Update app.py
Browse files
app.py
CHANGED
|
@@ -23,7 +23,10 @@ processor = None
|
|
| 23 |
|
| 24 |
|
| 25 |
def load_model():
|
| 26 |
-
"""
|
|
|
|
|
|
|
|
|
|
| 27 |
global model, processor
|
| 28 |
|
| 29 |
if model is not None and processor is not None:
|
|
@@ -31,27 +34,28 @@ def load_model():
|
|
| 31 |
|
| 32 |
print(f"[ZeroGPU] Loading model from: {MODEL_PATH}")
|
| 33 |
|
| 34 |
-
|
| 35 |
-
|
| 36 |
-
|
| 37 |
-
|
| 38 |
-
|
| 39 |
-
|
| 40 |
-
|
| 41 |
-
|
| 42 |
-
|
| 43 |
-
|
| 44 |
-
|
| 45 |
-
|
| 46 |
-
|
| 47 |
-
|
| 48 |
-
|
| 49 |
-
|
| 50 |
|
| 51 |
local_processor = Qwen3OmniMoeProcessor.from_pretrained(MODEL_PATH)
|
| 52 |
|
| 53 |
model = local_model
|
| 54 |
processor = local_processor
|
|
|
|
| 55 |
|
| 56 |
|
| 57 |
def build_messages_from_history(history, system_prompt, user_text, image, audio_path, video_path):
|
|
@@ -122,10 +126,14 @@ def qwen3_omni_inference(
|
|
| 122 |
top_p,
|
| 123 |
max_tokens,
|
| 124 |
):
|
| 125 |
-
"""
|
|
|
|
|
|
|
|
|
|
|
|
|
| 126 |
|
| 127 |
if not (user_text or image is not None or audio_path or video_path):
|
| 128 |
-
# لا يوجد مدخل
|
| 129 |
return history, None, "", None, None, None
|
| 130 |
|
| 131 |
load_model()
|
|
@@ -140,17 +148,20 @@ def qwen3_omni_inference(
|
|
| 140 |
video_path=video_path,
|
| 141 |
)
|
| 142 |
|
|
|
|
| 143 |
text_prompt = processor.apply_chat_template(
|
| 144 |
messages,
|
| 145 |
add_generation_prompt=True,
|
| 146 |
tokenize=False,
|
| 147 |
)
|
| 148 |
|
|
|
|
| 149 |
audios, images, videos = process_mm_info(
|
| 150 |
messages,
|
| 151 |
use_audio_in_video=USE_AUDIO_IN_VIDEO,
|
| 152 |
)
|
| 153 |
|
|
|
|
| 154 |
inputs = processor(
|
| 155 |
text=text_prompt,
|
| 156 |
audio=audios,
|
|
@@ -161,8 +172,13 @@ def qwen3_omni_inference(
|
|
| 161 |
use_audio_in_video=USE_AUDIO_IN_VIDEO,
|
| 162 |
)
|
| 163 |
|
| 164 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
| 165 |
|
|
|
|
| 166 |
gen_kwargs = dict(
|
| 167 |
temperature=float(temperature),
|
| 168 |
top_p=float(top_p),
|
|
@@ -171,6 +187,7 @@ def qwen3_omni_inference(
|
|
| 171 |
use_audio_in_video=USE_AUDIO_IN_VIDEO,
|
| 172 |
)
|
| 173 |
|
|
|
|
| 174 |
if not return_audio:
|
| 175 |
gen_kwargs["return_audio"] = False
|
| 176 |
text_ids, _ = model.generate(**inputs, **gen_kwargs)
|
|
@@ -179,15 +196,19 @@ def qwen3_omni_inference(
|
|
| 179 |
gen_kwargs["speaker"] = speaker
|
| 180 |
text_ids, audio_out = model.generate(**inputs, **gen_kwargs)
|
| 181 |
|
|
|
|
|
|
|
| 182 |
generated_text = processor.batch_decode(
|
| 183 |
-
text_ids.sequences[:,
|
| 184 |
skip_special_tokens=True,
|
| 185 |
clean_up_tokenization_spaces=False,
|
| 186 |
)[0]
|
| 187 |
|
|
|
|
| 188 |
user_display = user_text if (user_text and user_text.strip()) else "[Multimodal message]"
|
| 189 |
history = history + [[user_display, generated_text]]
|
| 190 |
|
|
|
|
| 191 |
gr_audio = None
|
| 192 |
if audio_out is not None:
|
| 193 |
audio_np = audio_out.reshape(-1).detach().cpu().numpy()
|
|
|
|
| 23 |
|
| 24 |
|
| 25 |
def load_model():
|
| 26 |
+
"""
|
| 27 |
+
تحميل Qwen3-Omni والمعالج عند أول استدعاء فقط (على ZeroGPU).
|
| 28 |
+
تم إلغاء flash_attention_2 و device_map='auto' لتجنب مشاكل الاستدعاء.
