Spaces:
Build error
Build error
Update app.py
Browse files
app.py
CHANGED
@@ -1,23 +1,24 @@
|
|
1 |
import spaces
|
2 |
import gradio as gr
|
3 |
-
import argparse
|
4 |
import sys
|
5 |
-
import time
|
6 |
import os
|
7 |
import random
|
8 |
import subprocess
|
9 |
-
from PIL import Image #
|
|
|
|
|
10 |
|
11 |
-
subprocess.run(['sh', './sky.sh']) # Keep this if needed for setup
|
12 |
sys.path.append("./SkyReels-V1")
|
13 |
|
14 |
-
|
15 |
-
from skyreelsinfer
|
16 |
-
|
|
|
17 |
from diffusers.utils import export_to_video
|
18 |
-
# from diffusers.utils import load_image # Removed: Use PIL directly
|
19 |
|
20 |
import torch
|
|
|
21 |
|
22 |
torch.backends.cuda.matmul.allow_tf32 = False
|
23 |
torch.backends.cuda.matmul.allow_bf16_reduced_precision_reduction = False
|
@@ -25,204 +26,58 @@ torch.backends.cuda.matmul.allow_fp16_reduced_precision_reduction = False
|
|
25 |
torch.backends.cudnn.allow_tf32 = False
|
26 |
torch.backends.cudnn.deterministic = False
|
27 |
torch.backends.cudnn.benchmark = False
|
28 |
-
# torch.backends.cuda.preferred_blas_library="cublas"
|
29 |
-
# torch.backends.cuda.preferred_linalg_library="cusolver"
|
30 |
torch.set_float32_matmul_precision("highest")
|
31 |
|
32 |
device = torch.device("cuda:0" if torch.cuda.is_available() else "cpu")
|
33 |
|
34 |
-
import logging # Correct: Keep logging
|
35 |
-
|
36 |
-
# --- Dummy Classes (Keep these for standalone execution) ---
|
37 |
-
class OffloadConfig:
|
38 |
-
def __init__(self, high_cpu_memory=False, parameters_level=False, compiler_transformer=False, compiler_cache=""):
|
39 |
-
self.high_cpu_memory = high_cpu_memory
|
40 |
-
self.parameters_level = parameters_level
|
41 |
-
self.compiler_transformer = compiler_transformer
|
42 |
-
self.compiler_cache = compiler_cache
|
43 |
-
|
44 |
-
class TaskType:
|
45 |
-
T2V = 0
|
46 |
-
I2V = 1
|
47 |
-
|
48 |
-
class LlamaModel:
|
49 |
-
@staticmethod
|
50 |
-
def from_pretrained(*args, **kwargs):
|
51 |
-
return LlamaModel()
|
52 |
-
def to(self, device):
|
53 |
-
return self
|
54 |
-
|
55 |
-
class HunyuanVideoTransformer3DModel:
|
56 |
-
@staticmethod
|
57 |
-
def from_pretrained(*args, **kwargs):
|
58 |
-
return HunyuanVideoTransformer3DModel()
|
59 |
-
def to(self, device):
|
60 |
-
return self
|
61 |
-
|
62 |
-
class SkyreelsVideoPipeline:
|
63 |
-
@staticmethod
|
64 |
-
def from_pretrained(*args, **kwargs):
|
65 |
-
return SkyreelsVideoPipeline()
|
66 |
-
def to(self, device):
|
67 |
-
return self
|
68 |
-
def __call__(self, *args, **kwargs):
|
69 |
-
frames = [torch.randn(1, 3, 512, 512)] # Dummy frames
|
70 |
-
return type('obj', (object,), {'frames' : frames})()
|
71 |
-
class vae:
|
72 |
-
@staticmethod
|
73 |
-
def enable_tiling():
|
74 |
-
return
|
75 |
-
|
76 |
-
def quantize_(*args, **kwargs):
|
77 |
-
return
|
78 |
-
|
79 |
-
def float8_weight_only():
|
80 |
-
return
|
81 |
-
# --- End of Dummy Classes/Functions ---
|
82 |
logger = logging.getLogger(__name__)
|
83 |
|
84 |
-
|
85 |
-
def __init__(
|
86 |
-
self,
|
87 |
-
task_type: TaskType,
|
88 |
-
model_id: str,
