John Ho commited on
Commit
2a9891d
·
1 Parent(s): 7c03996

trying to set device outsite of spaces.GPU decorator

Browse files
Files changed (1) hide show
  1. app.py +6 -3
app.py CHANGED
@@ -22,6 +22,10 @@ subprocess.run(
22
 
23
  # The model is trained on 8.0 FPS which we recommend for optimal inference
24
 
 
 
 
 
25
 
26
  def get_fps_ffmpeg(video_path: str):
27
  probe = ffmpeg.probe(video_path)
@@ -37,7 +41,6 @@ def get_fps_ffmpeg(video_path: str):
37
  return num / denom
38
 
39
 
40
- @spaces.GPU(duration=30)
41
  def load_model(
42
  model_name: str = "chancharikm/qwen2.5-vl-7b-cam-motion-preview",
43
  use_flash_attention: bool = True,
@@ -54,7 +57,7 @@ def load_model(
54
  else Qwen2_5_VLForConditionalGeneration.from_pretrained(
55
  model_name,
56
  torch_dtype=torch.bfloat16, # "auto",
57
- device_map="auto",
58
  )
59
  )
60
  return model
@@ -98,7 +101,7 @@ def inference(
98
  return_tensors="pt",
99
  **video_kwargs,
100
  )
101
- inputs = inputs.to("cuda" if torch.cuda.is_available() else "cpu")
102
 
103
  # Inference
104
  generated_ids = model.generate(**inputs, max_new_tokens=128)
 
22
 
23
  # The model is trained on 8.0 FPS which we recommend for optimal inference
24
 
25
+ DTYPE = torch.bfloat16
26
+ DEVICE = "cuda" if torch.cuda.is_available() else "cpu"
27
+ logger.info(f"Device: {device}, dtype: {dtype}")
28
+
29
 
30
  def get_fps_ffmpeg(video_path: str):
31
  probe = ffmpeg.probe(video_path)
 
41
  return num / denom
42
 
43
 
 
44
  def load_model(
45
  model_name: str = "chancharikm/qwen2.5-vl-7b-cam-motion-preview",
46
  use_flash_attention: bool = True,
 
57
  else Qwen2_5_VLForConditionalGeneration.from_pretrained(
58
  model_name,
59
  torch_dtype=torch.bfloat16, # "auto",
60
+ device_map=DEVICE,
61
  )
62
  )
63
  return model
 
101
  return_tensors="pt",
102
  **video_kwargs,
103
  )
104
+ # inputs = inputs.to("cuda" if torch.cuda.is_available() else "cpu")
105
 
106
  # Inference
107
  generated_ids = model.generate(**inputs, max_new_tokens=128)