Spaces:
Running
on
L40S
Running
on
L40S
Update app.py
Browse files
app.py
CHANGED
@@ -14,6 +14,7 @@ from huggingface_hub import hf_hub_download
|
|
14 |
# Ensure 'checkpoint' directory exists
|
15 |
os.makedirs("checkpoints", exist_ok=True)
|
16 |
|
|
|
17 |
hf_hub_download(
|
18 |
repo_id="wenqsun/DimensionX",
|
19 |
filename="orbit_left_lora_weights.safetensors",
|
@@ -26,93 +27,78 @@ hf_hub_download(
|
|
26 |
local_dir="checkpoints"
|
27 |
)
|
28 |
|
|
|
29 |
model_id = "THUDM/CogVideoX-5b-I2V"
|
30 |
-
transformer = CogVideoXTransformer3DModel.from_pretrained(model_id, subfolder="transformer", torch_dtype=torch.float16)
|
31 |
-
text_encoder = T5EncoderModel.from_pretrained(model_id, subfolder="text_encoder", torch_dtype=torch.float16)
|
32 |
-
vae = AutoencoderKLCogVideoX.from_pretrained(model_id, subfolder="vae", torch_dtype=torch.float16)
|
33 |
tokenizer = T5Tokenizer.from_pretrained(model_id, subfolder="tokenizer")
|
34 |
-
|
35 |
|
36 |
def find_and_move_object_to_cpu():
|
37 |
for obj in gc.get_objects():
|
38 |
try:
|
39 |
-
# Check if the object is a PyTorch model
|
40 |
if isinstance(obj, torch.nn.Module):
|
41 |
-
# Check if any parameter of the model is on CUDA
|
42 |
if any(param.is_cuda for param in obj.parameters()):
|
43 |
-
print(f"Found PyTorch model on CUDA: {type(obj).__name__}")
|
44 |
-
# Move the model to CPU
|
45 |
obj.to('cpu')
|
46 |
-
print(f"Moved {type(obj).__name__} to CPU.")
|
47 |
-
|
48 |
-
# Optionally check if buffers are on CUDA
|
49 |
if any(buf.is_cuda for buf in obj.buffers()):
|
50 |
-
print(f"Found buffer on CUDA in {type(obj).__name__}")
|
51 |
obj.to('cpu')
|
52 |
-
print(f"Moved buffers of {type(obj).__name__} to CPU.")
|
53 |
-
|
54 |
except Exception as e:
|
55 |
-
# Handle any exceptions if obj is not a torch model
|
56 |
pass
|
57 |
|
58 |
-
|
59 |
def clear_gpu():
|
60 |
-
|
61 |
-
|
62 |
-
|
63 |
-
print(f"Memory reserved before clearing: {torch.cuda.memory_reserved() / (1024 ** 2)} MB")
|
64 |
-
|
65 |
-
# Move any bound tensors back to CPU if needed
|
66 |
-
if torch.cuda.is_available():
|
67 |
-
torch.cuda.empty_cache()
|
68 |
-
torch.cuda.synchronize() # Ensure that all operations are completed
|
69 |
-
print("GPU memory cleared.")
|
70 |
-
|
71 |
-
print(f"Memory allocated after clearing: {torch.cuda.memory_allocated() / (1024 ** 2)} MB")
|
72 |
-
print(f"Memory reserved after clearing: {torch.cuda.memory_reserved() / (1024 ** 2)} MB")
|
73 |
-
|
74 |
|
75 |
def infer(image_path, prompt, orbit_type, progress=gr.Progress(track_tqdm=True)):
|
76 |
|
77 |
-
|
78 |
lora_path = "checkpoints/"
|
79 |
-
adapter_name = None
|
80 |
if orbit_type == "Left":
|
81 |
weight_name = "orbit_left_lora_weights.safetensors"
|
82 |
elif orbit_type == "Up":
|
83 |
weight_name = "orbit_up_lora_weights.safetensors"
|
84 |
lora_rank = 256
|
85 |
|
|
|
|
|
86 |
# Generate a timestamp for adapter_name
|
87 |
adapter_timestamp = datetime.now().strftime("%Y%m%d_%H%M%S")
|
88 |
|
|
|
89 |
pipe.load_lora_weights(lora_path, weight_name=weight_name, adapter_name=f"{adapter_timestamp}")
|
90 |
pipe.fuse_lora(lora_scale=1 / lora_rank)
|
|
|
|
|
91 |
pipe.to("cuda")
|
92 |
-
|
93 |
-
|
94 |
prompt = f"{prompt}. High quality, ultrarealistic detail and breath-taking movie-like camera shot."
