Update app.py

app.py

@@ -11,20 +11,20 @@ model_path = "deepseek-ai/deepseek-vl-1.3b-chat"
 vl_chat_processor = VLChatProcessor.from_pretrained(model_path)
 tokenizer = vl_chat_processor.tokenizer
 
-
-def describe_image(image, user_question="You are the best AP teacher in the world. Analyze the AP problem in the image, and solve it step by step to let a student who don't know how to solve it understand"):
+# Define the function for image description (CPU version)
+def describe_image(image, user_question="Describe this image in great detail."):
     try:
-        # Convert the PIL Image to a BytesIO object for compatibility
+        # Convert the PIL Image to a BytesIO object
         image_byte_arr = BytesIO()
-        image.save(image_byte_arr, format="PNG")  # Save image in PNG format
-        image_byte_arr.seek(0)  # Move pointer to the start
+        image.save(image_byte_arr, format="PNG")
+        image_byte_arr.seek(0)
 
-        # Define the conversation, using the user's question
+        # Define the conversation
         conversation = [
             {
                 "role": "User",
                 "content": f"<image_placeholder>{user_question}",
-                "images": [image_byte_arr]  # Pass the image byte array instead of an object
+                "images": [image_byte_arr]
             },
             {
                 "role": "Assistant",
@@ -32,42 +32,30 @@ def describe_image(image, user_question="You are the best AP teacher in the worl
             }
         ]
 
-        # Convert image byte array back to a PIL image for processing
-        pil_images = [Image.open(BytesIO(image_byte_arr.read()))]  # Convert byte back to PIL Image
-        image_byte_arr.seek(0)  # Reset the byte stream again for reuse
+        # Convert byte array back to PIL image
+        pil_images = [Image.open(BytesIO(image_byte_arr.read()))]
+        image_byte_arr.seek(0)
 
-        # Load images and prepare the inputs
+        # Prepare inputs with CPU and float32 type
         prepare_inputs = vl_chat_processor(
             conversations=conversation,
             images=pil_images,
             force_batchify=True
-        )
-
-        # Explicitly cast all tensors in prepare_inputs to torch.float16
-        prepare_inputs = {
-            k: v.to(torch.float16) if isinstance(v, torch.Tensor) else v
-            for k, v in prepare_inputs.items()
-        }
-
-        # Load and prepare the model
-        vl_gpt = AutoModelForCausalLM.from_pretrained(model_path, trust_remote_code=True).to(torch.float16).eval()
-        vl_gpt = vl_gpt.to(torch.float16)  # Explicitly ensure all components are in float16
+        ).to(torch.float32)  # Convert to float32 for CPU compatibility
 
-        # Generate embeddings from the image input
-        inputs_embeds = vl_gpt.prepare_inputs_embeds(**prepare_inputs).to(dtype=torch.float16)
+        # Load model with CPU and float32 weights
+        vl_gpt = AutoModelForCausalLM.from_pretrained(
+            model_path, 
+            trust_remote_code=True
+        ).float().eval()  # Convert all weights to float32
 
-        # Ensure attention mask is also in torch.float16
-        attention_mask = prepare_inputs["attention_mask"].to(vl_gpt.device).to(dtype=torch.float16)
+        # Generate embeddings with CPU
+        inputs_embeds = vl_gpt.prepare_inputs_embeds(**prepare_inputs)
 
-        # Debugging: Print tensor dtypes
-        print(f"Inputs Embeds dtype: {inputs_embeds.dtype}")
-        print(f"Attention Mask dtype: {attention_mask.dtype}")
-        print(f"Model dtype: {next(vl_gpt.parameters()).dtype}")
-
-        # Generate the model's response
+        # Generate response with CPU
         outputs = vl_gpt.language_model.generate(
-            inputs_embeds=inputs_embeds.to(torch.float16),
-            attention_mask=attention_mask.to(torch.float16),
+            inputs_embeds=inputs_embeds,
+            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,
@@ -76,36 +64,36 @@ def describe_image(image, user_question="You are the best AP teacher in the worl
             use_cache=True
         )
 
-        # Decode the generated tokens into text
-        answer = tokenizer.decode(outputs[0].cpu().tolist(), skip_special_tokens=True)
+        # Decode the response
+        answer = tokenizer.decode(outputs[0].tolist(), skip_special_tokens=True)
         return answer
+
     except Exception as e:
-        # Provide detailed error information
         return f"Error: {str(e)}"
+
 # Gradio interface
 def gradio_app():
     with gr.Blocks() as demo:
-        gr.Markdown("# Image Description with DeepSeek VL 1.3b 🐬\n### Upload an image and ask a question about it.")
+        gr.Markdown("# Image Description with DeepSeek VL 1.3b 🐬 (CPU Version)")
 
         with gr.Row():
             image_input = gr.Image(type="pil", label="Upload an Image")
             question_input = gr.Textbox(
                 label="Question (optional)", 
-                placeholder="Ask a question about the image (e.g., 'What is happening in this image?')",
+                placeholder="Ask a question about the image",
                 lines=2
             )
 
-        output_text = gr.Textbox(label="Solving the problem", interactive=False)
-
-        submit_btn = gr.Button("Solve")
+        output_text = gr.Textbox(label="Image Description", interactive=False)
+        submit_btn = gr.Button("Generate Description")
 
         submit_btn.click(
             fn=describe_image,
-            inputs=[image_input, question_input],  # Pass both image and question as inputs
+            inputs=[image_input, question_input],
             outputs=output_text
         )
 
     demo.launch()
 
-# Launch the Gradio app
-gradio_app()
+# Launch the app
+gradio_app()