Spaces:
Running
Running
Update app.py
Browse files
app.py
CHANGED
|
@@ -11,7 +11,7 @@ from collections import OrderedDict
|
|
| 11 |
|
| 12 |
app = FastAPI()
|
| 13 |
|
| 14 |
-
# 🔥
|
| 15 |
model_name = "Qwen/Qwen2.5-1.5B-Instruct"
|
| 16 |
|
| 17 |
tokenizer = AutoTokenizer.from_pretrained(model_name)
|
|
@@ -36,10 +36,39 @@ def split_text(text, max_len=100):
|
|
| 36 |
return "\n".join([text[i:i+max_len] for i in range(0, len(text), max_len)])
|
| 37 |
|
| 38 |
|
| 39 |
-
# 🔥
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 40 |
def generate_ai_stream(message: str):
|
|
|
|
| 41 |
messages = [
|
| 42 |
-
{
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 43 |
]
|
| 44 |
|
| 45 |
prompt = tokenizer.apply_chat_template(
|
|
@@ -55,7 +84,6 @@ def generate_ai_stream(message: str):
|
|
| 55 |
gen_kwargs = dict(
|
| 56 |
**inputs,
|
| 57 |
max_new_tokens=900,
|
| 58 |
-
min_new_tokens=90,
|
| 59 |
do_sample=True,
|
| 60 |
temperature=0.7,
|
| 61 |
top_p=0.9,
|
|
@@ -75,11 +103,17 @@ def generate_ai_stream(message: str):
|
|
| 75 |
if message in db:
|
| 76 |
db[message]["reply"] = split_text(partial)
|
| 77 |
|
| 78 |
-
# 🔥
|
|
|
|
|
|
|
|
|
|
| 79 |
try:
|
| 80 |
-
translated = GoogleTranslator(
|
|
|
|
|
|
|
|
|
|
| 81 |
except:
|
| 82 |
-
translated =
|
| 83 |
|
| 84 |
final_text = split_text(translated) + " full generated"
|
| 85 |
|
|
@@ -100,7 +134,6 @@ def worker():
|
|
| 100 |
continue
|
| 101 |
|
| 102 |
generate_ai_stream(message)
|
| 103 |
-
|
| 104 |
else:
|
| 105 |
time.sleep(0.01)
|
| 106 |
|
|
|
|
| 11 |
|
| 12 |
app = FastAPI()
|
| 13 |
|
| 14 |
+
# 🔥 MODEL
|
| 15 |
model_name = "Qwen/Qwen2.5-1.5B-Instruct"
|
| 16 |
|
| 17 |
tokenizer = AutoTokenizer.from_pretrained(model_name)
|
|
|
|
| 36 |
return "\n".join([text[i:i+max_len] for i in range(0, len(text), max_len)])
|
| 37 |
|
| 38 |
|
| 39 |
+
# 🔥 очистка мусора Qwen
|
| 40 |
+
def clean_output(text: str):
|
| 41 |
+
bad = [
|
| 42 |
+
"system",
|
| 43 |
+
"user",
|
| 44 |
+
"assistant",
|
| 45 |
+
"<|im_start|>",
|
| 46 |
+
"<|im_end|>",
|
| 47 |
+
"You are Qwen"
|
| 48 |
+
]
|
| 49 |
+
|
| 50 |
+
for b in bad:
|
| 51 |
+
text = text.replace(b, "")
|
| 52 |
+
|
| 53 |
+
return text.strip()
|
| 54 |
+
|
| 55 |
+
|
| 56 |
+
# 🔥 GENERATION
|
| 57 |
def generate_ai_stream(message: str):
|
| 58 |
+
|
| 59 |
messages = [
|
| 60 |
+
{
|
| 61 |
+
"role": "system",
|
| 62 |
+
"content": (
|
| 63 |
+
"Ты умный и точный ассистент. "
|
| 64 |
+
"Отвечай логично, кратко и понятно. "
|
| 65 |
+
"Если пользователь пишет на русском — отвечай на русском."
|
| 66 |
+
)
|
| 67 |
+
},
|
| 68 |
+
{
|
| 69 |
+
"role": "user",
|
| 70 |
+
"content": message
|
| 71 |
+
}
|
| 72 |
]
|
| 73 |
|
| 74 |
prompt = tokenizer.apply_chat_template(
|
|
|
|
| 84 |
gen_kwargs = dict(
|
| 85 |
**inputs,
|
| 86 |
max_new_tokens=900,
|
|
|
|
| 87 |
do_sample=True,
|
| 88 |
temperature=0.7,
|
| 89 |
top_p=0.9,
|
|
|
|
| 103 |
if message in db:
|
| 104 |
db[message]["reply"] = split_text(partial)
|
| 105 |
|
| 106 |
+
# 🔥 чистка
|
| 107 |
+
raw = clean_output(partial)
|
| 108 |
+
|
| 109 |
+
# 🔥 перевод (fallback)
|
| 110 |
try:
|
| 111 |
+
translated = GoogleTranslator(
|
| 112 |
+
source='auto',
|
| 113 |
+
target='ru'
|
| 114 |
+
).translate(raw)
|
| 115 |
except:
|
| 116 |
+
translated = raw
|
| 117 |
|
| 118 |
final_text = split_text(translated) + " full generated"
|
| 119 |
|
|
|
|
| 134 |
continue
|
| 135 |
|
| 136 |
generate_ai_stream(message)
|
|
|
|
| 137 |
else:
|
| 138 |
time.sleep(0.01)
|
| 139 |
|