Spaces:
Sleeping
Sleeping
Update app.py
Browse files
app.py
CHANGED
@@ -5,27 +5,31 @@ import json
|
|
5 |
import requests
|
6 |
|
7 |
import gradio as gr
|
8 |
-
import spaces
|
9 |
import torch
|
10 |
import transformers
|
11 |
from transformers import AutoModelForCausalLM, AutoTokenizer, TextIteratorStreamer
|
12 |
|
|
|
13 |
DESCRIPTION = """\
|
14 |
# Zero GPU Model Comparison Arena
|
15 |
Select two different models from the dropdowns and see how they perform on the same input.
|
16 |
"""
|
17 |
|
18 |
-
|
19 |
-
|
|
|
20 |
MAX_INPUT_TOKEN_LENGTH = int(os.getenv("MAX_INPUT_TOKEN_LENGTH", "4096"))
|
21 |
|
|
|
22 |
device = torch.device("cuda:0" if torch.cuda.is_available() else "cpu")
|
23 |
|
|
|
24 |
MODEL_OPTIONS = [
|
25 |
"sarvamai/OpenHathi-7B-Hi-v0.1-Base",
|
26 |
"TokenBender/Navarna_v0_1_OpenHermes_Hindi"
|
27 |
]
|
28 |
|
|
|
29 |
models = {}
|
30 |
tokenizers = {}
|
31 |
|
@@ -42,6 +46,7 @@ for model_id in MODEL_OPTIONS:
|
|
42 |
if tokenizers[model_id].pad_token_id is None:
|
43 |
tokenizers[model_id].pad_token_id = tokenizers[model_id].eos_token_id
|
44 |
|
|
|
45 |
def log_comparison(model1_name: str, model2_name: str, question: str, answer1: str, answer2: str, winner: str = None):
|
46 |
log_data = {
|
47 |
"question": question,
|
@@ -60,13 +65,27 @@ def log_comparison(model1_name: str, model2_name: str, question: str, answer1: s
|
|
60 |
except requests.RequestException as e:
|
61 |
print(f"Error sending log to server: {e}")
|
62 |
|
63 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
64 |
def generate(
|
65 |
model_id: str,
|
66 |
message: str,
|
67 |
chat_history: List[Tuple[str, str]],
|
68 |
max_new_tokens: int = DEFAULT_MAX_NEW_TOKENS,
|
69 |
-
temperature: float = 0.
|
70 |
top_p: float = 0.95,
|
71 |
) -> Iterator[str]:
|
72 |
model = models[model_id]
|
@@ -99,6 +118,7 @@ def generate(
|
|
99 |
outputs.append(text)
|
100 |
yield "".join(outputs)
|
101 |
|
|
|
102 |
def compare_models(
|
103 |
model1_name: str,
|
104 |
model2_name: str,
|
@@ -123,13 +143,13 @@ def compare_models(
|
|
123 |
|
124 |
return chat_history1, chat_history2, chat_history1, chat_history2
|
125 |
|
126 |
-
|
127 |
-
|
128 |
def vote_better(model1_name, model2_name, question, answer1, answer2, choice):
|
129 |
winner = model1_name if choice == "Model 1" else model2_name
|
130 |
log_comparison(model1_name, model2_name, question, answer1, answer2, winner)
|
131 |
return f"You voted that {winner} performs better. This has been logged."
|
132 |
|
|
|
133 |
with gr.Blocks(css="style.css") as demo:
|
134 |
gr.Markdown(DESCRIPTION)
|
135 |
|
@@ -174,9 +194,6 @@ with gr.Blocks(css="style.css") as demo:
|
|
174 |
outputs=[vote_output]
|
175 |
)
|
176 |
|
|
|
177 |
if __name__ == "__main__":
|
178 |
-
|
179 |
-
|
180 |
-
|
181 |
-
# Start Gradio app with public link
|
182 |
-
demo.queue(max_size=3).launch(share=True)
|
|
|
5 |
import requests
|
6 |
|
7 |
import gradio as gr
|
|
|
8 |
import torch
|
9 |
import transformers
|
10 |
from transformers import AutoModelForCausalLM, AutoTokenizer, TextIteratorStreamer
|
11 |
|
12 |
+
# Description for the Gradio Interface
|
13 |
DESCRIPTION = """\
|
14 |
# Zero GPU Model Comparison Arena
|
15 |
Select two different models from the dropdowns and see how they perform on the same input.
