Update app.py
Browse files
app.py
CHANGED
|
@@ -7,25 +7,46 @@ import spaces
|
|
| 7 |
from transformers import Qwen3OmniMoeForConditionalGeneration, Qwen3OmniMoeProcessor
|
| 8 |
from qwen_omni_utils import process_mm_info
|
| 9 |
|
| 10 |
-
#
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 11 |
# إعدادات عامة
|
| 12 |
-
#
|
| 13 |
|
| 14 |
MODEL_PATH = os.getenv("MODEL_PATH", "Qwen/Qwen3-Omni-30B-A3B-Instruct")
|
| 15 |
-
USE_AUDIO_IN_VIDEO = True # استخدام الصوت داخل الفيديو
|
| 16 |
|
| 17 |
VOICE_CHOICES = ["Ethan", "Chelsie", "Aiden"]
|
| 18 |
DEFAULT_VOICE = "Ethan"
|
| 19 |
|
| 20 |
-
#
|
| 21 |
model = None
|
| 22 |
processor = None
|
| 23 |
|
| 24 |
|
| 25 |
def load_model():
|
| 26 |
"""
|
| 27 |
-
تحميل Qwen3-Omni والمعالج عند أول استدعاء
|
| 28 |
-
|
|
|
|
| 29 |
"""
|
| 30 |
global model, processor
|
| 31 |
|
|
@@ -34,7 +55,7 @@ def load_model():
|
|
| 34 |
|
| 35 |
print(f"[ZeroGPU] Loading model from: {MODEL_PATH}")
|
| 36 |
|
| 37 |
-
#
|
| 38 |
if torch.cuda.is_available():
|
| 39 |
torch_dtype = torch.bfloat16
|
| 40 |
device = "cuda"
|
|
@@ -42,13 +63,12 @@ def load_model():
|
|
| 42 |
torch_dtype = torch.float32
|
| 43 |
device = "cpu"
|
| 44 |
|
| 45 |
-
# تحميل النموذج بدون flash_attention_2
|
| 46 |
local_model = Qwen3OmniMoeForConditionalGeneration.from_pretrained(
|
| 47 |
MODEL_PATH,
|
| 48 |
torch_dtype=torch_dtype,
|
| 49 |
-
attn_implementation="eager", #
|
| 50 |
)
|
| 51 |
-
|
| 52 |
local_model.to(device)
|
| 53 |
|
| 54 |
local_processor = Qwen3OmniMoeProcessor.from_pretrained(MODEL_PATH)
|
|
@@ -58,33 +78,47 @@ def load_model():
|
|
| 58 |
print(f"[ZeroGPU] Model loaded on {device} with dtype {torch_dtype}.")
|
| 59 |
|
| 60 |
|
| 61 |
-
def build_messages_from_history(
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 62 |
"""
|
| 63 |
-
تحويل تاريخ الدردشة + المدخل الحالي إلى
|
|
|
|
| 64 |
history: list of [user_text, assistant_text]
|
| 65 |
"""
|
| 66 |
messages = []
|
| 67 |
|
| 68 |
if system_prompt:
|
| 69 |
-
messages.append(
|
| 70 |
-
|
| 71 |
-
|
| 72 |
-
|
|
|
|
|
|
|
| 73 |
|
| 74 |
# تاريخ المحادثة
|
| 75 |
for user_msg, assistant_msg in history:
|
| 76 |
if user_msg:
|
| 77 |
-
messages.append(
|
| 78 |
-
|
| 79 |
-
|
| 80 |
-
|
|
|
|
|
|
|
| 81 |
if assistant_msg:
|
| 82 |
-
messages.append(
|
| 83 |
-
|
| 84 |
-
|
| 85 |
-
|
|
|
|
|
|
|
| 86 |
|
| 87 |
-
#
|
| 88 |
user_content = []
|
| 89 |
|
| 90 |
if image is not None:
|
|
@@ -100,17 +134,19 @@ def build_messages_from_history(history, system_prompt, user_text, image, audio_
|
|
| 100 |
user_content.append({"type": "text", "text": user_text.strip()})
|
| 101 |
|
| 102 |
if user_content:
|
| 103 |
-
messages.append(
|
| 104 |
-
|
| 105 |
-
|
| 106 |
-
|
|
|
|
|
|
|
| 107 |
|
| 108 |
return messages
|
| 109 |
|
| 110 |
|
| 111 |
-
#
|
| 112 |
-
