Update app.py (#5)

app.py

@@ -1,27 +1,46 @@
 import gradio as gr
+import spaces
 from transformers import AutoModelForCausalLM, AutoProcessor
+import torch
 from PIL import Image
+import subprocess
+subprocess.run('pip install flash-attn --no-build-isolation', env={'FLASH_ATTENTION_SKIP_CUDA_BUILD': "TRUE"}, shell=True)
 
-# Define constants
-MODEL_NAME = "microsoft/Phi-3.5-vision-instruct"
-DESCRIPTION = "# [Phi-3.5-vision Demo](https://huggingface.co/microsoft/Phi-3.5-vision-instruct)"
-DEVICE = "cuda"
+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()
 
-# Load model and processor
-model = AutoModelForCausalLM.from_pretrained(MODEL_NAME, trust_remote_code=True, torch_dtype="auto").to(DEVICE).eval()
-processor = AutoProcessor.from_pretrained(MODEL_NAME, trust_remote_code=True)
+}
 
-def run_example(image, text_input, model_id):
-    # Prepare prompt and image for processing
-    prompt = f"{text_input}\n"
+processors = {
+    "microsoft/Phi-3.5-vision-instruct": AutoProcessor.from_pretrained("microsoft/Phi-3.5-vision-instruct", trust_remote_code=True)
+}
+
+DESCRIPTION = "[Phi-3.5-vision Demo](https://huggingface.co/microsoft/Phi-3.5-vision-instruct)"
+
+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]
+
+    prompt = f"{user_prompt}<|image_1|>\n{text_input}{prompt_suffix}{assistant_prompt}"
     image = Image.fromarray(image).convert("RGB")
-    
-    # Process input
-    inputs = processor(prompt, image, return_tensors="pt").to(DEVICE)
-    generate_ids = model.generate(**inputs, max_new_tokens=1000, eos_token_id=processor.tokenizer.eos_token_id)
+
+    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 = 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 = """
@@ -32,17 +51,18 @@ css = """
   }
 """
 
-# Set up the Gradio interface
 with gr.Blocks(css=css) as demo:
     gr.Markdown(DESCRIPTION)
     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")
                 text_input = gr.Textbox(label="Question")
                 submit_btn = gr.Button(value="Submit")
             with gr.Column():
                 output_text = gr.Textbox(label="Output Text")
-        submit_btn.click(run_example, inputs=[input_img, text_input, MODEL_NAME], outputs=output_text)
+
+        submit_btn.click(run_example, [input_img, text_input, model_selector], [output_text])
 
 demo.launch(debug=True)