meepmoo commited on
Commit
7712d70
·
verified ·
1 Parent(s): e7ccefd

Update worker_runpod.py

Browse files
Files changed (1) hide show
  1. worker_runpod.py +12 -12
worker_runpod.py CHANGED
@@ -25,7 +25,7 @@ tokenxf = os.getenv("HF_API_TOKEN")
25
  # Low GPU memory mode
26
  low_gpu_memory_mode = False
27
  lora_path = "/content/shirtlift.safetensors"
28
-
29
  def to_pil(image):
30
  if isinstance(image, Image.Image):
31
  return image
@@ -75,13 +75,13 @@ with torch.inference_mode():
75
  model_id = "/runpod-volume/model"
76
  transformer = CogVideoXTransformer3DModel.from_pretrained_2d(
77
  model_id, subfolder="transformer"
78
- ).to(torch.bfloat16)
79
 
80
  vae = AutoencoderKLCogVideoX.from_pretrained(
81
  model_id, subfolder="vae"
82
- ).to(torch.bfloat16)
83
 
84
- text_encoder = T5EncoderModel.from_pretrained(model_id, subfolder="text_encoder")
85
 
86
  sampler_dict = {
87
  "Euler": EulerDiscreteScheduler,
@@ -102,7 +102,7 @@ with torch.inference_mode():
102
  text_encoder=text_encoder,
103
  transformer=transformer,
104
  scheduler=scheduler,
105
- torch_dtype=torch.bfloat16
106
  )
107
  else:
108
  pipeline = CogVideoX_Fun_Pipeline.from_pretrained(
@@ -111,16 +111,16 @@ with torch.inference_mode():
111
  text_encoder=text_encoder,
112
  transformer=transformer,
113
  scheduler=scheduler,
114
- torch_dtype=torch.bfloat16
115
  )
116
 
117
  pipeline = merge_lora(pipeline, lora_path, 1.00)
118
 
119
 
120
- # if low_gpu_memory_mode:
121
- # pipeline.enable_sequential_cpu_offload()
122
- # else:
123
- # pipeline.enable_model_cpu_offload()
124
 
125
 
126
 
@@ -130,13 +130,13 @@ def generate(input):
130
  prompt = values["prompt"]
131
  print("starting Generate function")
132
  print(prompt)
133
- negative_prompt = values.get("negative_prompt", "blurry, blurred, blurry face")
134
  guidance_scale = values.get("guidance_scale", 6.0)
135
  seed = values.get("seed", 42)
136
  num_inference_steps = values.get("num_inference_steps", 18)
137
  base_resolution = values.get("base_resolution", 512)
138
 
139
- video_length = values.get("video_length", 53)
140
  fps = values.get("fps", 10)
141
 
142
  save_path = "samples"
 
25
  # Low GPU memory mode
26
  low_gpu_memory_mode = False
27
  lora_path = "/content/shirtlift.safetensors"
28
+ weight_dtype = torch.bfloat16
29
  def to_pil(image):
30
  if isinstance(image, Image.Image):
31
  return image
 
75
  model_id = "/runpod-volume/model"
76
  transformer = CogVideoXTransformer3DModel.from_pretrained_2d(
77
  model_id, subfolder="transformer"
78
+ ).to(weight_dtype)
79
 
80
  vae = AutoencoderKLCogVideoX.from_pretrained(
81
  model_id, subfolder="vae"
82
+ ).to(weight_dtype)
83
 
84
+ text_encoder = T5EncoderModel.from_pretrained(model_id, subfolder="text_encoder", torch_dtype=weight_dtype)
85
 
86
  sampler_dict = {
87
  "Euler": EulerDiscreteScheduler,
 
102
  text_encoder=text_encoder,
103
  transformer=transformer,
104
  scheduler=scheduler,
105
+ torch_dtype=weight_dtype
106
  )
107
  else:
108
  pipeline = CogVideoX_Fun_Pipeline.from_pretrained(
 
111
  text_encoder=text_encoder,
112
  transformer=transformer,
113
  scheduler=scheduler,
114
+ torch_dtype=weight_dtype
115
  )
116
 
117
  pipeline = merge_lora(pipeline, lora_path, 1.00)
118
 
119
 
120
+ if low_gpu_memory_mode:
121
+ pipeline.enable_sequential_cpu_offload()
122
+ else:
123
+ pipeline.enable_model_cpu_offload()
124
 
125
 
126
 
 
130
  prompt = values["prompt"]
131
  print("starting Generate function")
132
  print(prompt)
133
+ negative_prompt = values.get("negative_prompt", "The video is not of a high quality, it has a low resolution. Watermark present in each frame. Strange motion trajectory. blurry, blurred, grainy, distortion, blurry face")
134
  guidance_scale = values.get("guidance_scale", 6.0)
135
  seed = values.get("seed", 42)
136
  num_inference_steps = values.get("num_inference_steps", 18)
137
  base_resolution = values.get("base_resolution", 512)
138
 
139
+ video_length = values.get("video_length", 49)
140
  fps = values.get("fps", 10)
141
 
142
  save_path = "samples"