Update app.py

app.py

@@ -1,100 +1,73 @@
 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)
 
-# Model and Processor Initialization
 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()
+    "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)
 }
 
-# Default question
-default_question = (
-    "You are an image-to-prompt converter. Your work is to observe each and every detail of the image and "
-    "craft a detailed prompt under 100 words in this format: [image content/subject, description of action, state, "
-    "and mood], [art form, style], [artist/photographer reference if needed], [additional settings such as camera "
-    "and lens settings, lighting, colors, effects, texture, background, rendering]."
-)
+DESCRIPTION = "[Phi-3.5-vision Demo](https://huggingface.co/microsoft/Phi-3.5-vision-instruct)"
 
-# Function to generate prompt
-def generate_caption(image):
-    model = models["microsoft/Phi-3.5-vision-instruct"]
-    processor = processors["microsoft/Phi-3.5-vision-instruct"]
+kwargs = {}
+kwargs['torch_dtype'] = torch.bfloat16
 
-    prompt = f"<|user|>\n<|image_1|>\n{default_question}<|end|>\n<|assistant|>\n"
+user_prompt = '<|user|>\n'
+assistant_prompt = '<|assistant|>\n'
+prompt_suffix = "<|end|>\n"
+
+default_question = "You are an image to prompt converter. Your work is to observe each and every detail of the image and craft a detailed prompt under 100 words in this format: [image content/subject, description of action, state, and mood], [art form, style], [artist/photographer reference if needed], [additional settings such as camera and lens settings, lighting, colors, effects, texture, background, rendering]."
+
+@spaces.GPU
+def run_example(image, text_input=default_question, 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")
 
     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
 
-# Enhanced CSS for streamlined UI
 css = """
-  #container {
-    background-color: #f9f9f9;
-    padding: 20px;
-    border-radius: 15px;
-    border: 2px solid #333;
-    box-shadow: 0 4px 8px rgba(0, 0, 0, 0.2);
-    max-width: 450px;
-    margin: auto;
-  }
-  #input_image {
-    margin-top: 15px;
-    border: 2px solid #333;
-    border-radius: 8px;
-    height: 180px;
-    object-fit: contain;
-  }
-  #output_caption {
-    margin-top: 15px;
-    border: 2px solid #333;
-    border-radius: 8px;
-    height: 180px;
-    overflow-y: auto;
-  }
-  #run_button {
-    background-color: #fff;
-    color: black;
-    border-radius: 10px;
-    padding: 10px;
-    cursor: pointer;
-    transition: background-color 0.3s ease;
-    margin-top: 15px;
+  #output {
+    height: 500px; 
+    overflow: auto; 
+    border: 1px solid #ccc; 
   }
-  #run_button:hover {
-    background-color: #333;
+  #model_selector, #text_input { 
+    display: none !important; 
   }
 """
 
-# Gradio Interface with Adjustments
 with gr.Blocks(css=css) as demo:
-    with gr.Column(elem_id="container"):
-        input_image = gr.Image(type="pil", elem_id="input_image", label="Upload Image")
-        run_button = gr.Button(value="Generate Prompt", elem_id="run_button")
-        output_caption = gr.Textbox(label="Generated Prompt", show_copy_button=True, elem_id="output_caption", lines=6)
+    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", visible=False)
+                text_input = gr.Textbox(label="Question", value=default_question, visible=False)
+                submit_btn = gr.Button(value="Submit")
+            with gr.Column():
+                output_text = gr.Textbox(label="Output Text")
 
-    run_button.click(
-        fn=generate_caption,
-        inputs=[input_image],
-        outputs=output_caption,
-    )
+        submit_btn.click(run_example, [input_img, text_input, model_selector], [output_text])
 
-demo.launch(share=False)
+demo.queue(api_open=False)
+demo.launch(debug=True, show_api=False)