# دالة الاستدلال (ZeroGPU)
|
| 113 |
-
#
|
| 114 |
|
| 115 |
@spaces.GPU(duration=120)
|
| 116 |
def qwen3_omni_inference(
|
|
@@ -127,13 +163,13 @@ def qwen3_omni_inference(
|
|
| 127 |
max_tokens,
|
| 128 |
):
|
| 129 |
"""
|
| 130 |
-
تنفيذ الاستدلال
|
| 131 |
-
- نص + صورة + صوت + فيديو
|
| 132 |
-
- مخرج نصي دائماً،
|
| 133 |
"""
|
| 134 |
|
|
|
|
| 135 |
if not (user_text or image is not None or audio_path or video_path):
|
| 136 |
-
# لا يوجد مدخل من المستخدم
|
| 137 |
return history, None, "", None, None, None
|
| 138 |
|
| 139 |
load_model()
|
|
@@ -148,14 +184,14 @@ def qwen3_omni_inference(
|
|
| 148 |
video_path=video_path,
|
| 149 |
)
|
| 150 |
|
| 151 |
-
# بناء
|
| 152 |
text_prompt = processor.apply_chat_template(
|
| 153 |
messages,
|
| 154 |
add_generation_prompt=True,
|
| 155 |
tokenize=False,
|
| 156 |
)
|
| 157 |
|
| 158 |
-
# تجهيز الوسائط المتعددة
|
| 159 |
audios, images, videos = process_mm_info(
|
| 160 |
messages,
|
| 161 |
use_audio_in_video=USE_AUDIO_IN_VIDEO,
|
|
@@ -172,13 +208,13 @@ def qwen3_omni_inference(
|
|
| 172 |
use_audio_in_video=USE_AUDIO_IN_VIDEO,
|
| 173 |
)
|
| 174 |
|
| 175 |
-
# نقل
|
| 176 |
first_param = next(model.parameters())
|
| 177 |
device = first_param.device
|
| 178 |
dtype = first_param.dtype
|
| 179 |
inputs = inputs.to(device=device, dtype=dtype)
|
| 180 |
|
| 181 |
-
#
|
| 182 |
gen_kwargs = dict(
|
| 183 |
temperature=float(temperature),
|
| 184 |
top_p=float(top_p),
|
|
@@ -187,16 +223,16 @@ def qwen3_omni_inference(
|
|
| 187 |
use_audio_in_video=USE_AUDIO_IN_VIDEO,
|
| 188 |
)
|
| 189 |
|
| 190 |
-
# توليد
|
| 191 |
if not return_audio:
|
| 192 |
gen_kwargs["return_audio"] = False
|
| 193 |
-
text_ids,
|
| 194 |
audio_out = None
|
| 195 |
else:
|
| 196 |
gen_kwargs["speaker"] = speaker
|
| 197 |
text_ids, audio_out = model.generate(**inputs, **gen_kwargs)
|
| 198 |
|
| 199 |
-
#
|
| 200 |
input_len = inputs["input_ids"].shape[1]
|
| 201 |
generated_text = processor.batch_decode(
|
| 202 |
text_ids.sequences[:, input_len:],
|
|
@@ -205,10 +241,12 @@ def qwen3_omni_inference(
|
|
| 205 |
)[0]
|
| 206 |
|
| 207 |
# تحديث تاريخ الدردشة
|
| 208 |
-
user_display =
|
|
|
|
|
|
|
| 209 |
history = history + [[user_display, generated_text]]
|
| 210 |
|
| 211 |
-
# تجهيز الصوت
|
| 212 |
gr_audio = None
|
| 213 |
if audio_out is not None:
|
| 214 |
audio_np = audio_out.reshape(-1).detach().cpu().numpy()
|
|
@@ -219,12 +257,12 @@ def qwen3_omni_inference(
|
|
| 219 |
return history, gr_audio, "", None, None, None
|
| 220 |
|
| 221 |
|
| 222 |
-
#
|
| 223 |
-
# دوال واجهة
|
| 224 |
-
#
|
| 225 |
|
| 226 |
def clear_chat():
|
| 227 |
-
"""
|
| 228 |
return [], None
|
| 229 |
|
| 230 |
|
|
@@ -236,15 +274,15 @@ def create_interface():
|
|
| 236 |
"""
|
| 237 |
<h1 style="text-align:center;">Qwen3-Omni-30B-A3B – ZeroGPU Chat</h1>
