Update app.py

app.py

@@ -36,12 +36,17 @@ def describe_image(image, user_question="Describe this image in great detail."):
         pil_images = [Image.open(BytesIO(image_byte_arr.read()))]
         image_byte_arr.seek(0)
 
-        # Prepare inputs with CPU and float32 type
+        # Prepare inputs
         prepare_inputs = vl_chat_processor(
             conversations=conversation,
             images=pil_images,
             force_batchify=True
-        ).to(torch.float32)  # Convert to float32 for CPU compatibility
+        )
+
+        # Convert all tensors in prepare_inputs to float32 for CPU compatibility
+        for key in prepare_inputs:
+            if isinstance(prepare_inputs[key], torch.Tensor):
+                prepare_inputs[key] = prepare_inputs[key].to(dtype=torch.float32)
 
         # Load model with CPU and float32 weights
         vl_gpt = AutoModelForCausalLM.from_pretrained(
@@ -55,7 +60,7 @@ def describe_image(image, user_question="Describe this image in great detail."):
         # Generate response with CPU
         outputs = vl_gpt.language_model.generate(
             inputs_embeds=inputs_embeds,
-            attention_mask=prepare_inputs.attention_mask,
+            attention_mask=prepare_inputs["attention_mask"],
             pad_token_id=tokenizer.eos_token_id,
             bos_token_id=tokenizer.bos_token_id,
             eos_token_id=tokenizer.eos_token_id,