Spaces:
Runtime error
Runtime error
Update app.py
Browse files
app.py
CHANGED
@@ -1,4 +1,3 @@
|
|
1 |
-
import spaces
|
2 |
import gradio as gr
|
3 |
import torch
|
4 |
from transformers import AutoModelForSeq2SeqLM, AutoTokenizer
|
@@ -11,7 +10,11 @@ print("Loading IndicBART tokenizer...")
|
|
11 |
tokenizer = AutoTokenizer.from_pretrained(model_name, do_lower_case=False, use_fast=False, keep_accents=True)
|
12 |
|
13 |
print("Loading IndicBART model on CPU...")
|
14 |
-
model = AutoModelForSeq2SeqLM.from_pretrained(
|
|
|
|
|
|
|
|
|
15 |
|
16 |
# Language mapping
|
17 |
LANGUAGE_CODES = {
|
@@ -29,11 +32,8 @@ LANGUAGE_CODES = {
|
|
29 |
"Telugu": "<2te>"
|
30 |
}
|
31 |
|
32 |
-
@spaces.GPU(duration=60)
|
33 |
def generate_response(input_text, source_lang, target_lang, task_type, max_length):
|
34 |
-
"""Generate response using IndicBART"""
|
35 |
-
device = "cuda" if torch.cuda.is_available() else "cpu"
|
36 |
-
model_gpu = model.to(device)
|
37 |
|
38 |
# Get language codes
|
39 |
src_code = LANGUAGE_CODES[source_lang]
|
@@ -51,43 +51,45 @@ def generate_response(input_text, source_lang, target_lang, task_type, max_lengt
|
|
51 |
formatted_input = f"{input_text} </s> {src_code}"
|
52 |
decoder_start_token = tgt_code
|
53 |
|
54 |
-
# Tokenize input
|
55 |
inputs = tokenizer(formatted_input, return_tensors="pt", padding=True, truncation=True, max_length=512)
|
56 |
-
inputs = {k: v.to(device) for k, v in inputs.items()}
|
57 |
|
58 |
# Get decoder start token id
|
59 |
-
|
|
|
|
|
|
|
|
|
60 |
|
61 |
-
# Generate
|
62 |
with torch.no_grad():
|
63 |
-
outputs =
|
64 |
**inputs,
|
65 |
decoder_start_token_id=decoder_start_token_id,
|
66 |
max_length=max_length,
|
67 |
-
num_beams=
|
68 |
early_stopping=True,
|
69 |
pad_token_id=tokenizer.pad_token_id,
|
70 |
eos_token_id=tokenizer.eos_token_id,
|
71 |
-
use_cache=True
|
|
|
72 |
)
|
73 |
|
74 |
# Decode output
|
75 |
generated_text = tokenizer.decode(outputs[0], skip_special_tokens=True, clean_up_tokenization_spaces=False)
|
76 |
|
77 |
-
# Move model back to CPU
|
78 |
-
model_gpu.cpu()
|
79 |
-
torch.cuda.empty_cache()
|
80 |
-
|
81 |
return generated_text
|
82 |
|
83 |
# Create Gradio interface
|
84 |
-
with gr.Blocks(title="IndicBART Multilingual Assistant", theme=gr.themes.Soft()) as demo:
|
85 |
gr.Markdown("""
|
86 |
-
# 🇮🇳 IndicBART Multilingual Assistant
|
87 |
|
88 |
Experience IndicBART - trained on **11 Indian languages**! Perfect for translation, text completion, and multilingual generation.