|
89 |
-
quant_model: bool = True,
|
90 |
-
is_offload: bool = True,
|
91 |
-
offload_config: OffloadConfig = OffloadConfig(),
|
92 |
-
enable_cfg_parallel: bool = True, # Remove world_size, local_rank
|
93 |
-
):
|
94 |
-
self.task_type = task_type
|
95 |
-
self.model_id = model_id
|
96 |
-
self.quant_model = quant_model
|
97 |
-
self.is_offload = is_offload
|
98 |
-
self.offload_config = offload_config
|
99 |
-
self.enable_cfg_parallel = enable_cfg_parallel # Keep this
|
100 |
-
self.pipe = None
|
101 |
-
self.is_initialized = False
|
102 |
-
self.gpu_device = None
|
103 |
-
|
104 |
-
def _load_model(self, model_id: str, base_model_id: str = "hunyuanvideo-community/HunyuanVideo", quant_model: bool = True):
|
105 |
-
logger.info(f"load model model_id:{model_id} quan_model:{quant_model}")
|
106 |
-
text_encoder = LlamaModel.from_pretrained(
|
107 |
-
base_model_id, subfolder="text_encoder", torch_dtype=torch.bfloat16
|
108 |
-
).to("cpu")
|
109 |
-
transformer = HunyuanVideoTransformer3DModel.from_pretrained(
|
110 |
-
model_id, torch_dtype=torch.bfloat16, device="cpu"
|
111 |
-
).to("cpu")
|
112 |
-
|
113 |
-
if quant_model:
|
114 |
-
quantize_(text_encoder, float8_weight_only())
|
115 |
-
text_encoder.to("cpu")
|
116 |
-
torch.cuda.empty_cache()
|
117 |
-
quantize_(transformer, float8_weight_only())
|
118 |
-
transformer.to("cpu")
|
119 |
-
torch.cuda.empty_cache()
|
120 |
-
|
121 |
-
pipe = SkyreelsVideoPipeline.from_pretrained(
|
122 |
-
base_model_id, transformer=transformer, text_encoder=text_encoder, torch_dtype=torch.bfloat16
|
123 |
-
).to("cpu")
|
124 |
-
pipe.vae.enable_tiling()
|
125 |
-
torch.cuda.empty_cache()
|
126 |
-
return pipe
|
127 |
-
|
128 |
-
def initialize(self):
|
129 |
-
"""Initializes the model and moves it to the GPU."""
|
130 |
-
if self.is_initialized:
|
131 |
-
return
|
132 |
-
|
133 |
-
if not torch.cuda.is_available():
|
134 |
-
raise RuntimeError("CUDA is not available. Cannot initialize model.")
|
135 |
-
|
136 |
-
self.gpu_device = "cuda:0" # Always cuda:0 in single-GPU case
|
137 |
-
self.pipe = self._load_model(model_id=self.model_id, quant_model=self.quant_model)
|
138 |
-
|
139 |
-
# Simplified: No need for max_batch_dim_size with single GPU
|
140 |
-
if self.is_offload:
|
141 |
-
pass
|
142 |
-
else:
|
143 |
-
self.pipe.to(self.gpu_device)
|
144 |
-
|
145 |
-
if self.offload_config.compiler_transformer:
|
146 |
-
torch._dynamo.config.suppress_errors = True
|
147 |
-
os.environ["TORCHINDUCTOR_FX_GRAPH_CACHE"] = "1"
|
148 |
-
# No world_size in cache directory name
|
149 |
-
os.environ["TORCHINDUCTOR_CACHE_DIR"] = f"{self.offload_config.compiler_cache}"
|
150 |
-
self.pipe.transformer = torch.compile(
|
151 |
-
self.pipe.transformer, mode="max-autotune-no-cudagraphs", dynamic=True
|
152 |
-
)
|
153 |
-
if self.offload_config.compiler_transformer: # Only warm up if compiling
|
154 |
-
self.warm_up()
|
155 |
-
self.is_initialized = True
|
156 |
|
157 |
-
|
158 |
-
if not self.is_initialized:
|
159 |
-
raise RuntimeError("Model must be initialized before warm-up.")