|
| 238 |
<p style="text-align:center;">
|
| 239 |
-
دردشة متعددة الوسائط (نص + صورة + صوت + فيديو) تعمل على ZeroGPU
|
| 240 |
-
|
| 241 |
(لإضافة سطر جديد استخدم Shift+Enter)
|
| 242 |
</p>
|
| 243 |
"""
|
| 244 |
)
|
| 245 |
|
| 246 |
with gr.Row():
|
| 247 |
-
#
|
| 248 |
with gr.Column(scale=3):
|
| 249 |
chatbot = gr.Chatbot(
|
| 250 |
label="المحادثة",
|
|
@@ -257,7 +295,6 @@ def create_interface():
|
|
| 257 |
autoplay=True,
|
| 258 |
)
|
| 259 |
|
| 260 |
-
# منطقة الإدخال أسفل الشات
|
| 261 |
with gr.Row():
|
| 262 |
user_text = gr.Textbox(
|
| 263 |
label="رسالتك",
|
|
@@ -286,7 +323,7 @@ def create_interface():
|
|
| 286 |
send_btn = gr.Button("إرسال", variant="primary", scale=2)
|
| 287 |
clear_btn = gr.Button("مسح المحادثة", variant="secondary")
|
| 288 |
|
| 289 |
-
#
|
| 290 |
with gr.Column(scale=1):
|
| 291 |
gr.Markdown("### إعدادات النموذج")
|
| 292 |
|
|
@@ -335,78 +372,78 @@ def create_interface():
|
|
| 335 |
|
| 336 |
gr.Markdown(
|
| 337 |
"""
|
| 338 |
-
|
| 339 |
-
- يمكنك إرسال نص فقط، أو نص مع صورة/صوت/فيديو في
|
| 340 |
-
-
|
| 341 |
-
- ZeroGPU قد يستغرق عدة ثوانٍ
|
| 342 |
"""
|
| 343 |
)
|
| 344 |
|
| 345 |
-
|
| 346 |
-
|
| 347 |
-
|
| 348 |
-
|
| 349 |
-
|
| 350 |
-
|
| 351 |
-
|
| 352 |
-
|
| 353 |
-
|
| 354 |
-
|
| 355 |
-
|
| 356 |
-
|
| 357 |
-
|
| 358 |
-
|
| 359 |
-
|
| 360 |
-
|
| 361 |
-
|
| 362 |
-
|
| 363 |
-
|
| 364 |
-
|
| 365 |
-
|
| 366 |
-
|
| 367 |
-
|
| 368 |
-
|
| 369 |
-
|
| 370 |
-
|
| 371 |
-
|
| 372 |
-
|
| 373 |
-
|
| 374 |
-
|
| 375 |
-
|
| 376 |
-
|
| 377 |
-
|
| 378 |
-
|
| 379 |
-
|
| 380 |
-
|
| 381 |
-
|
| 382 |
-
|
| 383 |
|
| 384 |
-
|
| 385 |
-
|
| 386 |
-
|
| 387 |
-
|
| 388 |
-
|
| 389 |
-
|
| 390 |
-
|
| 391 |
-
|
| 392 |
-
|
| 393 |
-
|
| 394 |
-
|
| 395 |
|
| 396 |
-
|
| 397 |
-
|
| 398 |
-
|
| 399 |
-
|
| 400 |
-
|
| 401 |
-
|
| 402 |
-
|
| 403 |
-
|
| 404 |
-
|
| 405 |
-
|
| 406 |
-
|
| 407 |
-
|
| 408 |
-
|
| 409 |
-
|
| 410 |
|
| 411 |
return demo
|
| 412 |
|
|
@@ -414,5 +451,5 @@ def create_interface():
|
|
| 414 |
demo = create_interface()
|
| 415 |
|
| 416 |
if __name__ == "__main__":
|
| 417 |
-
#
|
| 418 |
demo.launch(ssr_mode=False)
|
|
|
|
| 7 |
from transformers import Qwen3OmniMoeForConditionalGeneration, Qwen3OmniMoeProcessor
|
| 8 |
from qwen_omni_utils import process_mm_info
|
| 9 |
|
| 10 |
+
# =========================================================
|
| 11 |
+
# Patch لتجاوز مشكلة lm_head في Qwen3OmniMoeTalker*
|
| 12 |
+
# =========================================================
|
| 13 |
+
|
| 14 |
+
def _patched_mark_tied_weights_as_initialized(self):
|
| 15 |
+
"""
|
| 16 |
+
بعض إصدارات transformers + Qwen3-Omni تسبب خطأ:
|
| 17 |
+
Qwen3OmniMoeTalkerForConditionalGeneration has no attribute `lm_head`
|
| 18 |
+
عند محاولة ربط الـ tied weights.
