.to(torch.bfloat16) # Add .to(torch.bfloat16) here for explicit casting

app.py

@@ -94,7 +94,7 @@ def get_som_response(instruction, image_som):
     )
 
     inputs = magma_processor(images=[image_som], texts=prompt, return_tensors="pt")
-    inputs['pixel_values'] = inputs['pixel_values'].unsqueeze(0)
+    inputs['pixel_values'] = inputs['pixel_values'].unsqueeze(0).to(torch.bfloat16) # Add .to(torch.bfloat16) here for explicit casting
     inputs['image_sizes'] = inputs['image_sizes'].unsqueeze(0)
     # inputs = inputs.to("cuda")
     inputs = inputs.to("cuda", dtype=torch.bfloat16)
@@ -133,7 +133,7 @@ def get_qa_response(instruction, image):
     )
 
     inputs = magma_processor(images=[image], texts=prompt, return_tensors="pt")
-    inputs['pixel_values'] = inputs['pixel_values'].unsqueeze(0)
+    inputs['pixel_values'] = inputs['pixel_values'].unsqueeze(0).to(torch.bfloat16) # Add .to(torch.bfloat16) here for explicit casting
     inputs['image_sizes'] = inputs['image_sizes'].unsqueeze(0)
     # inputs = inputs.to("cuda")
     inputs = inputs.to("cuda", dtype=torch.bfloat16)