|
89 |
|
90 |
**Supported Languages**: Assamese, Bengali, Gujarati, Hindi, Kannada, Malayalam, Marathi, Oriya, Punjabi, Tamil, Telugu, English
|
|
|
|
|
91 |
""")
|
92 |
|
93 |
with gr.Row():
|
@@ -104,7 +106,9 @@ with gr.Blocks(title="IndicBART Multilingual Assistant", theme=gr.themes.Soft())
|
|
104 |
interactive=False
|
105 |
)
|
106 |
|
107 |
-
|
|
|
|
|
108 |
|
109 |
with gr.Column(scale=1):
|
110 |
task_type = gr.Dropdown(
|
@@ -126,9 +130,9 @@ with gr.Blocks(title="IndicBART Multilingual Assistant", theme=gr.themes.Soft())
|
|
126 |
)
|
127 |
|
128 |
max_length = gr.Slider(
|
129 |
-
minimum=
|
130 |
-
maximum=
|
131 |
-
value=
|
132 |
step=10,
|
133 |
label="Max Length"
|
134 |
)
|
@@ -137,11 +141,11 @@ with gr.Blocks(title="IndicBART Multilingual Assistant", theme=gr.themes.Soft())
|
|
137 |
gr.Markdown("### 💡 Try these examples:")
|
138 |
|
139 |
examples = [
|
140 |
-
["Hello, how are you?", "English", "Hindi", "Translation",
|
141 |
-
["मैं एक छात्र हूं", "Hindi", "English", "Translation",
|
142 |
-
["আমি ভাত খাই", "Bengali", "English", "Translation",
|
143 |
-
["भारत एक", "Hindi", "Hindi", "Text Completion",
|
144 |
-
["The capital of India", "English", "English", "Text Completion",
|
145 |
]
|
146 |
|
147 |
gr.Examples(
|
@@ -151,12 +155,21 @@ with gr.Blocks(title="IndicBART Multilingual Assistant", theme=gr.themes.Soft())
|
|
151 |
fn=generate_response
|
152 |
)
|
153 |
|
154 |
-
#
|
|
|
|
|
|
|
|
|
155 |
generate_btn.click(
|
156 |
generate_response,
|
157 |
inputs=[input_text, source_lang, target_lang, task_type, max_length],
|
158 |
outputs=output_text
|
159 |
)
|
|
|
|
|
|
|
|
|
|
|
160 |
|
161 |
if __name__ == "__main__":
|
162 |
-
demo.launch()
|
|
|
|
|
1 |
import gradio as gr
|
2 |
import torch
|
3 |
from transformers import AutoModelForSeq2SeqLM, AutoTokenizer
|
|
|
10 |
tokenizer = AutoTokenizer.from_pretrained(model_name, do_lower_case=False, use_fast=False, keep_accents=True)
|
11 |
|
12 |
print("Loading IndicBART model on CPU...")
|
13 |
+
model = AutoModelForSeq2SeqLM.from_pretrained(
|
14 |
+
model_name,
|
15 |
+
torch_dtype=torch.float32, # Use float32 for better CPU performance
|
16 |
+
device_map="cpu"
|
17 |
+
)
|
18 |
|
19 |
# Language mapping
|
20 |
LANGUAGE_CODES = {
|
|
|
32 |
"Telugu": "<2te>"
|
33 |
}
|
34 |
|
|
|
35 |
def generate_response(input_text, source_lang, target_lang, task_type, max_length):
|
36 |
+
"""Generate response using IndicBART on CPU"""
|
|
|
|
|
37 |
|
38 |
# Get language codes
|
39 |
src_code = LANGUAGE_CODES[source_lang]
|
|
|
51 |
formatted_input = f"{input_text} </s> {src_code}"
|
52 |
decoder_start_token = tgt_code
|
53 |
|
54 |
+
# Tokenize input (keep on CPU)
|
55 |
inputs = tokenizer(formatted_input, return_tensors="pt", padding=True, truncation=True, max_length=512)
|
|
|
56 |
|
57 |
# Get decoder start token id
|
58 |
+
try:
|
59 |
+
decoder_start_token_id = tokenizer._convert_token_to_id_with_added_voc(decoder_start_token)
|
60 |
+
except:
|
61 |