|
| 19 |
+
هذا الـ patch يجعل هذه الخطوة no-op (لا تقوم بشيء)،
|
| 20 |
+
وهو آمن للاستنتاج (inference) في معظم الحالات.
|
| 21 |
+
"""
|
| 22 |
+
return
|
| 23 |
+
|
| 24 |
+
# تطبيق الـ patch قبل أي استدعاء لـ from_pretrained
|
| 25 |
+
if hasattr(Qwen3OmniMoeForConditionalGeneration, "mark_tied_weights_as_initialized"):
|
| 26 |
+
Qwen3OmniMoeForConditionalGeneration.mark_tied_weights_as_initialized = (
|
| 27 |
+
_patched_mark_tied_weights_as_initialized
|
| 28 |
+
)
|
| 29 |
+
|
| 30 |
+
# =========================================================
|
| 31 |
# إعدادات عامة
|
| 32 |
+
# =========================================================
|
| 33 |
|
| 34 |
MODEL_PATH = os.getenv("MODEL_PATH", "Qwen/Qwen3-Omni-30B-A3B-Instruct")
|
| 35 |
+
USE_AUDIO_IN_VIDEO = True # استخدام الصوت داخل الفيديو إذا وجد
|
| 36 |
|
| 37 |
VOICE_CHOICES = ["Ethan", "Chelsie", "Aiden"]
|
| 38 |
DEFAULT_VOICE = "Ethan"
|
| 39 |
|
| 40 |
+
# سنحمّل النموذج كسولياً (عند أول استدعاء فقط)
|
| 41 |
model = None
|
| 42 |
processor = None
|
| 43 |
|
| 44 |
|
| 45 |
def load_model():
|
| 46 |
"""
|
| 47 |
+
تحميل Qwen3-Omni والمعالج عند أول استدعاء فقط.
|
| 48 |
+
- نستخدم attn_implementation="eager" لتفادي الحاجة لـ flash-attn.
|
| 49 |
+
- لا نستخدم device_map="auto" لتفادي مشاكل توزيع الذاكرة على ZeroGPU.
|
| 50 |
"""
|
| 51 |
global model, processor
|
| 52 |
|
|
|
|
| 55 |
|
| 56 |
print(f"[ZeroGPU] Loading model from: {MODEL_PATH}")
|
| 57 |
|
| 58 |
+
# اختيار نوع البيانات والجهاز
|
| 59 |
if torch.cuda.is_available():
|
| 60 |
torch_dtype = torch.bfloat16
|
| 61 |
device = "cuda"
|
|
|
|
| 63 |
torch_dtype = torch.float32
|
| 64 |
device = "cpu"
|
| 65 |
|
| 66 |
+
# تحميل النموذج (بدون flash_attention_2)
|
| 67 |
local_model = Qwen3OmniMoeForConditionalGeneration.from_pretrained(
|
| 68 |
MODEL_PATH,
|
| 69 |
torch_dtype=torch_dtype,
|
| 70 |
+
attn_implementation="eager", # آمن على ZeroGPU بدون flash-attn
|
| 71 |
)
|
|
|
|
| 72 |
local_model.to(device)
|
| 73 |
|
| 74 |
local_processor = Qwen3OmniMoeProcessor.from_pretrained(MODEL_PATH)
|
|
|
|
| 78 |
print(f"[ZeroGPU] Model loaded on {device} with dtype {torch_dtype}.")
