[Fix] Remove unused packing_position_embedding from PaddleOCRVL for better...

[Fix] Remove unused packing_position_embedding from PaddleOCRVL for better checkpoint compatibility (#38232)
Signed-off-by: zhangyue66 <zhangyue66@baidu.com>

vllm/model_executor/models/paddleocr_vl.py

@@ -409,7 +409,6 @@ class SiglipVisionEmbeddings(nn.Module):
         self.cache_position_embedding = dict()
         self.cache_position_count = dict()
         self.position_embedding = nn.Embedding(self.num_positions, self.embed_dim)
-        self.packing_position_embedding = nn.Embedding(32768, self.embed_dim)
 
         self.register_buffer(
             "position_ids",
@@ -501,7 +500,6 @@ class SiglipVisionEmbeddings(nn.Module):
             patch_embeds = self.patch_embedding(pixel_values.to(dtype=target_dtype))
             embeddings = patch_embeds.flatten(-2).squeeze(-1)
 
-            if interpolate_pos_encoding and image_grid_thw is not None:
             start = 0
             tmp_embeddings = list()
             for image_grid in image_grid_thw:
@@ -517,8 +515,7 @@ class SiglipVisionEmbeddings(nn.Module):
                 tmp_embeddings.append(image_embeddings)
                 start = end
             embeddings = torch.concat(tmp_embeddings, dim=0).unsqueeze(0)
-            else:
-                embeddings = embeddings + self.packing_position_embedding(position_ids)
+
             return embeddings
         else:
             raise ValueError(
@@ -939,6 +936,8 @@ class SiglipVisionModel(nn.Module):
                 continue
             if "head.mlp" in name or "head.probe" in name:
                 continue
+            if "packing_position_embedding" in name:
+                continue
             if self.quant_config is not None and (
                 scale_name := self.quant_config.get_cache_scale(name)
             ):