cuio commited on
Commit
0801856
·
verified ·
1 Parent(s): 78cbf2f

Update app.py

Browse files
Files changed (1) hide show
  1. app.py +49 -4
app.py CHANGED
@@ -179,8 +179,50 @@ pipe = URSAPipeline.from_pretrained(args.model, **model_args).to(device)
179
 
180
  # =========================
181
  # ✅ 第 7 步:真正跑推理的函数加 @spaces.GPU
 
 
 
 
182
  # =========================
183
- @spaces.GPU(duration=90)
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
184
  def generate_image(
185
  prompt,
186
  negative_prompt,
@@ -197,7 +239,7 @@ def generate_image(
197
  return [export_to_image(image, quality=95) for image in images] + [seed]
198
 
199
 
200
- @spaces.GPU(duration=300)
201
  def generate_video(
202
  prompt,
203
  negative_prompt,
@@ -274,7 +316,9 @@ if __name__ == "__main__":
274
  randomize_seed = gr.Checkbox(label="Randomize seed", value=True)
275
  guidance_scale = gr.Slider(label="Guidance scale", minimum=1, maximum=10.0, step=0.1, value=7.0)
276
  with gr.Row():
277
- num_inference_steps = gr.Slider(label="Inference steps", minimum=1, maximum=100, step=1, value=50)
 
 
278
  options.__exit__(), input_col.__exit__()
279
 
280
  # Results
@@ -316,4 +360,5 @@ if __name__ == "__main__":
316
  )
317
  # 限制并发,避免多个请求抢同一张卡
318
  app.__exit__(), app.queue(default_concurrency_limit=1).launch(
319
- share=False, ssr_mode=False, max_threads=1)
 
 
179
 
180
  # =========================
181
  # ✅ 第 7 步:真正跑推理的函数加 @spaces.GPU
182
+ # ZeroGPU 对单次 duration 有两道独立检查:
183
+ # 1) 档位上限:超过 → 「illegal duration」(与剩余配额无关,直接拒绝)
184
+ # 2) 配额预扣:声明值 > 访客剩余配额 → 「quota exceeded」
185
+ # 两道都要求声明值越小越好,所以用【动态 duration】按实际步数估算。
186
  # =========================
187
+
188
+ # 非 PRO 档位的单次上限。若仍报 illegal duration,逐级下调到 60 / 30 再试。
189
+ MAX_DURATION = 120
190
+
191
+ # 实测基准(half Blackwell, 512x320):T2I 25 步约 4s,其中首步含预热约 1.3s
192
+ IMAGE_OVERHEAD, IMAGE_PER_STEP = 15, 0.5
193
+ VIDEO_OVERHEAD, VIDEO_PER_STEP = 20, 2.0
194
+
195
+
196
+ def _clamped(value):
197
+ return int(max(20, min(value, MAX_DURATION)))
198
+
199
+
200
+ def estimate_image_duration(
201
+ prompt,
202
+ negative_prompt,
203
+ seed,
204
+ randomize_seed,
205
+ guidance_scale,
206
+ num_inference_steps=25,
207
+ ):
208
+ return _clamped(IMAGE_OVERHEAD + num_inference_steps * IMAGE_PER_STEP)
209
+
210
+
211
+ def estimate_video_duration(
212
+ prompt,
213
+ negative_prompt,
214
+ image,
215
+ motion_score,
216
+ seed,
217
+ randomize_seed,
218
+ guidance_scale,
219
+ num_inference_steps,
220
+ output_type="np",
221
+ ):
222
+ return _clamped(VIDEO_OVERHEAD + num_inference_steps * VIDEO_PER_STEP)
223
+
224
+
225
+ @spaces.GPU(duration=estimate_image_duration)
226
  def generate_image(
227
  prompt,
228
  negative_prompt,
 
239
  return [export_to_image(image, quality=95) for image in images] + [seed]
240
 
241
 
242
+ @spaces.GPU(duration=estimate_video_duration)
243
  def generate_video(
244
  prompt,
245
  negative_prompt,
 
316
  randomize_seed = gr.Checkbox(label="Randomize seed", value=True)
317
  guidance_scale = gr.Slider(label="Guidance scale", minimum=1, maximum=10.0, step=0.1, value=7.0)
318
  with gr.Row():
319
+ # 上限从 100 收到 50:与 estimate_video_duration 的估算保持自洽,
320
+ # 避免访客拉满步数后必然撞上档位上限而被直接拒绝
321
+ num_inference_steps = gr.Slider(label="Inference steps", minimum=1, maximum=50, step=1, value=30)
322
  options.__exit__(), input_col.__exit__()
323
 
324
  # Results
 
360
  )
361
  # 限制并发,避免多个请求抢同一张卡
362
  app.__exit__(), app.queue(default_concurrency_limit=1).launch(
363
+ share=False, ssr_mode=False, max_threads=1
364
+ )