|
| 79 |
|
| 80 |
|
| 81 |
+
def build_messages_from_history(
|
| 82 |
+
history,
|
| 83 |
+
system_prompt,
|
| 84 |
+
user_text,
|
| 85 |
+
image,
|
| 86 |
+
audio_path,
|
| 87 |
+
video_path,
|
| 88 |
+
):
|
| 89 |
"""
|
| 90 |
+
تحويل تاريخ الدردشة + المدخل الحالي إلى conversation بالـ format
|
| 91 |
+
المطلوب من Qwen3-Omni.
|
| 92 |
history: list of [user_text, assistant_text]
|
| 93 |
"""
|
| 94 |
messages = []
|
| 95 |
|
| 96 |
if system_prompt:
|
| 97 |
+
messages.append(
|
| 98 |
+
{
|
| 99 |
+
"role": "system",
|
| 100 |
+
"content": [{"type": "text", "text": system_prompt}],
|
| 101 |
+
}
|
| 102 |
+
)
|
| 103 |
|
| 104 |
# تاريخ المحادثة
|
| 105 |
for user_msg, assistant_msg in history:
|
| 106 |
if user_msg:
|
| 107 |
+
messages.append(
|
| 108 |
+
{
|
| 109 |
+
"role": "user",
|
| 110 |
+
"content": [{"type": "text", "text": user_msg}],
|
| 111 |
+
}
|
| 112 |
+
)
|
| 113 |
if assistant_msg:
|
| 114 |
+
messages.append(
|
| 115 |
+
{
|
| 116 |
+
"role": "assistant",
|
| 117 |
+
"content": [{"type": "text", "text": assistant_msg}],
|
| 118 |
+
}
|
| 119 |
+
)
|
| 120 |
|
| 121 |
+
# محتوى رسالة المستخدم الحالية
|
| 122 |
user_content = []
|
| 123 |
|
| 124 |
if image is not None:
|
|
|
|
| 134 |
user_content.append({"type": "text", "text": user_text.strip()})
|
| 135 |
|
| 136 |
if user_content:
|
| 137 |
+
messages.append(
|
| 138 |
+
{
|
| 139 |
+
"role": "user",
|
| 140 |
+
"content": user_content,
|
| 141 |
+
}
|
| 142 |
+
)
|
| 143 |
|
| 144 |
return messages
|
| 145 |
|
| 146 |
|
| 147 |
+
# =========================================================
|
| 148 |
+
# دالة الاستدلال (تعمل على ZeroGPU)
|
| 149 |
+
# =========================================================
|
| 150 |
|
| 151 |
@spaces.GPU(duration=120)
|
| 152 |
def qwen3_omni_inference(
|
|
|
|
| 163 |
max_tokens,
|
| 164 |
):
|
| 165 |
"""
|
| 166 |
+
- تنفيذ الاستدلال على ZeroGPU.
|
| 167 |
+
- يدعم نص + صورة + صوت + فيديو في نفس الرسالة.
|
| 168 |
+
- مخرج نصي دائماً، ومخرج صوتي اختياري.
