Spaces:
Running
on
Zero
Running
on
Zero
Update app.py
Browse files
app.py
CHANGED
@@ -1,57 +1,67 @@
|
|
1 |
import os
|
2 |
-
from collections.abc import Iterator
|
3 |
-
from threading import Thread
|
4 |
-
|
5 |
-
import gradio as gr
|
6 |
-
import spaces
|
7 |
import torch
|
|
|
8 |
from transformers import AutoModelForCausalLM, AutoTokenizer, TextIteratorStreamer
|
9 |
|
|
|
10 |
DESCRIPTION = """\
|
11 |
-
# Llama 3.2 3B Instruct
|
12 |
-
|
13 |
-
Llama 3.2 3B is Meta's latest iteration of open LLMs.
|
14 |
-
This is a demo of [`meta-llama/Llama-3.2-3B-Instruct`](https://huggingface.co/meta-llama/Llama-3.2-3B-Instruct), fine-tuned for instruction following.
|
15 |
-
For more details, please check [our post](https://huggingface.co/blog/llama32).
|
16 |
"""
|
17 |
|
|
|
18 |
MAX_MAX_NEW_TOKENS = 2048
|
19 |
-
DEFAULT_MAX_NEW_TOKENS =
|
20 |
-
MAX_INPUT_TOKEN_LENGTH =
|
21 |
|
22 |
-
|
|
|
23 |
|
|
|
24 |
model_id = "meta-llama/Llama-3.2-3B-Instruct"
|
|
|
|
|
25 |
tokenizer = AutoTokenizer.from_pretrained(model_id)
|
|
|
|
|
26 |
model = AutoModelForCausalLM.from_pretrained(
|
27 |
model_id,
|
28 |
-
device_map=
|
29 |
-
torch_dtype=torch.
|
30 |
)
|
31 |
model.eval()
|
|
|
32 |
|
33 |
-
|
34 |
-
@spaces.GPU(duration=90)
|
35 |
def generate(
|
36 |
message: str,
|
37 |
chat_history: list[dict],
|
38 |
-
max_new_tokens: int =
|
39 |
temperature: float = 0.6,
|
40 |
top_p: float = 0.9,
|
41 |
top_k: int = 50,
|
42 |
repetition_penalty: float = 1.2,
|
43 |
-
)
|
|
|
44 |
conversation = [*chat_history, {"role": "user", "content": message}]
|
45 |
-
|
46 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
47 |
if input_ids.shape[1] > MAX_INPUT_TOKEN_LENGTH:
|
48 |
input_ids = input_ids[:, -MAX_INPUT_TOKEN_LENGTH:]
|
49 |
-
|
50 |
-
input_ids = input_ids.to(
|
51 |
|
|
|
52 |
streamer = TextIteratorStreamer(tokenizer, timeout=20.0, skip_prompt=True, skip_special_tokens=True)
|
53 |
generate_kwargs = dict(
|
54 |
-
|
55 |
streamer=streamer,
|
56 |
max_new_tokens=max_new_tokens,
|
57 |
do_sample=True,
|
@@ -61,15 +71,17 @@ def generate(
|
|
61 |
num_beams=1,
|
62 |
repetition_penalty=repetition_penalty,
|
63 |
)
|
64 |
-
|
65 |
-
|
66 |
-
|
|
|
67 |
outputs = []
|
|
|
68 |
for text in streamer:
|
69 |
outputs.append(text)
|
70 |
yield "".join(outputs)
|
71 |
|
72 |
-
|
73 |
demo = gr.ChatInterface(
|
74 |
fn=generate,
|
75 |
additional_inputs=[
|
@@ -120,10 +132,10 @@ demo = gr.ChatInterface(
|
|
120 |
cache_examples=False,
|
121 |
type="messages",
|
122 |
description=DESCRIPTION,
|
123 |
-
css_paths="style.css",
|
124 |
fill_height=True,
|
125 |
)
|
126 |
|
127 |
-
|
128 |
if __name__ == "__main__":
|
|
|
129 |
demo.queue(max_size=20).launch()
|
|
|
1 |
import os
|
|
|
|
|
|
|
|
|
|
|
2 |
import torch
|
3 |
+
import gradio as gr
|
4 |
from transformers import AutoModelForCausalLM, AutoTokenizer, TextIteratorStreamer
|
5 |
|
6 |
+
# وصف التطبيق
|
7 |
DESCRIPTION = """\
|
8 |
+
# Llama 3.2 3B Instruct (CPU-Only)
|
9 |
+
هذا نموذج توضيحي لـ [`meta-llama/Llama-3.2-3B-Instruct`](https://huggingface.co/meta-llama/Llama-3.2-3B-Instruct) يعمل باستخدام الـ CPU فقط.