+
# Fallback if the method doesn't exist
|
62 |
+
decoder_start_token_id = tokenizer.convert_tokens_to_ids(decoder_start_token)
|
63 |
|
64 |
+
# Generate on CPU
|
65 |
with torch.no_grad():
|
66 |
+
outputs = model.generate(
|
67 |
**inputs,
|
68 |
decoder_start_token_id=decoder_start_token_id,
|
69 |
max_length=max_length,
|
70 |
+
num_beams=2, # Reduced for faster CPU inference
|
71 |
early_stopping=True,
|
72 |
pad_token_id=tokenizer.pad_token_id,
|
73 |
eos_token_id=tokenizer.eos_token_id,
|
74 |
+
use_cache=True,
|
75 |
+
do_sample=False # Deterministic for CPU
|
76 |
)
|
77 |
|
78 |
# Decode output
|
79 |
generated_text = tokenizer.decode(outputs[0], skip_special_tokens=True, clean_up_tokenization_spaces=False)
|
80 |
|
|
|
|
|
|
|
|
|
81 |
return generated_text
|
82 |
|
83 |
# Create Gradio interface
|
84 |
+
with gr.Blocks(title="IndicBART CPU Multilingual Assistant", theme=gr.themes.Soft()) as demo:
|
85 |
gr.Markdown("""
|
86 |
+
# 🇮🇳 IndicBART Multilingual Assistant (CPU Version)
|
87 |
|
88 |
Experience IndicBART - trained on **11 Indian languages**! Perfect for translation, text completion, and multilingual generation.
|
89 |
|
90 |
**Supported Languages**: Assamese, Bengali, Gujarati, Hindi, Kannada, Malayalam, Marathi, Oriya, Punjabi, Tamil, Telugu, English
|
91 |
+
|
92 |
+
*Note: Running on CPU - responses may take longer than GPU version.*
|
93 |
""")
|
94 |
|
95 |
with gr.Row():
|
|
|
106 |
interactive=False
|
107 |
)
|
108 |
|
109 |
+
with gr.Row():
|
110 |
+
generate_btn = gr.Button("Generate", variant="primary", size="lg")
|
111 |
+
clear_btn = gr.Button("Clear", variant="secondary")
|
112 |
|
113 |
with gr.Column(scale=1):
|
114 |
task_type = gr.Dropdown(
|
|
|
130 |
)
|
131 |
|
132 |
max_length = gr.Slider(
|
133 |
+
minimum=20,
|
134 |
+
maximum=200, # Reduced for faster CPU processing
|
135 |
+
value=80,
|
136 |
step=10,
|
137 |
label="Max Length"
|
138 |
)
|
|
|
141 |
gr.Markdown("### 💡 Try these examples:")
|
142 |
|
143 |
examples = [
|
144 |
+
["Hello, how are you?", "English", "Hindi", "Translation", 80],
|
145 |
+
["मैं एक छात्र हूं", "Hindi", "English", "Translation", 80],
|
146 |
+
["আমি ভাত খাই", "Bengali", "English", "Translation", 80],
|
147 |
+
["भारत एक", "Hindi", "Hindi", "Text Completion", 100],
|
148 |
+
["The capital of India", "English", "English", "Text Completion", 80]
|
149 |
]
|
150 |
|
151 |
gr.Examples(
|
|
|
155 |
fn=generate_response
|
156 |
)
|
157 |
|
158 |
+
# Event handlers
|
159 |
+
def clear_fields():
|
160 |
+
return "", ""
|
161 |
+
|
162 |
+
# Connect buttons
|
163 |
generate_btn.click(
|
164 |
generate_response,
|
165 |
inputs=[input_text, source_lang, target_lang, task_type, max_length],
|
166 |
outputs=output_text
|
167 |
)
|
168 |
+
|
169 |
+
clear_btn.click(
|
170 |
+
clear_fields,
|
171 |
+
outputs=[input_text, output_text]
|
172 |
+
)
|
173 |
|
174 |
if __name__ == "__main__":
|
175 |
+
demo.launch(share=True) # Added share=True for easier access
|