Spaces:
Running
on
Zero
Running
on
Zero
Dynamic Axis
#10
by
multimodalart
HF Staff
- opened
- app.py +53 -25
- optimization.py +21 -45
app.py
CHANGED
@@ -19,8 +19,11 @@ from optimization import optimize_pipeline_
|
|
19 |
|
20 |
MODEL_ID = "Wan-AI/Wan2.2-I2V-A14B-Diffusers"
|
21 |
|
22 |
-
|
23 |
-
|
|
|
|
|
|
|
24 |
MAX_SEED = np.iinfo(np.int32).max
|
25 |
|
26 |
FIXED_FPS = 16
|
@@ -50,11 +53,14 @@ for i in range(3):
|
|
50 |
torch.cuda.synchronize()
|
51 |
torch.cuda.empty_cache()
|
52 |
|
|
|
|
|
|
|
53 |
optimize_pipeline_(pipe,
|
54 |
-
image=Image.new('RGB', (
|
55 |
prompt='prompt',
|
56 |
-
height=
|
57 |
-
width=
|
58 |
num_frames=MAX_FRAMES_MODEL,
|
59 |
)
|
60 |
|
@@ -62,28 +68,51 @@ optimize_pipeline_(pipe,
|
|
62 |
default_prompt_i2v = "make this image come alive, cinematic motion, smooth animation"
|
63 |
default_negative_prompt = "色调艳丽, 过曝, 静态, 细节模糊不清, 字幕, 风格, 作品, 画作, 画面, 静止, 整体发灰, 最差质量, 低质量, JPEG压缩残留, 丑陋的, 残缺的, 多余的手指, 画得不好的手部, 画得不好的脸部, 畸形的, 毁容的, 形态畸形的肢体, 手指融合, 静止不动的画面, 杂乱的背景, 三条腿, 背景人很多, 倒着走"
|
64 |
|
65 |
-
|
66 |
def resize_image(image: Image.Image) -> Image.Image:
|
67 |
-
|
68 |
-
|
69 |
-
|
70 |
-
|
71 |
-
return resize_image_landscape(image)
|
72 |
|
|
|
|
|
|
|
73 |
|
74 |
-
|
75 |
-
|
76 |
-
|
77 |
-
|
78 |
-
|
79 |
-
|
80 |
-
|
81 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
82 |
else:
|
83 |
-
|
84 |
-
|
85 |
-
|
86 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
87 |
|
88 |
def get_duration(
|
89 |
input_image,
|
@@ -147,7 +176,6 @@ def generate_video(
|
|
147 |
gr.Error: If input_image is None (no image uploaded).
|
148 |
|
149 |
Note:
|
150 |
-
- The function automatically resizes the input image to the target dimensions
|
151 |
- Frame count is calculated as duration_seconds * FIXED_FPS (24)
|
152 |
- Output dimensions are adjusted to be multiples of MOD_VALUE (32)
|
153 |
- The function uses GPU acceleration via the @spaces.GPU decorator
|
@@ -185,7 +213,7 @@ with gr.Blocks() as demo:
|
|
185 |
gr.Markdown("run Wan 2.2 in just 4-8 steps, with [Lightning LoRA](https://huggingface.co/Kijai/WanVideo_comfy/tree/main/Wan22-Lightning), fp8 quantization & AoT compilation - compatible with 🧨 diffusers and ZeroGPU⚡️")
|
186 |
with gr.Row():
|
187 |
with gr.Column():
|
188 |
-
input_image_component = gr.Image(type="pil", label="Input Image
|
189 |
prompt_input = gr.Textbox(label="Prompt", value=default_prompt_i2v)
|
190 |
duration_seconds_input = gr.Slider(minimum=MIN_DURATION, maximum=MAX_DURATION, step=0.1, value=3.5, label="Duration (seconds)", info=f"Clamped to model's {MIN_FRAMES_MODEL}-{MAX_FRAMES_MODEL} frames at {FIXED_FPS}fps.")
