Spaces:
Sleeping
Sleeping
TenzinGayche
commited on
Create app.py
Browse files
app.py
ADDED
@@ -0,0 +1,54 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import gradio as gr
|
2 |
+
import torch
|
3 |
+
from transformers import AutoModelForCausalLM, GemmaTokenizerFast, StoppingCriteria, StoppingCriteriaList, TextIteratorStreamer
|
4 |
+
from threading import Thread
|
5 |
+
|
6 |
+
# Load tokenizer and model
|
7 |
+
tokenizer = GemmaTokenizerFast.from_pretrained("buddhist-nlp/gemma2-mitra-bo-instruct")
|
8 |
+
model = AutoModelForCausalLM.from_pretrained("buddhist-nlp/gemma2-mitra-bo-instruct", torch_dtype=torch.float16).to('cuda:0')
|
9 |
+
|
10 |
+
# Define custom stopping criteria
|
11 |
+
class StopOnTokens(StoppingCriteria):
|
12 |
+
def __call__(self, input_ids: torch.LongTensor, scores: torch.FloatTensor, **kwargs) -> bool:
|
13 |
+
# Define stop tokens (adjust based on your model's tokenizer)
|
14 |
+
stop_ids = [29, 0] # These should be the token IDs for end of response or similar tokens
|
15 |
+
for stop_id in stop_ids:
|
16 |
+
if input_ids[0][-1] == stop_id:
|
17 |
+
return True
|
18 |
+
return False
|
19 |
+
|
20 |
+
# Define prediction function for the chat interface
|
21 |
+
def predict(message, history):
|
22 |
+
# Prepare the conversation in the required format
|
23 |
+
history_transformer_format = history + [[message, ""]]
|
24 |
+
stop = StopOnTokens()
|
25 |
+
|
26 |
+
# Concatenate previous messages and the user's input
|
27 |
+
messages = "".join([f"\n### user : {item[0]} \n### bot : {item[1]}" for item in history_transformer_format])
|
28 |
+
|
29 |
+
# Tokenize the input
|
30 |
+
model_inputs = tokenizer([messages], return_tensors="pt").to("cuda")
|
31 |
+
|
32 |
+
# Set up the streamer for partial message output
|
33 |
+
streamer = TextIteratorStreamer(tokenizer, timeout=10., skip_prompt=True, skip_special_tokens=True)
|
34 |
+
|
35 |
+
# Generate settings
|
36 |
+
generate_kwargs = dict(
|
37 |
+
model_inputs,
|
38 |
+
streamer=streamer,
|
39 |
+
max_new_tokens=1024,
|
40 |
+
)
|
41 |
+
|
42 |
+
# Run generation in a separate thread
|
43 |
+
t = Thread(target=model.generate, kwargs=generate_kwargs)
|
44 |
+
t.start()
|
45 |
+
|
46 |
+
# Stream partial messages as they are generated
|
47 |
+
partial_message = ""
|
48 |
+
for new_token in streamer:
|
49 |
+
if new_token != '<': # Skip specific tokens if necessary
|
50 |
+
partial_message += new_token
|
51 |
+
yield partial_message
|
52 |
+
|
53 |
+
# Create the chat interface using Gradio
|
54 |
+
gr.ChatInterface(fn=predict, title="Gemma LLM Chatbot", description="Chat with the Gemma model using real-time generation and streaming.").launch(share=True)
|