|
160 |
|
161 |
-
|
162 |
-
|
163 |
-
"height": 544,
|
164 |
-
"width": 960,
|
165 |
-
"guidance_scale": 6,
|
166 |
-
"num_inference_steps": 1,
|
167 |
-
"negative_prompt": "bad quality",
|
168 |
-
"num_frames": 16,
|
169 |
-
"generator": torch.Generator(self.gpu_device).manual_seed(42),
|
170 |
-
"embedded_guidance_scale": 1.0,
|
171 |
-
}
|
172 |
-
if self.task_type == TaskType.I2V:
|
173 |
-
init_kwargs["image"] = Image.new("RGB", (544,960), color="black") #Dummy
|
174 |
-
self.pipe(**init_kwargs)
|
175 |
-
logger.info("Warm-up complete.")
|
176 |
|
177 |
-
|
178 |
-
"""Handles inference requests."""
|
179 |
-
if not self.is_initialized:
|
180 |
-
self.initialize()
|
181 |
-
if "seed" in kwargs:
|
182 |
-
kwargs["generator"] = torch.Generator(self.gpu_device).manual_seed(kwargs["seed"])
|
183 |
-
del kwargs["seed"]
|
184 |
-
assert (self.task_type == TaskType.I2V and "image" in kwargs) or self.task_type == TaskType.T2V
|
185 |
-
result = self.pipe(**kwargs).frames[0]
|
186 |
-
return result
|
187 |
-
|
188 |
-
|
189 |
-
# --- Spaces Integration ---
|
190 |
-
|
191 |
-
_predictor = None # Global variable to hold the predictor
|
192 |
-
|
193 |
-
@spaces.GPU(duration=90) # We DO need @spaces.GPU on init_predictor
|
194 |
def init_predictor():
|
195 |
global _predictor
|
196 |
-
|
|
|
197 |
|
198 |
if _predictor is None:
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
199 |
_predictor = SkyReelsVideoSingleGpuInfer(
|
200 |
-
task_type=
|
201 |
-
model_id=
|
202 |
quant_model=True,
|
203 |
is_offload=True,
|
204 |
offload_config=OffloadConfig(
|
205 |
high_cpu_memory=True,
|
206 |
parameters_level=True,
|
207 |
-
compiler_transformer=False,
|
208 |
),
|
209 |
)
|
210 |
-
_predictor.initialize()
|
211 |
logger.info("Predictor initialized")
|
212 |
else:
|
213 |
logger.warning("Predictor already initialized (should be rare).")
|
214 |
|
215 |
-
|
216 |
-
|
217 |
-
@spaces.GPU(duration=90) # Now needed, because we write files.
|
218 |
def generate_video(prompt, seed, image=None):
|
|
|
219 |
global task_type
|
220 |
-
global _predictor # Correct: Access global _predictor
|
221 |
|
222 |
-
print(f"image:{type(image)}")
|
223 |
if seed == -1:
|
224 |
-
random.seed(time
|
225 |
seed = int(random.randrange(4294967294))
|
|
|
226 |
kwargs = {
|
227 |
"prompt": prompt,
|
228 |
"height": 512,
|
@@ -235,38 +90,60 @@ def generate_video(prompt, seed, image=None):
|
|
235 |
"negative_prompt": "Aerial view, aerial view, overexposed, low quality, deformation, a poor composition, bad hands, bad teeth, bad eyes, bad limbs, distortion",
|
236 |
"cfg_for": False,
|
237 |
}
|
238 |
-
|
239 |
-
|
240 |
-
|
|
|
|
|
|
|
|
|
|
|
241 |
|
242 |
if _predictor is None:
|
243 |
init_predictor()
|
244 |
|
245 |
-
output = _predictor.infer(**kwargs)
|
246 |
|
247 |
-
save_dir = f"./result/{task_type}"
|
248 |
os.makedirs(save_dir, exist_ok=True)
|
249 |
-
video_out_file = f"{save_dir}/{prompt[:100].replace('/','')}_{seed}.mp4"
|
250 |
print(f"generate video, local path: {video_out_file}")
|
251 |
export_to_video(output, video_out_file, fps=24)
|
252 |
-
return video_out_file, kwargs
|
|
|
253 |
|
254 |
def create_gradio_interface():
|
255 |
-
|
