Workpam commited on
Commit
9cbf52e
·
verified ·
1 Parent(s): 3fb0573

Update app.py

Browse files
Files changed (1) hide show
  1. app.py +28 -50
app.py CHANGED
@@ -1,64 +1,41 @@
 
1
  import gradio as gr
2
- from PIL import Image
3
- import torch
4
- from transformers import AutoProcessor, Qwen2_5_VLForConditionalGeneration, BitsAndBytesConfig
5
 
6
  # --- Configuration ---
7
- MODEL_NAME = "Qwen/Qwen2.5-VL-7B-Instruct"
8
- DEVICE = "cuda" if torch.cuda.is_available() else "cpu"
 
9
 
10
- # Quantization: 4-bit NF4 (reduces GPU memory footprint)
11
- bnb_config = BitsAndBytesConfig(
12
- load_in_4bit=True,
13
- bnb_4bit_use_double_quant=True,
14
- bnb_4bit_quant_type="nf4",
15
- bnb_4bit_compute_dtype=torch.float16
16
  )
17
 
18
- # --- Model & Processor Initialization ---
19
- processor = AutoProcessor.from_pretrained(
20
- MODEL_NAME,
21
- trust_remote_code=True
22
- )
23
- model = Qwen2_5_VLForConditionalGeneration.from_pretrained(
24
- MODEL_NAME,
25
- quantization_config=bnb_config,
26
- device_map="auto",
27
- torch_dtype=torch.float16,
28
- trust_remote_code=True
29
- )
30
-
31
- # --- Alt Text Generation Function ---
32
  def generate_alt_text(files, max_new_tokens):
33
  captions = []
34
  for f in files or []:
35
- img = Image.open(f.name).convert("RGB")
36
- prompt = "Generate a concise, detailed alt text for the image."
37
-
38
- # Preprocess both image and prompt
39
- inputs = processor(
40
- images=img,
41
- text=prompt,
42
- return_tensors="pt"
43
- ).to(model.device)
44
-
45
- # Generate with quantized model
46
- outputs = model.generate(
47
- **inputs,
48
- max_new_tokens=int(max_new_tokens)
49
  )
50
-
51
- # Decode and format
52
- text = processor.decode(outputs[0], skip_special_tokens=True).strip()
53
  if text:
54
  text = text[0].upper() + text[1:]
55
  captions.append(text)
56
  return "\n\n".join(captions)
57
 
58
  # --- Gradio Interface ---
59
- with gr.Blocks(title="Qwen2.5-VL-7B-Instruct (4-bit) Alt Text Generator") as demo:
60
- gr.Markdown("# Qwen2.5-VL-7B-Instruct (4-bit) Alt Text Generator")
61
- gr.Markdown(f"**Model:** {MODEL_NAME} (4-bit NF4 quantized)")
62
 
63
  with gr.Row():
64
  file_input = gr.File(
@@ -83,14 +60,15 @@ with gr.Blocks(title="Qwen2.5-VL-7B-Instruct (4-bit) Alt Text Generator") as dem
83
  show_copy_button=True
84
  )
85
 
86
- # Preview images
87
  file_input.change(
88
- fn=lambda files: [Image.open(f.name).convert("RGB") for f in (files or [])],
89
  inputs=file_input,
90
- outputs=gallery
 
91
  )
92
 
93
- # Generate alt text
94
  generate_btn.click(
95
  fn=generate_alt_text,
96
  inputs=[file_input, max_new_tokens],
@@ -104,4 +82,4 @@ with gr.Blocks(title="Qwen2.5-VL-7B-Instruct (4-bit) Alt Text Generator") as dem
104
  outputs=[file_input, max_new_tokens, gallery, output]
105
  )
106
 
107
- demo.launch(server_name="0.0.0.0", server_port=7860)
 
1
+ import os
2
  import gradio as gr
3
+ from huggingface_hub import InferenceApi
 
 
4
 
5
  # --- Configuration ---
6
+ # Make sure to set HF_TOKEN in your Space Settings → Secrets
7
+ hf_token = os.getenv("HF_TOKEN")
8
+ repo_id = "Qwen/Qwen2.5-VL-7B-Instruct"
9
 
10
+ # Initialize remote inference client
11
+ inference = InferenceApi(
12
+ repo_id=repo_id,
13
+ token=hf_token,
 
 
14
  )
15
 
16
+ # --- Core Generation Function ---
 
 
 
 
 
 
 
 
 
 
 
 
 
17
  def generate_alt_text(files, max_new_tokens):
18
  captions = []
19
  for f in files or []:
20
+ # Read raw image bytes
21
+ img_bytes = f.read()
22
+ # Call the hosted model
23
+ response = inference(
24
+ inputs=img_bytes,
25
+ parameters={"max_new_tokens": int(max_new_tokens)},
26
+ request_format="binary"
 
 
 
 
 
 
 
27
  )
28
+ text = response[0].get("generated_text", "").strip()
29
+ # Capitalize first letter for readability
 
30
  if text:
31
  text = text[0].upper() + text[1:]
32
  captions.append(text)
33
  return "\n\n".join(captions)
34
 
35
  # --- Gradio Interface ---
36
+ with gr.Blocks(title="Qwen2.5-VL-7B via Inference API") as demo:
37
+ gr.Markdown("# Qwen2.5-VL-7B-Instruct Alt Text Generator (Inference API)")
38
+ gr.Markdown("Uses Hugging Face Inference API, no GPU memory on the Space required.")
39
 
40
  with gr.Row():
41
  file_input = gr.File(
 
60
  show_copy_button=True
61
  )
62
 
63
+ # Show image previews
64
  file_input.change(
65
+ fn=lambda files: [f.read() and None or None for f in (files or [])],
66
  inputs=file_input,
67
+ outputs=gallery,
68
+ preprocess=False
69
  )
70
 
71
+ # Generate captions
72
  generate_btn.click(
73
  fn=generate_alt_text,
74
  inputs=[file_input, max_new_tokens],
 
82
  outputs=[file_input, max_new_tokens, gallery, output]
83
  )
84
 
85
+ demo.launch()