multimodalart HF Staff commited on
Commit
0ea8f8a
·
verified ·
1 Parent(s): 30b6bc4

Upload app.py with huggingface_hub

Browse files
Files changed (1) hide show
  1. app.py +5 -5
app.py CHANGED
@@ -82,8 +82,11 @@ def _latent_frames(seconds: float, variant: str) -> int:
82
 
83
 
84
  def _duration(prompt, mode="chunkwise", seconds=5, seed=0, randomize_seed=True, *args, **kwargs):
85
- lat = _latent_frames(float(seconds), mode if mode in PIPES else "chunkwise")
86
- return int(min(170, 20 + lat * 1.2))
 
 
 
87
 
88
 
89
  @spaces.GPU(duration=_duration)
@@ -119,9 +122,6 @@ def generate(
119
  num_latents = _latent_frames(float(seconds), mode)
120
 
121
  t0 = time.perf_counter()
122
- free, total = torch.cuda.mem_get_info()
123
- print(f"[SGF+] GPU {torch.cuda.get_device_name()} free={free/2**30:.1f}G total={total/2**30:.1f}G "
124
- f"allocated={torch.cuda.memory_allocated()/2**30:.1f}G")
125
  pipe.kv_cache1 = None # fresh caches for every request
126
  torch.manual_seed(seed)
127
  noise = torch.randn([1, num_latents, 16, 60, 104], device="cuda", dtype=torch.bfloat16)
 
82
 
83
 
84
  def _duration(prompt, mode="chunkwise", seconds=5, seed=0, randomize_seed=True, *args, **kwargs):
85
+ mode = mode if mode in PIPES else "chunkwise"
86
+ lat = _latent_frames(float(seconds), mode)
87
+ # measured on ZeroGPU (MIG 2g.48gb): chunkwise ~0.75 s / latent, framewise ~1.45 s / latent
88
+ per_latent = 0.75 if mode == "chunkwise" else 1.45
89
+ return int(8 + lat * per_latent * 1.15)
90
 
91
 
92
  @spaces.GPU(duration=_duration)
 
122
  num_latents = _latent_frames(float(seconds), mode)
123
 
124
  t0 = time.perf_counter()
 
 
 
125
  pipe.kv_cache1 = None # fresh caches for every request
126
  torch.manual_seed(seed)
127
  noise = torch.randn([1, num_latents, 16, 60, 104], device="cuda", dtype=torch.bfloat16)