Update app.py
Browse files
app.py
CHANGED
@@ -1,24 +1,41 @@
|
|
1 |
import gradio as gr
|
2 |
from transformers import AutoTokenizer, AutoModelForCausalLM
|
3 |
import torch
|
|
|
|
|
|
|
4 |
|
|
|
5 |
model_name = "Qwen/Qwen2.5-0.5B-Instruct"
|
|
|
6 |
|
7 |
-
|
8 |
-
|
9 |
-
|
10 |
-
)
|
11 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
12 |
def predict(history, message):
|
13 |
-
|
14 |
-
history
|
15 |
-
message: new user input string
|
16 |
-
"""
|
17 |
-
# Add the latest user message to the conversation
|
18 |
-
history = history or [] # make sure it's a list
|
19 |
history.append((message, ""))
|
20 |
|
21 |
-
# Convert to
|
22 |
messages = []
|
23 |
for human, bot in history:
|
24 |
if human:
|
@@ -26,28 +43,47 @@ def predict(history, message):
|
|
26 |
if bot:
|
27 |
messages.append({"role": "assistant", "content": bot})
|
28 |
|
29 |
-
# Apply chat template
|
30 |
text = tokenizer.apply_chat_template(
|
31 |
messages, tokenize=False, add_generation_prompt=True
|
32 |
)
|
33 |
-
|
34 |
model_inputs = tokenizer([text], return_tensors="pt").to(model.device)
|
35 |
|
36 |
-
|
37 |
-
|
38 |
-
|
39 |
-
|
40 |
-
|
41 |
-
|
42 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
43 |
|
44 |
-
# Update last message with bot reply
|
45 |
history[-1] = (message, reply)
|
46 |
-
return history, ""
|
47 |
|
|
|
|
|
|
|
|
|
|
|
48 |
with gr.Blocks() as demo:
|
49 |
-
|
50 |
-
|
51 |
-
|
|
|
|
|
|
|
|
|
|
|
52 |
|
53 |
-
|
|
|
|
|
|
1 |
import gradio as gr
|
2 |
from transformers import AutoTokenizer, AutoModelForCausalLM
|
3 |
import torch
|
4 |
+
import threading
|
5 |
+
import time
|
6 |
+
import os
|
7 |
|
8 |
+
# Model config
|
9 |
model_name = "Qwen/Qwen2.5-0.5B-Instruct"
|
10 |
+
offload_dir = "offload"
|
11 |
|
12 |
+
# Global variables
|
13 |
+
tokenizer = None
|
14 |
+
model = None
|
15 |
+
model_lock = threading.Lock()
|
16 |
|
17 |
+
# Lazy-load the model with quantization & offloading
|
18 |
+
def load_model():
|
19 |
+
global tokenizer, model
|
20 |
+
if model is None:
|
21 |
+
tokenizer = AutoTokenizer.from_pretrained(model_name)
|
22 |
+
# Ensure offload folder exists
|
23 |
+
os.makedirs(offload_dir, exist_ok=True)
|
24 |
+
model = AutoModelForCausalLM.from_pretrained(
|
25 |
+
model_name,
|
26 |
+
load_in_8bit=True, # Quantize to 8-bit
|
27 |
+
device_map="auto",
|
28 |
+
offload_folder=offload_dir, # Offload some weights to disk
|
29 |
+
torch_dtype=torch.float16
|
30 |
+
)
|
31 |
+
|
32 |
+
# Chatbot prediction function
|
33 |
def predict(history, message):
|
34 |
+
load_model()
|
35 |
+
history = history or []
|
|
|
|
|
|
|
|
|
36 |
history.append((message, ""))
|
37 |
|
38 |
+
# Convert to Qwen message format
|
39 |
messages = []
|
40 |
for human, bot in history:
|
41 |
if human:
|
|
|
43 |
if bot:
|
44 |
messages.append({"role": "assistant", "content": bot})
|
45 |
|
|
|
46 |
text = tokenizer.apply_chat_template(
|
47 |
messages, tokenize=False, add_generation_prompt=True
|
48 |
)
|
|
|
49 |
model_inputs = tokenizer([text], return_tensors="pt").to(model.device)
|
50 |
|
51 |
+
reply = ""
|
52 |
+
try:
|
53 |
+
with model_lock: # Serialize CPU inference safely
|
54 |
+
with torch.no_grad():
|
55 |
+
start = time.time()
|
56 |
+
generated_ids = model.generate(**model_inputs, max_new_tokens=256)
|
57 |
+
if time.time() - start > 30: # 30s timeout
|
58 |
+
reply = "[Response timed out]"
|
59 |
+
else:
|
60 |
+
# Remove input_ids from output
|
61 |
+
generated_ids = [
|
62 |
+
output_ids[len(input_ids):]
|
63 |
+
for input_ids, output_ids in zip(model_inputs.input_ids, generated_ids)
|
64 |
+
]
|
65 |
+
reply = tokenizer.batch_decode(generated_ids, skip_special_tokens=True)[0]
|
66 |
+
except Exception as e:
|
67 |
+
reply = f"[Error: {str(e)}]"
|
68 |
|
|
|
69 |
history[-1] = (message, reply)
|
70 |
+
return history, ""
|
71 |
|
72 |
+
# Keep-alive endpoint for local client ping
|
73 |
+
def keep_alive(msg="ping"):
|
74 |
+
return "pong"
|
75 |
+
|
76 |
+
# Gradio UI
|
77 |
with gr.Blocks() as demo:
|
78 |
+
with gr.Tab("Chatbot"):
|
79 |
+
chatbot = gr.Chatbot()
|
80 |
+
msg = gr.Textbox(placeholder="Type your message here...")
|
81 |
+
msg.submit(predict, [chatbot, msg], [chatbot, msg])
|
82 |
+
|
83 |
+
with gr.Tab("Keep Alive"):
|
84 |
+
gr.Textbox(label="Ping", value="ping", interactive=False)
|
85 |
+
gr.Button("Ping").click(keep_alive, inputs=None, outputs=None)
|
86 |
|
87 |
+
# Multi-user queue with concurrency
|
88 |
+
demo.queue(concurrency_count=4, max_size=8) # 4 simultaneous, 8 waiting
|
89 |
+
demo.launch(server_name="0.0.0.0", server_port=7860, share=False, show_error=True)
|