|
95 |
image = load_image(image_path)
|
96 |
seed = random.randint(0, 2**8 - 1)
|
97 |
|
|
|
98 |
video = pipe(
|
99 |
image,
|
100 |
prompt,
|
101 |
-
num_inference_steps=50,
|
102 |
-
guidance_scale=7.0,
|
103 |
use_dynamic_cfg=True,
|
104 |
generator=torch.Generator(device="cpu").manual_seed(seed)
|
105 |
)
|
106 |
-
|
107 |
-
find_and_move_object_to_cpu()
|
108 |
-
clear_gpu()
|
109 |
|
110 |
-
# Generate
|
111 |
timestamp = datetime.now().strftime("%Y%m%d_%H%M%S")
|
112 |
export_to_video(video.frames[0], f"output_{timestamp}.mp4", fps=8)
|
|
|
|
|
|
|
|
|
113 |
|
114 |
return f"output_{timestamp}.mp4"
|
115 |
|
|
|
116 |
with gr.Blocks(analytics_enabled=False) as demo:
|
117 |
with gr.Column(elem_id="col-container"):
|
118 |
gr.Markdown("# DimensionX")
|
|
|
14 |
# Ensure 'checkpoint' directory exists
|
15 |
os.makedirs("checkpoints", exist_ok=True)
|
16 |
|
17 |
+
# Download LoRA weights
|
18 |
hf_hub_download(
|
19 |
repo_id="wenqsun/DimensionX",
|
20 |
filename="orbit_left_lora_weights.safetensors",
|
|
|
27 |
local_dir="checkpoints"
|
28 |
)
|
29 |
|
30 |
+
# Load models in the global scope
|
31 |
model_id = "THUDM/CogVideoX-5b-I2V"
|
32 |
+
transformer = CogVideoXTransformer3DModel.from_pretrained(model_id, subfolder="transformer", torch_dtype=torch.float16).to("cpu")
|
33 |
+
text_encoder = T5EncoderModel.from_pretrained(model_id, subfolder="text_encoder", torch_dtype=torch.float16).to("cpu")
|
34 |
+
vae = AutoencoderKLCogVideoX.from_pretrained(model_id, subfolder="vae", torch_dtype=torch.float16).to("cpu")
|
35 |
tokenizer = T5Tokenizer.from_pretrained(model_id, subfolder="tokenizer")
|
36 |
+
pipe = CogVideoXImageToVideoPipeline.from_pretrained(model_id, tokenizer=tokenizer, text_encoder=text_encoder, transformer=transformer, vae=vae, torch_dtype=torch.float16)
|
37 |
|
38 |
def find_and_move_object_to_cpu():
|
39 |
for obj in gc.get_objects():
|
40 |
try:
|
|
|
41 |
if isinstance(obj, torch.nn.Module):
|
|
|
42 |
if any(param.is_cuda for param in obj.parameters()):
|
|
|
|
|
43 |
obj.to('cpu')
|
|
|
|
|
|
|
44 |
if any(buf.is_cuda for buf in obj.buffers()):
|
|
|
45 |
obj.to('cpu')
|
|
|
|
|
46 |
except Exception as e:
|
|
|
47 |
pass
|
48 |
|
|
|
49 |
def clear_gpu():
|
50 |
+
torch.cuda.empty_cache()
|
51 |
+
torch.cuda.synchronize()
|
52 |
+
gc.collect()
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
53 |
|
54 |
def infer(image_path, prompt, orbit_type, progress=gr.Progress(track_tqdm=True)):
|
55 |
|
56 |
+
|
57 |
lora_path = "checkpoints/"
|
|
|
58 |
if orbit_type == "Left":
|
59 |
weight_name = "orbit_left_lora_weights.safetensors"
|
60 |
elif orbit_type == "Up":
|
61 |
weight_name = "orbit_up_lora_weights.safetensors"
|
62 |
lora_rank = 256
|
63 |
|
64 |
+
pipe.unload_lora_weights()
|
65 |
+
|
66 |
# Generate a timestamp for adapter_name
|
67 |
adapter_timestamp = datetime.now().strftime("%Y%m%d_%H%M%S")
|
68 |
|
69 |
+
# Load LoRA weights on CPU, move to GPU afterward
|
70 |
pipe.load_lora_weights(lora_path, weight_name=weight_name, adapter_name=f"{adapter_timestamp}")
|
71 |
pipe.fuse_lora(lora_scale=1 / lora_rank)
|
72 |
+
|
73 |
+
# Move the pipeline to GPU for inference
|
74 |
pipe.to("cuda")
|
75 |
+
|
76 |
+
# Set the inference prompt
|
77 |
prompt = f"{prompt}. High quality, ultrarealistic detail and breath-taking movie-like camera shot."
|
78 |
image = load_image(image_path)
|
79 |
seed = random.randint(0, 2**8 - 1)
|
80 |
|
81 |
+
|
82 |
video = pipe(
|
83 |
image,
|
84 |
prompt,
|
85 |
+
num_inference_steps=50,
|
86 |
+
guidance_scale=7.0,
|
87 |
use_dynamic_cfg=True,
|
88 |
generator=torch.Generator(device="cpu").manual_seed(seed)
|
89 |
)
|
|
|
|
|
|
|
90 |
|
91 |
+
# Generate and save output video
|
92 |
timestamp = datetime.now().strftime("%Y%m%d_%H%M%S")
|
93 |
export_to_video(video.frames[0], f"output_{timestamp}.mp4", fps=8)
|
94 |
+
|
95 |
+
# Move objects to CPU and clear GPU memory immediately after inference
|
96 |
+
find_and_move_object_to_cpu()
|
97 |
+
clear_gpu()
|
98 |
|
99 |
return f"output_{timestamp}.mp4"
|
100 |
|
101 |
+
# Set up Gradio UI
|
102 |
with gr.Blocks(analytics_enabled=False) as demo:
|
103 |
with gr.Column(elem_id="col-container"):
|
104 |
gr.Markdown("# DimensionX")
|