MaziyarPanahi commited on
Commit
21fcfe6
1 Parent(s): 153a98d
Files changed (1) hide show
  1. app.py +38 -18
app.py CHANGED
@@ -1,27 +1,46 @@
1
  import gradio as gr
 
2
  from transformers import AutoModelForCausalLM, AutoProcessor
 
3
  from PIL import Image
 
 
4
 
5
- # Define constants
6
- MODEL_NAME = "microsoft/Phi-3.5-vision-instruct"
7
- DESCRIPTION = "# [Phi-3.5-vision Demo](https://huggingface.co/microsoft/Phi-3.5-vision-instruct)"
8
- DEVICE = "cuda"
9
 
10
- # Load model and processor
11
- model = AutoModelForCausalLM.from_pretrained(MODEL_NAME, trust_remote_code=True, torch_dtype="auto").to(DEVICE).eval()
12
- processor = AutoProcessor.from_pretrained(MODEL_NAME, trust_remote_code=True)
13
 
14
- def run_example(image, text_input, model_id):
15
- # Prepare prompt and image for processing
16
- prompt = f"{text_input}\n"
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
17
  image = Image.fromarray(image).convert("RGB")
18
-
19
- # Process input
20
- inputs = processor(prompt, image, return_tensors="pt").to(DEVICE)
21
- generate_ids = model.generate(**inputs, max_new_tokens=1000, eos_token_id=processor.tokenizer.eos_token_id)
 
 
22
  generate_ids = generate_ids[:, inputs['input_ids'].shape[1]:]
23
- response = processor.batch_decode(generate_ids, skip_special_tokens=True, clean_up_tokenization_spaces=False)[0]
24
-
 
25
  return response
26
 
27
  css = """
@@ -32,17 +51,18 @@ css = """
32
  }
33
  """
34
 
35
- # Set up the Gradio interface
36
  with gr.Blocks(css=css) as demo:
37
  gr.Markdown(DESCRIPTION)
38
  with gr.Tab(label="Phi-3.5 Input"):
39
  with gr.Row():
40
  with gr.Column():
41
  input_img = gr.Image(label="Input Picture")
 
42
  text_input = gr.Textbox(label="Question")
43
  submit_btn = gr.Button(value="Submit")
44
  with gr.Column():
45
  output_text = gr.Textbox(label="Output Text")
46
- submit_btn.click(run_example, inputs=[input_img, text_input, MODEL_NAME], outputs=output_text)
 
47
 
48
  demo.launch(debug=True)
 
1
  import gradio as gr
2
+ import spaces
3
  from transformers import AutoModelForCausalLM, AutoProcessor
4
+ import torch
5
  from PIL import Image
6
+ import subprocess
7
+ subprocess.run('pip install flash-attn --no-build-isolation', env={'FLASH_ATTENTION_SKIP_CUDA_BUILD': "TRUE"}, shell=True)
8
 
9
+ models = {
10
+ "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()
 
 
11
 
12
+ }
 
 
13
 
14
+ processors = {
15
+ "microsoft/Phi-3.5-vision-instruct": AutoProcessor.from_pretrained("microsoft/Phi-3.5-vision-instruct", trust_remote_code=True)
16
+ }
17
+
18
+ DESCRIPTION = "[Phi-3.5-vision Demo](https://huggingface.co/microsoft/Phi-3.5-vision-instruct)"
19
+
20
+ kwargs = {}
21
+ kwargs['torch_dtype'] = torch.bfloat16
22
+
23
+ user_prompt = '<|user|>\n'
24
+ assistant_prompt = '<|assistant|>\n'
25
+ prompt_suffix = "<|end|>\n"
26
+
27
+ @spaces.GPU
28
+ def run_example(image, text_input=None, model_id="microsoft/Phi-3.5-vision-instruct"):
29
+ model = models[model_id]
30
+ processor = processors[model_id]
31
+
32
+ prompt = f"{user_prompt}<|image_1|>\n{text_input}{prompt_suffix}{assistant_prompt}"
33
  image = Image.fromarray(image).convert("RGB")
34
+
35
+ inputs = processor(prompt, image, return_tensors="pt").to("cuda:0")
36
+ generate_ids = model.generate(**inputs,
37
+ max_new_tokens=1000,
38
+ eos_token_id=processor.tokenizer.eos_token_id,
39
+ )
40
  generate_ids = generate_ids[:, inputs['input_ids'].shape[1]:]
41
+ response = processor.batch_decode(generate_ids,
42
+ skip_special_tokens=True,
43
+ clean_up_tokenization_spaces=False)[0]
44
  return response
45
 
46
  css = """
 
51
  }
52
  """
53
 
 
54
  with gr.Blocks(css=css) as demo:
55
  gr.Markdown(DESCRIPTION)
56
  with gr.Tab(label="Phi-3.5 Input"):
57
  with gr.Row():
58
  with gr.Column():
59
  input_img = gr.Image(label="Input Picture")
60
+ model_selector = gr.Dropdown(choices=list(models.keys()), label="Model", value="microsoft/Phi-3.5-vision-instruct")
61
  text_input = gr.Textbox(label="Question")
62
  submit_btn = gr.Button(value="Submit")
63
  with gr.Column():
64
  output_text = gr.Textbox(label="Output Text")
65
+
66
+ submit_btn.click(run_example, [input_img, text_input, model_selector], [output_text])
67
 
68
  demo.launch(debug=True)