256 |
-
|
|
|
257 |
image = gr.Image(label="Upload Image", type="filepath")
|
258 |
prompt = gr.Textbox(label="Input Prompt")
|
259 |
-
seed = gr.Number(label="Random Seed", value=-1)
|
260 |
-
|
261 |
-
|
262 |
-
|
263 |
-
|
264 |
-
|
265 |
-
|
266 |
-
|
267 |
-
|
268 |
-
|
|
|
|
|
|
|
269 |
|
270 |
if __name__ == "__main__":
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
271 |
demo = create_gradio_interface()
|
272 |
-
demo.queue().launch() # Add queue
|
|
|
1 |
import spaces
|
2 |
import gradio as gr
|
3 |
+
import argparse # Import argparse
|
4 |
import sys
|
|
|
5 |
import os
|
6 |
import random
|
7 |
import subprocess
|
8 |
+
from PIL import Image # Keep PIL import
|
9 |
+
|
10 |
+
subprocess.run(['sh', './sky.sh']) # Keep if needed
|
11 |
|
|
|
12 |
sys.path.append("./SkyReels-V1")
|
13 |
|
14 |
+
# Corrected Relative Imports
|
15 |
+
from SkyReels-V1.skyreelsinfer import TaskType # Now imported correctly
|
16 |
+
from SkyReels-V1.skyreelsinfer.offload import OffloadConfig
|
17 |
+
from SkyReels-V1.skyreelsinfer.skyreels_video_infer import SkyReelsVideoSingleGpuInfer # Import the class
|
18 |
from diffusers.utils import export_to_video
|
|
|
19 |
|
20 |
import torch
|
21 |
+
import logging
|
22 |
|
23 |
torch.backends.cuda.matmul.allow_tf32 = False
|
24 |
torch.backends.cuda.matmul.allow_bf16_reduced_precision_reduction = False
|
|
|
26 |
torch.backends.cudnn.allow_tf32 = False
|
27 |
torch.backends.cudnn.deterministic = False
|
28 |
torch.backends.cudnn.benchmark = False
|
|
|
|
|
29 |
torch.set_float32_matmul_precision("highest")
|
30 |
|
31 |
device = torch.device("cuda:0" if torch.cuda.is_available() else "cpu")
|
32 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
33 |
logger = logging.getLogger(__name__)
|
34 |
|
35 |
+
# --- Dummy Classes (Moved to skyreelsinfer/__init__.py) ---
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
36 |
|
37 |
+
# --- Global Variables and Argument Parsing ---
|
|
|
|
|
38 |
|
39 |
+
_predictor = None
|
40 |
+
task_type = TaskType.I2V # Default task type. IMPORTANT: Set a default.
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
41 |
|
42 |
+
@spaces.GPU(duration=90)
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
43 |
def init_predictor():
|
44 |
global _predictor
|
45 |
+
global task_type # Access the global task_type
|
46 |
+
logger = logging.getLogger(__name__)
|
47 |
|
48 |
if _predictor is None:
|
49 |
+
if task_type == TaskType.I2V:
|
50 |
+
model_id = "Skywork/SkyReels-V1-Hunyuan-I2V"
|
51 |
+
elif task_type == TaskType.T2V:
|
52 |
+
model_id = "your_t2v_model_id" # Replace with your T2V model ID
|
53 |
+
else:
|
54 |
+
raise ValueError(f"Invalid task_type: {task_type}")
|
55 |
+
|
56 |
_predictor = SkyReelsVideoSingleGpuInfer(
|
57 |
+
task_type=task_type,
|
58 |
+
model_id=model_id,
|
59 |
quant_model=True,
|
60 |
is_offload=True,
|
61 |
offload_config=OffloadConfig(
|
62 |
high_cpu_memory=True,
|
63 |
parameters_level=True,
|
64 |
+
compiler_transformer=False,
|
65 |
),
|
66 |
)
|
67 |
+
_predictor.initialize()
|
68 |
logger.info("Predictor initialized")
|
69 |
else:
|
70 |
logger.warning("Predictor already initialized (should be rare).")