|
| 29 |
+
"""
|
| 30 |
global model, processor
|
| 31 |
|
| 32 |
if model is not None and processor is not None:
|
|
|
|
| 34 |
|
| 35 |
print(f"[ZeroGPU] Loading model from: {MODEL_PATH}")
|
| 36 |
|
| 37 |
+
# نحدد نوع البيانات والجهاز
|
| 38 |
+
if torch.cuda.is_available():
|
| 39 |
+
torch_dtype = torch.bfloat16
|
| 40 |
+
device = "cuda"
|
| 41 |
+
else:
|
| 42 |
+
torch_dtype = torch.float32
|
| 43 |
+
device = "cpu"
|
| 44 |
+
|
| 45 |
+
# تحميل النموذج بدون flash_attention_2 ولا device_map="auto"
|
| 46 |
+
local_model = Qwen3OmniMoeForConditionalGeneration.from_pretrained(
|
| 47 |
+
MODEL_PATH,
|
| 48 |
+
torch_dtype=torch_dtype,
|
| 49 |
+
attn_implementation="eager", # الأكثر أماناً في هذه البيئة
|
| 50 |
+
)
|
| 51 |
+
|
| 52 |
+
local_model.to(device)
|
| 53 |
|
| 54 |
local_processor = Qwen3OmniMoeProcessor.from_pretrained(MODEL_PATH)
|
| 55 |
|
| 56 |
model = local_model
|
| 57 |
processor = local_processor
|
| 58 |
+
print(f"[ZeroGPU] Model loaded on {device} with dtype {torch_dtype}.")
|
| 59 |
|
| 60 |
|
| 61 |
def build_messages_from_history(history, system_prompt, user_text, image, audio_path, video_path):
|
|
|
|
| 126 |
top_p,
|
| 127 |
max_tokens,
|
| 128 |
):
|
| 129 |
+
"""
|
| 130 |
+
تنفيذ الاستدلال الفعلي على ZeroGPU:
|
| 131 |
+
- نص + صورة + صوت + فيديو
|
| 132 |
+
- مخرج نصي دائماً، وصوتي عند الحاجة
|
| 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 |
video_path=video_path,
|
| 149 |
)
|
| 150 |
|
| 151 |
+
# بناء النص من المحادثة (chat template)
|
| 152 |
text_prompt = processor.apply_chat_template(
|
| 153 |
messages,
|
| 154 |
add_generation_prompt=True,
|
| 155 |
tokenize=False,
|
| 156 |
)
|
| 157 |
|
| 158 |
+
# تجهيز الوسائط المتعددة (صوت/صورة/فيديو)
|
| 159 |
audios, images, videos = process_mm_info(
|
| 160 |
messages,
|
| 161 |
use_audio_in_video=USE_AUDIO_IN_VIDEO,
|
| 162 |
)
|
| 163 |
|
| 164 |
+
# تحويل إلى تينسورات
|
| 165 |
inputs = processor(
|
| 166 |
text=text_prompt,
|
| 167 |
audio=audios,
|
|
|
|
| 172 |
use_audio_in_video=USE_AUDIO_IN_VIDEO,
|
| 173 |
)
|
| 174 |
|
| 175 |
+
# نقل المدخلات إلى نفس الجهاز ونفس dtype للنموذج
|
| 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 |
+
# بارامترات التوليد
|
| 182 |
gen_kwargs = dict(
|
| 183 |
temperature=float(temperature),
|
| 184 |
top_p=float(top_p),
|
|
|
|
| 187 |
use_audio_in_video=USE_AUDIO_IN_VIDEO,
|
| 188 |
)
|
| 189 |
|
| 190 |
+
# توليد النص فقط أو نص + صوت
|
| 191 |
if not return_audio:
|
| 192 |
gen_kwargs["return_audio"] = False
|
| 193 |
text_ids, _ = model.generate(**inputs, **gen_kwargs)
|
|
|
|
| 196 |
gen_kwargs["speaker"] = speaker
|
| 197 |
text_ids, audio_out = model.generate(**inputs, **gen_kwargs)
|
| 198 |
|
| 199 |
+
# فك ترميز النص الناتج
|
| 200 |
+
input_len = inputs["input_ids"].shape[1]
|
| 201 |
generated_text = processor.batch_decode(
|
| 202 |
+
text_ids.sequences[:, input_len:],
|
| 203 |
skip_special_tokens=True,
|
| 204 |
clean_up_tokenization_spaces=False,
|
| 205 |
)[0]
|
| 206 |
|
| 207 |
+
# تحديث تاريخ الدردشة
|
| 208 |
user_display = user_text if (user_text and user_text.strip()) else "[Multimodal message]"
|
| 209 |
history = history + [[user_display, generated_text]]
|
| 210 |
|
| 211 |
+
# تجهيز الصوت لمخرج Gradio (إن وجد)
|
| 212 |
gr_audio = None
|
| 213 |
if audio_out is not None:
|
| 214 |
audio_np = audio_out.reshape(-1).detach().cpu().numpy()
|