_EMBD_PATCH1    = auto() # qwen2vl
     V_ENC_EMBD_POS       = auto()
     V_ENC_ATTN_Q         = auto()
     V_ENC_ATTN_K         = auto()
@@ -739,7 +738,6 @@ class MODEL_TENSOR(IntEnum):
     MODEL_TENSOR.V_MMPROJ_PEG:              "mm.model.peg.{bid}",
     MODEL_TENSOR.V_ENC_EMBD_CLS:            "v.class_embd",
     MODEL_TENSOR.V_ENC_EMBD_PATCH:          "v.patch_embd",
-    MODEL_TENSOR.V_ENC_EMBD_PATCH1:         "v.patch_embd.weight.1", # qwen2vl
     MODEL_TENSOR.V_ENC_EMBD_POS:            "v.position_embd",
     MODEL_TENSOR.V_ENC_ATTN_Q:              "v.blk.{bid}.attn_q",
     MODEL_TENSOR.V_ENC_ATTN_K:              "v.blk.{bid}.attn_k",
@@ -778,7 +776,6 @@ class MODEL_TENSOR(IntEnum):
         MODEL_TENSOR.V_MMPROJ_PEG,
         MODEL_TENSOR.V_ENC_EMBD_CLS,
         MODEL_TENSOR.V_ENC_EMBD_PATCH,
-        MODEL_TENSOR.V_ENC_EMBD_PATCH1,
         MODEL_TENSOR.V_ENC_EMBD_POS,
         MODEL_TENSOR.V_ENC_ATTN_Q,
         MODEL_TENSOR.V_ENC_ATTN_K,

gguf-py/gguf/tensor_mapping.py

@@ -923,10 +923,6 @@ class TensorNameMap:
             "visual.patch_embed.proj", # qwen2vl
         ),
 
-        MODEL_TENSOR.V_ENC_EMBD_PATCH1: (
-            "visual.patch_embed.proj.weight.1", # qwen2vl, generated
-        ),
-
         MODEL_TENSOR.V_ENC_EMBD_POS: (
             "vision_tower.vision_model.embeddings.position_embedding",
             "vpm.embeddings.position_embedding",