Bug Fix for issue huggingface#34294 (huggingface#34295)

Update SiglipVisionEmbeddings.forward to cast input to correct dtype before embedding it.

src/transformers/models/siglip/modeling_siglip.py

@@ -308,7 +308,8 @@ def interpolate_pos_encoding(self, embeddings: torch.Tensor, height: int, width:
 
     def forward(self, pixel_values: torch.FloatTensor, interpolate_pos_encoding=False) -> torch.Tensor:
         _, _, height, width = pixel_values.shape
-        patch_embeds = self.patch_embedding(pixel_values)  # shape = [*, width, grid, grid]
+        target_dtype = self.patch_embedding.weight.dtype
+        patch_embeds = self.patch_embedding(pixel_values.to(dtype=target_dtype))  # shape = [*, width, grid, grid]
         embeddings = patch_embeds.flatten(2).transpose(1, 2)
 
         if interpolate_pos_encoding: