fix: cast input pixels to appropriate dtype for image_to_text pipelines (#24947)

* fix: cast input pixels to appropriate dtype for image_to_text tasks

* fix: add casting to pixel inputs of additional models after running copy checks

src/transformers/models/altclip/modeling_altclip.py

@@ -1022,7 +1022,8 @@ class AltCLIPVisionEmbeddings(nn.Module):
 
     def forward(self, pixel_values: torch.FloatTensor) -> torch.Tensor:
         batch_size = pixel_values.shape[0]
-        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]
         patch_embeds = patch_embeds.flatten(2).transpose(1, 2)
 
         class_embeds = self.class_embedding.expand(batch_size, 1, -1)

src/transformers/models/blip/modeling_blip.py

@@ -246,7 +246,7 @@ class BlipVisionEmbeddings(nn.Module):
     def forward(self, pixel_values: torch.FloatTensor) -> torch.Tensor:
         batch_size = pixel_values.shape[0]
         target_dtype = self.patch_embedding.weight.dtype
-        patch_embeds = self.patch_embedding(pixel_values)  # shape = [*, width, grid, grid]
+        patch_embeds = self.patch_embedding(pixel_values.to(dtype=target_dtype))  # shape = [*, width, grid, grid]
         patch_embeds = patch_embeds.flatten(2).transpose(1, 2)
 
         class_embeds = self.class_embedding.expand(batch_size, 1, -1).to(target_dtype)

src/transformers/models/blip_2/modeling_blip_2.py

@@ -109,7 +109,7 @@ class Blip2VisionEmbeddings(nn.Module):
     def forward(self, pixel_values: torch.FloatTensor) -> torch.Tensor:
         batch_size = pixel_values.shape[0]
         target_dtype = self.patch_embedding.weight.dtype
-        patch_embeds = self.patch_embedding(pixel_values)  # shape = [*, width, grid, grid]
+        patch_embeds = self.patch_embedding(pixel_values.to(dtype=target_dtype))  # shape = [*, width, grid, grid]
         patch_embeds = patch_embeds.flatten(2).transpose(1, 2)
 
         class_embeds = self.class_embedding.expand(batch_size, 1, -1).to(target_dtype)

src/transformers/models/bridgetower/modeling_bridgetower.py

@@ -284,7 +284,8 @@ class BridgeTowerVisionEmbeddings(nn.Module):
 
     def forward(self, pixel_values: torch.FloatTensor) -> torch.Tensor:
         batch_size = pixel_values.shape[0]
-        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]
         patch_embeds = patch_embeds.flatten(2).transpose(1, 2)
 
         class_embeds = self.class_embedding.expand(batch_size, 1, -1)

src/transformers/models/chinese_clip/modeling_chinese_clip.py

@@ -196,7 +196,8 @@ class ChineseCLIPVisionEmbeddings(nn.Module):
 
     def forward(self, pixel_values: torch.FloatTensor) -> torch.Tensor:
         batch_size = pixel_values.shape[0]
-        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]
         patch_embeds = patch_embeds.flatten(2).transpose(1, 2)
 
         class_embeds = self.class_embedding.expand(batch_size, 1, -1)

src/transformers/models/clip/modeling_clip.py

@@ -192,7 +192,8 @@ class CLIPVisionEmbeddings(nn.Module):
 
     def forward(self, pixel_values: torch.FloatTensor) -> torch.Tensor:
         batch_size = pixel_values.shape[0]
-        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]
         patch_embeds = patch_embeds.flatten(2).transpose(1, 2)
 
         class_embeds = self.class_embedding.expand(batch_size, 1, -1)

src/transformers/models/git/modeling_git.py

@@ -628,7 +628,8 @@ class GitVisionEmbeddings(nn.Module):
 
     def forward(self, pixel_values: torch.FloatTensor) -> torch.Tensor:
         batch_size = pixel_values.shape[0]
-        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]
         patch_embeds = patch_embeds.flatten(2).transpose(1, 2)
 
         class_embeds = self.class_embedding.expand(batch_size, 1, -1)

src/transformers/models/instructblip/modeling_instructblip.py

@@ -110,7 +110,7 @@ class InstructBlipVisionEmbeddings(nn.Module):
     def forward(self, pixel_values: torch.FloatTensor) -> torch.Tensor:
         batch_size = pixel_values.shape[0]
         target_dtype = self.patch_embedding.weight.dtype
-        patch_embeds = self.patch_embedding(pixel_values)  # shape = [*, width, grid, grid]
+        patch_embeds = self.patch_embedding(pixel_values.to(dtype=target_dtype))  # shape = [*, width, grid, grid]
         patch_embeds = patch_embeds.flatten(2).transpose(1, 2)
 
         class_embeds = self.class_embedding.expand(batch_size, 1, -1).to(target_dtype)

src/transformers/models/x_clip/modeling_x_clip.py

@@ -143,7 +143,8 @@ class XCLIPVisionEmbeddings(nn.Module):
 
     def forward(self, pixel_values: torch.FloatTensor) -> torch.Tensor:
         batch_size = pixel_values.shape[0]
-        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]
         patch_embeds = patch_embeds.flatten(2).transpose(1, 2)
 
         class_embeds = self.class_embedding.expand(batch_size, 1, -1)