snjev310 commited on
Commit
ab98cf2
·
verified ·
1 Parent(s): 58fff40

Create app.py

Browse files
Files changed (1) hide show
  1. app.py +104 -0
app.py ADDED
@@ -0,0 +1,104 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ import gradio as gr
2
+ import torch
3
+ import spaces
4
+ from transformers import AutoModelForSeq2SeqLM, AutoTokenizer
5
+ from peft import PeftModel
6
+
7
+ # --- Configuration ---
8
+ BASE_MODEL_ID = "CohereForAI/aya-101"
9
+
10
+ # Map the dropdown options to your 3 Hugging Face Model IDs
11
+ MODEL_MAP = {
12
+ "English to Angika": "snjev310/aya-101-english-angika",
13
+ "Hindi to Angika": "snjev310/aya-101-hindi-angika",
14
+ "Angika to English": "snjev310/aya-101-angika-english"
15
+ }
16
+
17
+ # Load Tokenizer globally (it's small and stays in CPU RAM)
18
+ tokenizer = AutoTokenizer.from_pretrained(BASE_MODEL_ID)
19
+
20
+ @spaces.GPU(duration=180) # 3 minutes to allow for 13B model loading + inference
21
+ def translate(text, model_choice):
22
+ if not text.strip():
23
+ return "Please enter text to translate."
24
+
25
+ adapter_id = MODEL_MAP[model_choice]
26
+
27
+ # 1. Load Base Model in bfloat16 (Standard for Aya-101)
28
+ # Pro ZeroGPU has ~70GB VRAM, so we don't need 4-bit quantization
29
+ base_model = AutoModelForSeq2SeqLM.from_pretrained(
30
+ BASE_MODEL_ID,
31
+ torch_dtype=torch.bfloat16,
32
+ low_cpu_mem_usage=True,
33
+ device_map="auto"
34
+ )
35
+
36
+ # 2. Load the specific PEFT Adapter
37
+ model = PeftModel.from_pretrained(base_model, adapter_id)
38
+ model.eval()
39
+
40
+ # 3. Prepare Input
41
+ # Using a prompt format helps the model understand the task
42
+ prompt = f"{model_choice}: {text}"
43
+ inputs = tokenizer(prompt, return_tensors="pt").to(model.device)
44
+
45
+ # 4. Generate
46
+ with torch.no_grad():
47
+ outputs = model.generate(
48
+ **inputs,
49
+ max_new_tokens=256,
50
+ do_sample=True,
51
+ temperature=0.3,
52
+ top_p=0.9
53
+ )
54
+
55
+ result = tokenizer.decode(outputs[0], skip_special_tokens=True)
56
+
57
+ # 5. Cleanup (CRITICAL for ZeroGPU to release resources)
58
+ del model
59
+ del base_model
60
+ torch.cuda.empty_cache()
61
+
62
+ return result
63
+
64
+ # --- Gradio UI ---
65
+ with gr.Blocks(theme=gr.themes.Soft()) as demo:
66
+ gr.Markdown("# 🗣️ Angika Multi-Translator")
67
+ gr.Markdown("Powered by **Aya-101** and **ZeroGPU**. Select your translation direction below.")
68
+
69
+ with gr.Row():
70
+ with gr.Column():
71
+ model_dropdown = gr.Dropdown(
72
+ choices=list(MODEL_MAP.keys()),
73
+ value="English to Angika",
74
+ label="Select Translation Mode"
75
+ )
76
+ input_text = gr.Textbox(
77
+ label="Input Text",
78
+ placeholder="Type here...",
79
+ lines=5
80
+ )
81
+ submit_btn = gr.Button("Translate", variant="primary")
82
+
83
+ with gr.Column():
84
+ output_text = gr.Textbox(
85
+ label="Translated Text",
86
+ lines=5,
87
+ interactive=False
88
+ )
89
+
90
+ submit_btn.click(
91
+ fn=translate,
92
+ inputs=[input_text, model_dropdown],
93
+ outputs=output_text
94
+ )
95
+
96
+ gr.Examples(
97
+ examples=[
98
+ ["How are you doing today?", "English to Angika"],
99
+ ["आप कैसे हैं?", "Hindi to Angika"],
100
+ ],
101
+ inputs=[input_text, model_dropdown]
102
+ )
103
+
104
+ demo.launch()