|
|
|
|
|
|
|
10 |
"""
|
11 |
|
12 |
+
# إعداد الثوابت
|
13 |
MAX_MAX_NEW_TOKENS = 2048
|
14 |
+
DEFAULT_MAX_NEW_TOKENS = 512
|
15 |
+
MAX_INPUT_TOKEN_LENGTH = 4096 # الحد الأقصى لعدد التوكنات في المدخلات
|
16 |
|
17 |
+
# تحديد الجهاز: استخدام CPU فقط
|
18 |
+
device = torch.device("cpu")
|
19 |
|
20 |
+
# تحديد معرف النموذج وتحميله
|
21 |
model_id = "meta-llama/Llama-3.2-3B-Instruct"
|
22 |
+
|
23 |
+
# تحميل التوكن الخاص بالنموذج
|
24 |
tokenizer = AutoTokenizer.from_pretrained(model_id)
|
25 |
+
|
26 |
+
# تحميل النموذج على CPU مع استخدام torch.float32
|
27 |
model = AutoModelForCausalLM.from_pretrained(
|
28 |
model_id,
|
29 |
+
device_map=None, # عدم استخدام GPU
|
30 |
+
torch_dtype=torch.float32
|
31 |
)
|
32 |
model.eval()
|
33 |
+
model.to(device)
|
34 |
|
|
|
|
|
35 |
def generate(
|
36 |
message: str,
|
37 |
chat_history: list[dict],
|
38 |
+
max_new_tokens: int = DEFAULT_MAX_NEW_TOKENS,
|
39 |
temperature: float = 0.6,
|
40 |
top_p: float = 0.9,
|
41 |
top_k: int = 50,
|
42 |
repetition_penalty: float = 1.2,
|
43 |
+
):
|
44 |
+
# دمج سجل المحادثة مع الرسالة الجديدة
|
45 |
conversation = [*chat_history, {"role": "user", "content": message}]
|
46 |
+
|
47 |
+
# تحويل المحادثة إلى مدخلات للنموذج
|
48 |
+
inputs = tokenizer.apply_chat_template(
|
49 |
+
conversation,
|
50 |
+
add_generation_prompt=True,
|
51 |
+
return_tensors="pt"
|
52 |
+
)
|
53 |
+
input_ids = inputs["input_ids"]
|
54 |
+
|
55 |
+
# قص التوكنز إذا تجاوز طولها الحد المسموح
|
56 |
if input_ids.shape[1] > MAX_INPUT_TOKEN_LENGTH:
|
57 |
input_ids = input_ids[:, -MAX_INPUT_TOKEN_LENGTH:]
|
58 |
+
|
59 |
+
input_ids = input_ids.to(device)
|
60 |
|
61 |
+
# إعداد البث التدريجي للنص باستخدام TextIteratorStreamer
|
62 |
streamer = TextIteratorStreamer(tokenizer, timeout=20.0, skip_prompt=True, skip_special_tokens=True)
|
63 |
generate_kwargs = dict(
|
64 |
+
input_ids=input_ids,
|
65 |
streamer=streamer,
|
66 |
max_new_tokens=max_new_tokens,
|
67 |
do_sample=True,
|
|
|
71 |
num_beams=1,
|
72 |
repetition_penalty=repetition_penalty,
|
73 |
)
|
74 |
+
|
75 |
+
# تشغيل عملية التوليد على نفس الخيط (CPU)
|
76 |
+
model.generate(**generate_kwargs)
|
77 |
+
|
78 |
outputs = []
|
79 |
+
# بث النص تدريجيًا أثناء توليد النموذج
|
80 |
for text in streamer:
|
81 |
outputs.append(text)
|
82 |
yield "".join(outputs)
|
83 |
|
84 |
+
# إنشاء واجهة الدردشة باستخدام Gradio
|
85 |
demo = gr.ChatInterface(
|
86 |
fn=generate,
|
87 |
additional_inputs=[
|
|
|
132 |
cache_examples=False,
|
133 |
type="messages",
|
134 |
description=DESCRIPTION,
|
135 |
+
css_paths="style.css", # تأكدي من رفع ملف style.css إذا كان موجوداً
|
136 |
fill_height=True,
|
137 |
)
|
138 |
|
|
|
139 |
if __name__ == "__main__":
|
140 |
+
# استخدام queue() لإدارة الطلبات المتزامنة
|
141 |
demo.queue(max_size=20).launch()
|