|
| 169 |
"""
|
| 170 |
|
| 171 |
+
# في حالة عدم وجود مداخل من المستخدم
|
| 172 |
if not (user_text or image is not None or audio_path or video_path):
|
|
|
|
| 173 |
return history, None, "", None, None, None
|
| 174 |
|
| 175 |
load_model()
|
|
|
|
| 184 |
video_path=video_path,
|
| 185 |
)
|
| 186 |
|
| 187 |
+
# بناء نص المحادثة باستخدام chat_template
|
| 188 |
text_prompt = processor.apply_chat_template(
|
| 189 |
messages,
|
| 190 |
add_generation_prompt=True,
|
| 191 |
tokenize=False,
|
| 192 |
)
|
| 193 |
|
| 194 |
+
# تجهيز الوسائط المتعددة
|
| 195 |
audios, images, videos = process_mm_info(
|
| 196 |
messages,
|
| 197 |
use_audio_in_video=USE_AUDIO_IN_VIDEO,
|
|
|
|
| 208 |
use_audio_in_video=USE_AUDIO_IN_VIDEO,
|
| 209 |
)
|
| 210 |
|
| 211 |
+
# نقل إلى جهاز النموذج ونفس dtype
|
| 212 |
first_param = next(model.parameters())
|
| 213 |
device = first_param.device
|
| 214 |
dtype = first_param.dtype
|
| 215 |
inputs = inputs.to(device=device, dtype=dtype)
|
| 216 |
|
| 217 |
+
# إعدادات التوليد
|
| 218 |
gen_kwargs = dict(
|
| 219 |
temperature=float(temperature),
|
| 220 |
top_p=float(top_p),
|
|
|
|
| 223 |
use_audio_in_video=USE_AUDIO_IN_VIDEO,
|
| 224 |
)
|
| 225 |
|
| 226 |
+
# توليد نص فقط أو نص + صوت
|
| 227 |
if not return_audio:
|
| 228 |
gen_kwargs["return_audio"] = False
|
| 229 |
+
text_ids, audio_out = model.generate(**inputs, **gen_kwargs)
|
| 230 |
audio_out = None
|
| 231 |
else:
|
| 232 |
gen_kwargs["speaker"] = speaker
|
| 233 |
text_ids, audio_out = model.generate(**inputs, **gen_kwargs)
|
| 234 |
|
| 235 |
+
# استخراج النص الناتج (بدون مدخل prompt)
|
| 236 |
input_len = inputs["input_ids"].shape[1]
|
| 237 |
generated_text = processor.batch_decode(
|
| 238 |
text_ids.sequences[:, input_len:],
|
|
|
|
| 241 |
)[0]
|
| 242 |
|
| 243 |
# تحديث تاريخ الدردشة
|
| 244 |
+
user_display = (
|
| 245 |
+
user_text if (user_text and user_text.strip()) else "[Multimodal message]"
|
| 246 |
+
)
|
| 247 |
history = history + [[user_display, generated_text]]
|
| 248 |
|
| 249 |
+
# تجهيز الصوت الناتج إن وجد
|
| 250 |
gr_audio = None
|
| 251 |
if audio_out is not None:
|
| 252 |
audio_np = audio_out.reshape(-1).detach().cpu().numpy()
|
|
|
|
| 257 |
return history, gr_audio, "", None, None, None
|
| 258 |
|
| 259 |
|
| 260 |
+
# =========================================================
|
| 261 |
+
# دوال واجهة Gradio
|
| 262 |
+
# =========================================================
|
| 263 |
|
| 264 |
def clear_chat():
|
| 265 |
+
"""إعادة تعيين المحادثة ومخرج الصوت."""
|
| 266 |
return [], None
|
| 267 |
|
| 268 |
|
|
|
|
| 274 |
"""
|
| 275 |
<h1 style="text-align:center;">Qwen3-Omni-30B-A3B – ZeroGPU Chat</h1>
|
| 276 |
<p style="text-align:center;">
|
| 277 |
+
دردشة متعددة الوسائط (نص + صورة + صوت + فيديو) تعمل على ZeroGPU.<br/>
|
| 278 |
+
اكتب رسالتك، ويمكنك إضافة صورة/صوت/فيديو، ثم اضغط <b>إرسال</b> أو Enter.<br/>
|
| 279 |
(لإضافة سطر جديد استخدم Shift+Enter)
|
| 280 |
</p>
|
| 281 |
"""
|
| 282 |
)
|
| 283 |
|
| 284 |
with gr.Row():
|
| 285 |
+
# العمود الأيسر: المحادثة
|
| 286 |
with gr.Column(scale=3):
|
| 287 |
chatbot = gr.Chatbot(
|
| 288 |
label="المحادثة",
|
|
|
|
| 295 |
autoplay=True,
|
| 296 |
)
|
| 297 |
|
|
|
|
| 298 |
with gr.Row():
|
| 299 |
user_text = gr.Textbox(
|
| 300 |
label="رسالتك",
|
|
|
|
| 323 |
send_btn = gr.Button("إرسال", variant="primary", scale=2)
|
| 324 |
clear_btn = gr.Button("مسح المحادثة", variant="secondary")
|
| 325 |
|
| 326 |
+
# العمود الأيمن: الإعدادات
|
| 327 |
with gr.Column(scale=1):
|
| 328 |
gr.Markdown("### إعدادات النموذج")
|
| 329 |
|
|
|
|
| 372 |
|
| 373 |
gr.Markdown(
|
| 374 |
"""
|
| 375 |
+
**ملاحظات:**
|
| 376 |
+
- يمكنك إرسال نص فقط، أو نص مع صورة/صوت/فيديو في رسالة واحدة.