|
16 |
"""
|
17 |
|
18 |
+
# Constants
|
19 |
+
MAX_MAX_NEW_TOKENS = 256
|
20 |
+
DEFAULT_MAX_NEW_TOKENS = 128
|
21 |
MAX_INPUT_TOKEN_LENGTH = int(os.getenv("MAX_INPUT_TOKEN_LENGTH", "4096"))
|
22 |
|
23 |
+
# Device configuration
|
24 |
device = torch.device("cuda:0" if torch.cuda.is_available() else "cpu")
|
25 |
|
26 |
+
# Model options
|
27 |
MODEL_OPTIONS = [
|
28 |
"sarvamai/OpenHathi-7B-Hi-v0.1-Base",
|
29 |
"TokenBender/Navarna_v0_1_OpenHermes_Hindi"
|
30 |
]
|
31 |
|
32 |
+
# Load models and tokenizers
|
33 |
models = {}
|
34 |
tokenizers = {}
|
35 |
|
|
|
46 |
if tokenizers[model_id].pad_token_id is None:
|
47 |
tokenizers[model_id].pad_token_id = tokenizers[model_id].eos_token_id
|
48 |
|
49 |
+
# Function to log comparisons
|
50 |
def log_comparison(model1_name: str, model2_name: str, question: str, answer1: str, answer2: str, winner: str = None):
|
51 |
log_data = {
|
52 |
"question": question,
|
|
|
65 |
except requests.RequestException as e:
|
66 |
print(f"Error sending log to server: {e}")
|
67 |
|
68 |
+
# Function to prepare input
|
69 |
+
def prepare_input(model_id: str, message: str, chat_history: List[Tuple[str, str]]):
|
70 |
+
tokenizer = tokenizers[model_id]
|
71 |
+
# Prepare inputs for the model
|
72 |
+
inputs = tokenizer(
|
73 |
+
[x[1] for x in chat_history] + [message],
|
74 |
+
return_tensors="pt",
|
75 |
+
truncation=True,
|
76 |
+
padding=True,
|
77 |
+
max_length=MAX_INPUT_TOKEN_LENGTH,
|
78 |
+
)
|
79 |
+
return inputs
|
80 |
+
|
81 |
+
# Function to generate responses from models
|
82 |
+
@spaces.GPU(duration=120)
|
83 |
def generate(
|
84 |
model_id: str,
|
85 |
message: str,
|
86 |
chat_history: List[Tuple[str, str]],
|
87 |
max_new_tokens: int = DEFAULT_MAX_NEW_TOKENS,
|
88 |
+
temperature: float = 0.4,
|
89 |
top_p: float = 0.95,
|
90 |
) -> Iterator[str]:
|
91 |
model = models[model_id]
|
|
|
118 |
outputs.append(text)
|
119 |
yield "".join(outputs)
|
120 |
|
121 |
+
# Function to compare two models
|
122 |
def compare_models(
|
123 |
model1_name: str,
|
124 |
model2_name: str,
|
|
|
143 |
|
144 |
return chat_history1, chat_history2, chat_history1, chat_history2
|
145 |
|
146 |
+
# Function to log the voting result
|
|
|
147 |
def vote_better(model1_name, model2_name, question, answer1, answer2, choice):
|
148 |
winner = model1_name if choice == "Model 1" else model2_name
|
149 |
log_comparison(model1_name, model2_name, question, answer1, answer2, winner)
|
150 |
return f"You voted that {winner} performs better. This has been logged."
|
151 |
|
152 |
+
# Gradio UI setup
|
153 |
with gr.Blocks(css="style.css") as demo:
|
154 |
gr.Markdown(DESCRIPTION)
|
155 |
|
|
|
194 |
outputs=[vote_output]
|
195 |
)
|
196 |
|
197 |
+
# Main function to run the Gradio app
|
198 |
if __name__ == "__main__":
|
199 |
+
demo.queue(max_size=3).launch(share=True)
|
|
|
|
|
|
|
|