Spaces:
Running
on
Zero
Running
on
Zero
Update app.py
Browse files* 設定pad_token
* 使用return_dict=True獲取attention_mask
* 在generate_kwargs中加入必要參數
* 增加錯誤處理和記憶體清理
* 增加streamer timeout
app.py
CHANGED
@@ -1,18 +1,13 @@
|
|
1 |
import gradio as gr
|
2 |
import spaces
|
3 |
import os
|
4 |
-
import spaces
|
5 |
import torch
|
6 |
-
from transformers import GemmaTokenizer, AutoModelForCausalLM
|
7 |
from transformers import AutoModelForCausalLM, AutoTokenizer, TextIteratorStreamer
|
8 |
from threading import Thread
|
9 |
|
10 |
# Set an environment variable
|
11 |
HF_TOKEN = os.environ.get("HF_TOKEN", None)
|
12 |
|
13 |
-
zero = torch.Tensor([0]).cuda()
|
14 |
-
print(zero.device) # <-- 'cpu' 🤔
|
15 |
-
|
16 |
DESCRIPTION = '''
|
17 |
<div>
|
18 |
<h1 style="text-align: center;">TAIDE/Llama3-TAIDE-LX-8B-Chat-Alpha1</h1>
|
@@ -41,7 +36,12 @@ h1 {
|
|
41 |
|
42 |
# Load the tokenizer and model
|
43 |
tokenizer = AutoTokenizer.from_pretrained("taide/Llama3-TAIDE-LX-8B-Chat-Alpha1")
|
44 |
-
model = AutoModelForCausalLM.from_pretrained("taide/Llama3-TAIDE-LX-8B-Chat-Alpha1")
|
|
|
|
|
|
|
|
|
|
|
45 |
terminators = [
|
46 |
tokenizer.eos_token_id,
|
47 |
tokenizer.convert_tokens_to_ids("<|eot_id|>")
|
@@ -55,47 +55,60 @@ def chat_taide_8b(message: str,
|
|
55 |
) -> str:
|
56 |
"""
|
57 |
Generate a streaming response using the llama3-8b model.
|
58 |
-
Args:
|
59 |
-
message (str): The input message.
|
60 |
-
history (list): The conversation history used by ChatInterface.
|
61 |
-
temperature (float): The temperature for generating the response.
|
62 |
-
max_new_tokens (int): The maximum number of new tokens to generate.
|
63 |
-
Returns:
|
64 |
-
str: The generated response.
|
65 |
"""
|
66 |
-
|
67 |
-
|
68 |
-
|
69 |
-
|
70 |
-
|
71 |
-
input_ids = tokenizer.apply_chat_template(conversation, return_tensors="pt").to(model.device)
|
72 |
-
|
73 |
-
streamer = TextIteratorStreamer(tokenizer, timeout=10.0, skip_prompt=True, skip_special_tokens=True)
|
74 |
|
75 |
-
|
76 |
-
|
77 |
-
|
78 |
-
|
79 |
-
|
80 |
-
|
81 |
-
|
82 |
-
|
83 |
-
|
84 |
-
|
85 |
-
|
|
|
86 |
|
87 |
-
|
88 |
-
t.start()
|
89 |
|
90 |
-
|
91 |
-
|
92 |
-
|
93 |
-
|
94 |
-
|
|
|
|
|
|
|
|
|
|
|
95 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
96 |
|
97 |
# Gradio block
|
98 |
-
chatbot=gr.Chatbot(height=450, label='Gradio ChatInterface')
|
99 |
|
100 |
with gr.Blocks(fill_height=True, css=css) as demo:
|
101 |
|
@@ -118,15 +131,15 @@ with gr.Blocks(fill_height=True, css=css) as demo:
|
|
118 |
step=1,
|
119 |
value=512,
|
120 |
label="Max new tokens",
|
121 |
-
render=False
|
122 |
-
|
123 |
examples=[
|
124 |
['請以以下內容為基礎,寫一篇文章:撰寫一篇作文,題目為《一張舊照片》,內容要求為:選擇一張令你印象深刻的照片,說明令你印象深刻的原因,並描述照片中的影像及背後的故事。記錄成長的過程、與他人的情景、環境變遷和美麗的景色。'],
|
125 |
['請以品牌經理的身份,給廣告公司的創意總監寫一封信,提出對於新產品廣告宣傳活動的創意建議。'],
|
126 |
['以下提供英文內容,請幫我翻譯成中文。Dongshan coffee is famous for its unique position, and the constant refinement of production methods. The flavor is admired by many caffeine afficionados.'],
|
127 |
-
|
128 |
cache_examples=False,
|
129 |
-
|
130 |
|
131 |
gr.Markdown(LICENSE)
|
132 |
|
|
|
1 |
import gradio as gr
|
2 |
import spaces
|
3 |
import os
|
|
|
4 |
import torch
|
|
|
5 |
from transformers import AutoModelForCausalLM, AutoTokenizer, TextIteratorStreamer
|
6 |
from threading import Thread
|
7 |
|
8 |
# Set an environment variable
|
9 |
HF_TOKEN = os.environ.get("HF_TOKEN", None)
|
10 |
|
|
|
|
|
|
|
11 |
DESCRIPTION = '''
|
12 |
<div>
|
13 |
<h1 style="text-align: center;">TAIDE/Llama3-TAIDE-LX-8B-Chat-Alpha1</h1>
|
|
|
36 |
|
37 |
# Load the tokenizer and model
|
38 |
tokenizer = AutoTokenizer.from_pretrained("taide/Llama3-TAIDE-LX-8B-Chat-Alpha1")
|
39 |
+
model = AutoModelForCausalLM.from_pretrained("taide/Llama3-TAIDE-LX-8B-Chat-Alpha1")
|
40 |
+
|
41 |
+
# 設定pad_token_id(關鍵修正)
|
42 |
+
if tokenizer.pad_token is None:
|
43 |
+
tokenizer.pad_token = tokenizer.eos_token
|
44 |
+
|
45 |
terminators = [
|
46 |
tokenizer.eos_token_id,
|
47 |
tokenizer.convert_tokens_to_ids("<|eot_id|>")
|
|
|
55 |
) -> str:
|
56 |
"""
|
57 |
Generate a streaming response using the llama3-8b model.