|
| 377 |
+
- Enter للإرسال، وShift+Enter لسطر جديد.
|
| 378 |
+
- تشغيل النموذج على ZeroGPU قد يستغرق عدة ثوانٍ حسب طول الرسالة.
|
| 379 |
"""
|
| 380 |
)
|
| 381 |
|
| 382 |
+
# حالة المحادثة
|
| 383 |
+
history_state = gr.State([])
|
| 384 |
+
|
| 385 |
+
# مدخلات دالة الإرسال
|
| 386 |
+
send_inputs = [
|
| 387 |
+
history_state,
|
| 388 |
+
user_text,
|
| 389 |
+
image_input,
|
| 390 |
+
audio_input,
|
| 391 |
+
video_input,
|
| 392 |
+
system_prompt,
|
| 393 |
+
return_audio,
|
| 394 |
+
speaker,
|
| 395 |
+
temperature,
|
| 396 |
+
top_p,
|
| 397 |
+
max_tokens,
|
| 398 |
+
]
|
| 399 |
+
|
| 400 |
+
send_outputs = [
|
| 401 |
+
history_state,
|
| 402 |
+
audio_output,
|
| 403 |
+
user_text,
|
| 404 |
+
image_input,
|
| 405 |
+
audio_input,
|
| 406 |
+
video_input,
|
| 407 |
+
]
|
| 408 |
+
|
| 409 |
+
# إرسال بالزر
|
| 410 |
+
send_btn.click(
|
| 411 |
+
fn=qwen3_omni_inference,
|
| 412 |
+
inputs=send_inputs,
|
| 413 |
+
outputs=send_outputs,
|
| 414 |
+
queue=True,
|
| 415 |
+
).then(
|
| 416 |
+
lambda h: h,
|
| 417 |
+
inputs=history_state,
|
| 418 |
+
outputs=chatbot,
|
| 419 |
+
)
|
| 420 |
|
| 421 |
+
# إرسال بالـ Enter من Textbox
|
| 422 |
+
user_text.submit(
|
| 423 |
+
fn=qwen3_omni_inference,
|
| 424 |
+
inputs=send_inputs,
|
| 425 |
+
outputs=send_outputs,
|
| 426 |
+
queue=True,
|
| 427 |
+
).then(
|
| 428 |
+
lambda h: h,
|
| 429 |
+
inputs=history_state,
|
| 430 |
+
outputs=chatbot,
|
| 431 |
+
)
|
| 432 |
|
| 433 |
+
# مسح المحادثة
|
| 434 |
+
clear_btn.click(
|
| 435 |
+
fn=clear_chat,
|
| 436 |
+
inputs=None,
|
| 437 |
+
outputs=[history_state, audio_output],
|
| 438 |
+
).then(
|
| 439 |
+
lambda: [],
|
| 440 |
+
inputs=None,
|
| 441 |
+
outputs=chatbot,
|
| 442 |
+
).then(
|
| 443 |
+
lambda: ("", None, None),
|
| 444 |
+
inputs=None,
|
| 445 |
+
outputs=[user_text, image_input, audio_input],
|
| 446 |
+
)
|
| 447 |
|
| 448 |
return demo
|
| 449 |
|
|
|
|
| 451 |
demo = create_interface()
|
| 452 |
|
| 453 |
if __name__ == "__main__":
|
| 454 |
+
# إيقاف SSR لأنه تجريبي وقد يسبب مشكلة "Starting..."
|
| 455 |
demo.launch(ssr_mode=False)
|