cuio commited on
Commit
ea7eb88
·
verified ·
1 Parent(s): 43349a0

Update app.py

Browse files
Files changed (1) hide show
  1. app.py +43 -26
app.py CHANGED
@@ -14,8 +14,26 @@
14
  # ------------------------------------------------------------------------
15
  """URSA TI2V application."""
16
 
17
- import argparse
 
 
18
  import os
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
19
  import sys
20
 
21
  import gradio as gr
@@ -66,16 +84,6 @@ _patch_diffnext_qwen3()
66
  # =========================
67
 
68
 
69
- # ✅ HF Token(从环境变量读取,不写死)
70
- os.environ["HF_TOKEN"] = os.environ.get("HF_TOKEN", "")
71
- os.environ["HUGGINGFACE_HUB_TOKEN"] = os.environ.get("HF_TOKEN", "")
72
- os.environ["HF_HUB_ENABLE_HUGGINGFACE_TRANSFER"] = "1"
73
- os.environ["HF_HUB_DOWNLOAD_TIMEOUT"] = "300"
74
-
75
- os.environ["TOKENIZERS_PARALLELISM"] = "true"
76
- os.environ["PYTORCH_CUDA_ALLOC_CONF"] = "expandable_segments:True"
77
-
78
-
79
  def parse_args():
80
  parser = argparse.ArgumentParser(description="Serve URSA TI2V application")
81
  parser.add_argument("--model", default="BAAI/URSA-0.6B-FSQ320", help="model path")
@@ -98,6 +106,25 @@ def crop_image(image, target_h, target_w):
98
  return np.array(image.resize((target_w, target_h), PIL.Image.Resampling.BILINEAR))
99
 
100
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
101
  def generate_image(
102
  prompt,
103
  negative_prompt,
@@ -114,6 +141,7 @@ def generate_image(
114
  return [export_to_image(image, quality=95) for image in images] + [seed]
115
 
116
 
 
117
  def generate_video(
118
  prompt,
119
  negative_prompt,
@@ -161,20 +189,6 @@ examples = [list(x) for x in zip(prompts, motion_scores)]
161
 
162
 
163
  if __name__ == "__main__":
164
- args = parse_args()
165
-
166
- device = torch.device("cuda" if torch.cuda.is_available() else "cpu", args.device)
167
- model_args = {
168
- "torch_dtype": getattr(torch, args.precision.lower()),
169
- "trust_remote_code": True,
170
- }
171
-
172
- pipe = URSAPipeline.from_pretrained(
173
- args.model,
174
- **model_args,
175
- resume_download=True,
176
- ).to(device)
177
-
178
  # Application
179
  app = gr.Blocks(css=css, theme="origin").__enter__()
180
  container = gr.Column(elem_id="col-container").__enter__()
@@ -244,4 +258,7 @@ if __name__ == "__main__":
244
  ],
245
  outputs=[result, seed],
246
  )
247
- app.__exit__(), app.launch(share=False, ssr_mode=False, max_threads=1)
 
 
 
 
14
  # ------------------------------------------------------------------------
15
  """URSA TI2V application."""
16
 
17
+ # =========================
18
+ # ✅ 改 1:环境变量必须最先设置(早于 torch / spaces / diffnext)
19
+ # =========================
20
  import os
21
+
22
+ # 只在 token 存在时设置,避免把 HF_TOKEN 污染成空字符串
23
+ if os.environ.get("HF_TOKEN"):
24
+ os.environ.setdefault("HUGGINGFACE_HUB_TOKEN", os.environ["HF_TOKEN"])
25
+ os.environ["HF_HUB_ENABLE_HUGGINGFACE_TRANSFER"] = "1"
26
+ os.environ["HF_HUB_DOWNLOAD_TIMEOUT"] = "300"
27
+ os.environ["TOKENIZERS_PARALLELISM"] = "true"
28
+ # 必须在 CUDA 初始化前设置,否则 expandable_segments 不生效
29
+ os.environ["PYTORCH_CUDA_ALLOC_CONF"] = "expandable_segments:True"
30
+
31
+ # =========================
32
+ # ✅ 改 2:import spaces 必须早于 torch(ZeroGPU 硬性要求)
33
+ # =========================
34
+ import spaces
35
+
36
+ import argparse
37
  import sys
38
 
39
  import gradio as gr
 
84
  # =========================
85
 
86
 
 
 
 
 
 
 
 
 
 
 
87
  def parse_args():
88
  parser = argparse.ArgumentParser(description="Serve URSA TI2V application")
89
  parser.add_argument("--model", default="BAAI/URSA-0.6B-FSQ320", help="model path")
 
106
  return np.array(image.resize((target_w, target_h), PIL.Image.Resampling.BILINEAR))
107
 
108
 
109
+ # =========================
110
+ # ✅ 改 3:模型加载放在【模块顶层】,不能放进 if __name__ == "__main__"
111
+ # ZeroGPU 只在模块级启用 CUDA 模拟并打包权重
112
+ # =========================
113
+ args = parse_args()
114
+
115
+ device = torch.device("cuda" if torch.cuda.is_available() else "cpu", args.device)
116
+ model_args = {
117
+ "torch_dtype": getattr(torch, args.precision.lower()),
118
+ "trust_remote_code": True,
119
+ }
120
+
121
+ pipe = URSAPipeline.from_pretrained(args.model, **model_args).to(device)
122
+
123
+
124
+ # =========================
125
+ # ✅ 改 4:给真正跑推理的函数加 @spaces.GPU
126
+ # =========================
127
+ @spaces.GPU(duration=60)
128
  def generate_image(
129
  prompt,
130
  negative_prompt,
 
141
  return [export_to_image(image, quality=95) for image in images] + [seed]
142
 
143
 
144
+ @spaces.GPU(duration=180)
145
  def generate_video(
146
  prompt,
147
  negative_prompt,
 
189
 
190
 
191
  if __name__ == "__main__":
 
 
 
 
 
 
 
 
 
 
 
 
 
 
192
  # Application
193
  app = gr.Blocks(css=css, theme="origin").__enter__()
194
  container = gr.Column(elem_id="col-container").__enter__()
 
258
  ],
259
  outputs=[result, seed],
260
  )
261
+ # ✅ 改 5:加 queue(default_concurrency_limit=1),避免多请求抢同一张卡
262
+ app.__exit__(), app.queue(default_concurrency_limit=1).launch(
263
+ share=False, ssr_mode=False, max_threads=1
264
+ )