Update app.py

app.py

@@ -4,77 +4,84 @@ from transformers import AutoModelForCausalLM, AutoProcessor
 import torch
 from PIL import Image
 import subprocess
-# Ensure flash-attn is installed but skip CUDA build (to avoid 'flash_attn_2_cuda' error)
+
+# Install flash-attn but skip CUDA build (to avoid 'flash_attn_2_cuda' error)
 subprocess.run(
-    'pip install flash-attn --no-build-isolation', 
-    env={'FLASH_ATTENTION_SKIP_CUDA_BUILD': "TRUE"}, 
+    'pip install flash-attn --no-build-isolation',
+    env={'FLASH_ATTENTION_SKIP_CUDA_BUILD': "TRUE"},
     shell=True
 )
 
-models = {
-    "microsoft/Phi-3.5-vision-instruct": AutoModelForCausalLM.from_pretrained("microsoft/Phi-3.5-vision-instruct", trust_remote_code=True, torch_dtype="auto", _attn_implementation="flash_attention_2").cuda().eval()
-
-}
-
-processors = {
-    "microsoft/Phi-3.5-vision-instruct": AutoProcessor.from_pretrained("microsoft/Phi-3.5-vision-instruct", trust_remote_code=True)
-}
-
-kwargs = {}
-kwargs['torch_dtype'] = torch.bfloat16
-
 user_prompt = '<|user|>\n'
 assistant_prompt = '<|assistant|>\n'
 prompt_suffix = "<|end|>\n"
 
-@spaces.GPU
-def run_example(image, text_input=None, model_id="microsoft/Phi-3.5-vision-instruct"):
-    model = models[model_id]
-    processor = processors[model_id]
+model_name = "microsoft/Phi-3.5-vision-instruct"
+
+# Lazy-load the model and processor at runtime
+def get_model_and_processor(model_id):
+    model = AutoModelForCausalLM.from_pretrained(
+        model_id,
+        trust_remote_code=True,
+        torch_dtype=torch.bfloat16  # safer than 'auto'
+    ).cuda().eval()
+    processor = AutoProcessor.from_pretrained(
+        model_id,
+        trust_remote_code=True
+    )
+    return model, processor
 
+@spaces.GPU
+def run_example(image, text_input=None, model_id=model_name):
+    model, processor = get_model_and_processor(model_id)
     prompt = f"{user_prompt}<|image_1|>\n{text_input}{prompt_suffix}{assistant_prompt}"
     image = Image.fromarray(image).convert("RGB")
 
     inputs = processor(prompt, image, return_tensors="pt").to("cuda:0")
-    generate_ids = model.generate(**inputs, 
-                                max_new_tokens=1000,
-                                eos_token_id=processor.tokenizer.eos_token_id,
-                                )
+    generate_ids = model.generate(
+        **inputs,
+        max_new_tokens=1000,
+        eos_token_id=processor.tokenizer.eos_token_id
+    )
     generate_ids = generate_ids[:, inputs['input_ids'].shape[1]:]
-    response = processor.batch_decode(generate_ids, 
-                                    skip_special_tokens=True, 
-                                    clean_up_tokenization_spaces=False)[0]
+    response = processor.batch_decode(
+        generate_ids,
+        skip_special_tokens=True,
+        clean_up_tokenization_spaces=False
+    )[0]
     return response
 
 css = """
   #output {
-    height: 500px; 
-    overflow: auto; 
-    border: 1px solid #ccc; 
+    height: 500px;
+    overflow: auto;
+    border: 1px solid #ccc;
   }
 """
 
 with gr.Blocks(css=css) as demo:
     gr.Markdown("## Phi-3.5 Vision Instruct Demo with Example Inputs")
-    
+
     with gr.Tab(label="Phi-3.5 Input"):
         with gr.Row():
             with gr.Column():
                 input_img = gr.Image(label="Input Picture")
-                model_selector = gr.Dropdown(choices=list(models.keys()), label="Model", value="microsoft/Phi-3.5-vision-instruct")
+                model_selector = gr.Dropdown(
+                    choices=[model_name],
+                    label="Model",
+                    value=model_name
+                )
                 text_input = gr.Textbox(label="Question")
                 submit_btn = gr.Button(value="Submit")
             with gr.Column():
                 output_text = gr.Textbox(label="Output Text")
 
-        # Example images and text
         examples = [
             ["image1.jpeg", "What does this painting tell us explain in detail?"],
             ["image2.jpg", "What does this painting tell us explain in detail?"],
             ["image3.jpg", "Describe the scene in this picture."]
         ]
 
-        # Adding Examples
         gr.Examples(
             examples=examples,
             inputs=[input_img, text_input],
@@ -83,6 +90,5 @@ with gr.Blocks(css=css) as demo:
 
         submit_btn.click(run_example, [input_img, text_input, model_selector], [output_text])
 
-# Launch the demo
 demo.queue(api_open=False)
-demo.launch(debug=True, show_api=False)
+demo.launch(debug=True, show_api=False)