[Models] Replace all nn.Conv2d with vLLM's Conv2dLayer (#28842)

Signed-off-by: Isotr0py <mozf@mail2.sysu.edu.cn>

vllm/model_executor/models/siglip2navit.py

@@ -16,6 +16,7 @@ from vllm.attention.backends.registry import AttentionBackendEnum
 from vllm.attention.layer import maybe_get_vit_flash_attn_backend
 from vllm.distributed import divide, get_tensor_model_parallel_world_size
 from vllm.model_executor.layers.activation import get_act_fn
+from vllm.model_executor.layers.conv import Conv2dLayer
 from vllm.model_executor.layers.linear import (
     ColumnParallelLinear,
     LinearBase,
@@ -67,7 +68,7 @@ class Siglip2VisionEmbeddings(nn.Module):
                 self.position_embedding = nn.Embedding(self.num_patches, self.embed_dim)
 
         else:
-            self.patch_embedding = nn.Conv2d(
+            self.patch_embedding = Conv2dLayer(
                 in_channels=config.num_channels,
                 out_channels=self.embed_dim,
                 kernel_size=self.patch_size,
@@ -99,7 +100,7 @@ class Siglip2VisionEmbeddings(nn.Module):
         target_dtype = self.patch_embedding.weight.dtype
         if isinstance(self.patch_embedding, LinearBase):
             patch_embeds = self.patch_embedding(pixel_values.to(dtype=target_dtype))
-        elif isinstance(self.patch_embedding, nn.Conv2d):
+        elif isinstance(self.patch_embedding, Conv2dLayer):
             pixel_values = pixel_values.view(
                 -1,
                 self.config.num_channels * self.config.temporal_patch_size,