|
71 |
|
72 |
+
@spaces.GPU(duration=90) # Needed, because we are saving a file
|
|
|
|
|
73 |
def generate_video(prompt, seed, image=None):
|
74 |
+
global _predictor
|
75 |
global task_type
|
|
|
76 |
|
|
|
77 |
if seed == -1:
|
78 |
+
random.seed() # Use system time for randomness if seed is -1
|
79 |
seed = int(random.randrange(4294967294))
|
80 |
+
|
81 |
kwargs = {
|
82 |
"prompt": prompt,
|
83 |
"height": 512,
|
|
|
90 |
"negative_prompt": "Aerial view, aerial view, overexposed, low quality, deformation, a poor composition, bad hands, bad teeth, bad eyes, bad limbs, distortion",
|
91 |
"cfg_for": False,
|
92 |
}
|
93 |
+
|
94 |
+
if task_type == TaskType.I2V:
|
95 |
+
assert image is not None, "Please input an image for I2V task."
|
96 |
+
kwargs["image"] = Image.open(image) # Use PIL.Image.open
|
97 |
+
elif task_type == TaskType.T2V:
|
98 |
+
pass # No image needed.
|
99 |
+
else:
|
100 |
+
raise ValueError("Invalid Tasktype")
|
101 |
|
102 |
if _predictor is None:
|
103 |
init_predictor()
|
104 |
|
105 |
+
output = _predictor.infer(**kwargs)
|
106 |
|
107 |
+
save_dir = f"./result/{task_type.name}" # Use task_type.name for directory
|
108 |
os.makedirs(save_dir, exist_ok=True)
|
109 |
+
video_out_file = f"{save_dir}/{prompt[:100].replace('/', '')}_{seed}.mp4"
|
110 |
print(f"generate video, local path: {video_out_file}")
|
111 |
export_to_video(output, video_out_file, fps=24)
|
112 |
+
return video_out_file, kwargs # Return the file path
|
113 |
+
|
114 |
|
115 |
def create_gradio_interface():
|
116 |
+
with gr.Blocks() as demo:
|
117 |
+
with gr.Row():
|
118 |
+
with gr.Column():
|
119 |
image = gr.Image(label="Upload Image", type="filepath")
|
120 |
prompt = gr.Textbox(label="Input Prompt")
|
121 |
+
seed = gr.Number(label="Random Seed", value=-1) # Default to -1
|
122 |
+
with gr.Column():
|
123 |
+
submit_button = gr.Button("Generate Video")
|
124 |
+
output_video = gr.Video(label="Generated Video")
|
125 |
+
output_params = gr.Textbox(label="Output Parameters")
|
126 |
+
|
127 |
+
submit_button.click(
|
128 |
+
fn=generate_video,
|
129 |
+
inputs=[prompt, seed, image],
|
130 |
+
outputs=[output_video, output_params],
|
131 |
+
)
|
132 |
+
return demo
|
133 |
+
|
134 |
|
135 |
if __name__ == "__main__":
|
136 |
+
parser = argparse.ArgumentParser()
|
137 |
+
parser.add_argument("--task_type", type=str, default="i2v", choices=["t2v", "i2v"],
|
138 |
+
help="Task type, 't2v' for text-to-video, 'i2v' for image-to-video.")
|
139 |
+
args = parser.parse_args()
|
140 |
+
|
141 |
+
# Set the global task_type based on command-line arguments
|
142 |
+
if args.task_type == "t2v":
|
143 |
+
task_type = TaskType.T2V
|
144 |
+
elif args.task_type == "i2v":
|
145 |
+
task_type = TaskType.I2V
|
146 |
+
# No else needed, default is already set
|
147 |
+
|
148 |
demo = create_gradio_interface()
|
149 |
+
demo.queue().launch() # Add queue
|