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
58 |
"""
|
59 |
+
try:
|
60 |
+
conversation = []
|
61 |
+
for user, assistant in history:
|
62 |
+
conversation.extend([{"role": "user", "content": user}, {"role": "assistant", "content": assistant}])
|
63 |
+
conversation.append({"role": "user", "content": message})
|
|
|
|
|
|
|
64 |
|
65 |
+
# 使用return_dict=True來獲取attention_mask(關鍵修正)
|
66 |
+
inputs = tokenizer.apply_chat_template(
|
67 |
+
conversation,
|
68 |
+
return_tensors="pt",
|
69 |
+
return_dict=True,
|
70 |
+
add_generation_prompt=True
|
71 |
+
)
|
72 |
+
|
73 |
+
input_ids = inputs["input_ids"].to(model.device)
|
74 |
+
attention_mask = inputs.get("attention_mask", None)
|
75 |
+
if attention_mask is not None:
|
76 |
+
attention_mask = attention_mask.to(model.device)
|
77 |
|
78 |
+
streamer = TextIteratorStreamer(tokenizer, timeout=30.0, skip_prompt=True, skip_special_tokens=True)
|
|
|
79 |
|
80 |
+
generate_kwargs = dict(
|
81 |
+
input_ids=input_ids,
|
82 |
+
attention_mask=attention_mask, # 加入attention_mask
|
83 |
+
streamer=streamer,
|
84 |
+
max_new_tokens=max_new_tokens,
|
85 |
+
do_sample=True,
|
86 |
+
temperature=temperature,
|
87 |
+
eos_token_id=terminators,
|
88 |
+
pad_token_id=tokenizer.pad_token_id, # 明確設定pad_token_id
|
89 |
+
)
|
90 |
|
91 |
+
# This will enforce greedy generation (do_sample=False) when the temperature is passed 0, avoiding the crash.
|
92 |
+
if temperature == 0:
|
93 |
+
generate_kwargs['do_sample'] = False
|
94 |
+
|
95 |
+
t = Thread(target=model.generate, kwargs=generate_kwargs)
|
96 |
+
t.start()
|
97 |
+
|
98 |
+
outputs = []
|
99 |
+
for text in streamer:
|
100 |
+
outputs.append(text)
|
101 |
+
yield "".join(outputs)
|
102 |
+
|
103 |
+
except Exception as e:
|
104 |
+
yield f"生成過程中發生錯誤: {str(e)}"
|
105 |
+
finally:
|
106 |
+
# 清理GPU記憶體
|
107 |
+
if torch.cuda.is_available():
|
108 |
+
torch.cuda.empty_cache()
|
109 |
|
110 |
# Gradio block
|
111 |
+
chatbot = gr.Chatbot(height=450, label='Gradio ChatInterface')
|
112 |
|
113 |
with gr.Blocks(fill_height=True, css=css) as demo:
|
114 |
|
|
|
131 |
step=1,
|
132 |
value=512,
|
133 |
label="Max new tokens",
|
134 |
+
render=False),
|
135 |
+
],
|
136 |
examples=[
|
137 |
['請以以下內容為基礎,寫一篇文章:撰寫一篇作文,題目為《一張舊照片》,內容要求為:選擇一張令你印象深刻的照片,說明令你印象深刻的原因,並描述照片中的影像及背後的故事。記錄成長的過程、與他人的情景、環境變遷和美麗的景色。'],
|
138 |
['請以品牌經理的身份,給廣告公司的創意總監寫一封信,提出對於新產品廣告宣傳活動的創意建議。'],
|
139 |
['以下提供英文內容,請幫我翻譯成中文。Dongshan coffee is famous for its unique position, and the constant refinement of production methods. The flavor is admired by many caffeine afficionados.'],
|
140 |
+
],
|
141 |
cache_examples=False,
|
142 |
+
)
|
143 |
|
144 |
gr.Markdown(LICENSE)
|
145 |
|