multimodalart HF Staff commited on
Commit
929cee3
Β·
verified Β·
1 Parent(s): 81df4d4

Upload app.py with huggingface_hub

Browse files
Files changed (1) hide show
  1. app.py +2 -1
app.py CHANGED
@@ -188,7 +188,8 @@ def _caller_ip_token() -> str | None:
188
  # ── GPU inference ─────────────────────────────────────────────────────────────
189
 
190
 
191
- @spaces.GPU(duration=240, size=GPU_SIZE)
 
192
  @torch.no_grad() # render() is grad-free end to end; without this the VAE decode of the merge
193
  # builds a full autograd graph (~24 GB of retained activations) and OOMs the worker
194
  def _generate(prompt_embeds, seed, steps, switch, hr_steps, feather, alpha, use_head, box):
 
188
  # ── GPU inference ─────────────────────────────────────────────────────────────
189
 
190
 
191
+ @spaces.GPU(duration=60, size=GPU_SIZE) # measured 33s end to end (sampler 32s); 60 covers a cold
192
+ # worker attach + the sliders' max settings (steps 30)
193
  @torch.no_grad() # render() is grad-free end to end; without this the VAE decode of the merge
194
  # builds a full autograd graph (~24 GB of retained activations) and OOMs the worker
195
  def _generate(prompt_embeds, seed, steps, switch, hr_steps, feather, alpha, use_head, box):