|
191 |
|
|
|
19 |
|
20 |
MODEL_ID = "Wan-AI/Wan2.2-I2V-A14B-Diffusers"
|
21 |
|
22 |
+
MAX_DIM = 832
|
23 |
+
MIN_DIM = 480
|
24 |
+
SQUARE_DIM = 640
|
25 |
+
MULTIPLE_OF = 16
|
26 |
+
|
27 |
MAX_SEED = np.iinfo(np.int32).max
|
28 |
|
29 |
FIXED_FPS = 16
|
|
|
53 |
torch.cuda.synchronize()
|
54 |
torch.cuda.empty_cache()
|
55 |
|
56 |
+
OPTIMIZE_WIDTH = 832
|
57 |
+
OPTIMIZE_HEIGHT = 624
|
58 |
+
|
59 |
optimize_pipeline_(pipe,
|
60 |
+
image=Image.new('RGB', (OPTIMIZE_WIDTH, OPTIMIZE_HEIGHT)),
|
61 |
prompt='prompt',
|
62 |
+
height=OPTIMIZE_HEIGHT,
|
63 |
+
width=OPTIMIZE_WIDTH,
|
64 |
num_frames=MAX_FRAMES_MODEL,
|
65 |
)
|
66 |
|
|
|
68 |
default_prompt_i2v = "make this image come alive, cinematic motion, smooth animation"
|
69 |
default_negative_prompt = "色调艳丽, 过曝, 静态, 细节模糊不清, 字幕, 风格, 作品, 画作, 画面, 静止, 整体发灰, 最差质量, 低质量, JPEG压缩残留, 丑陋的, 残缺的, 多余的手指, 画得不好的手部, 画得不好的脸部, 畸形的, 毁容的, 形态畸形的肢体, 手指融合, 静止不动的画面, 杂乱的背景, 三条腿, 背景人很多, 倒着走"
|
70 |
|
|
|
71 |
def resize_image(image: Image.Image) -> Image.Image:
|
72 |
+
"""
|
73 |
+
Resizes an image to fit within the model's constraints, preserving aspect ratio as much as possible.
|
74 |
+
"""
|
75 |
+
width, height = image.size
|
|
|
76 |
|
77 |
+
# Handle square case
|
78 |
+
if width == height:
|
79 |
+
return image.resize((SQUARE_DIM, SQUARE_DIM), Image.LANCZOS)
|
80 |
|
81 |
+
aspect_ratio = width / height
|
82 |
+
|
83 |
+
MAX_ASPECT_RATIO = MAX_DIM / MIN_DIM
|
84 |
+
MIN_ASPECT_RATIO = MIN_DIM / MAX_DIM
|
85 |
+
|
86 |
+
image_to_resize = image
|
87 |
+
|
88 |
+
if aspect_ratio > MAX_ASPECT_RATIO:
|
89 |
+
# Very wide image -> crop width to fit 832x480 aspect ratio
|
90 |
+
target_w, target_h = MAX_DIM, MIN_DIM
|
91 |
+
crop_width = int(round(height * MAX_ASPECT_RATIO))
|
92 |
+
left = (width - crop_width) // 2
|
93 |
+
image_to_resize = image.crop((left, 0, left + crop_width, height))
|
94 |
+
elif aspect_ratio < MIN_ASPECT_RATIO:
|
95 |
+
# Very tall image -> crop height to fit 480x832 aspect ratio
|
96 |
+
target_w, target_h = MIN_DIM, MAX_DIM
|
97 |
+
crop_height = int(round(width / MIN_ASPECT_RATIO))
|
98 |
+
top = (height - crop_height) // 2
|
99 |
+
image_to_resize = image.crop((0, top, width, top + crop_height))
|
100 |
else:
|
101 |
+
if width > height: # Landscape
|
102 |
+
target_w = MAX_DIM
|
103 |
+
target_h = int(round(target_w / aspect_ratio))
|
104 |
+
else: # Portrait
|
105 |
+
target_h = MAX_DIM
|
106 |
+
target_w = int(round(target_h * aspect_ratio))
|
107 |
+
|
108 |
+
final_w = round(target_w / MULTIPLE_OF) * MULTIPLE_OF
|
109 |
+
final_h = round(target_h / MULTIPLE_OF) * MULTIPLE_OF
|
110 |
+
|
111 |
+
final_w = max(MIN_DIM, min(MAX_DIM, final_w))
|
112 |
+
final_h = max(MIN_DIM, min(MAX_DIM, final_h))
|
113 |
+
|
114 |
+
return image_to_resize.resize((final_w, final_h), Image.LANCZOS)
|
115 |
+
|
116 |
|
117 |
def get_duration(
|
118 |
input_image,
|
|
|
176 |
gr.Error: If input_image is None (no image uploaded).
|
177 |
|
178 |
Note:
|
|
|
179 |
- Frame count is calculated as duration_seconds * FIXED_FPS (24)
|
180 |
- Output dimensions are adjusted to be multiples of MOD_VALUE (32)
|
181 |
- The function uses GPU acceleration via the @spaces.GPU decorator
|
|
|
213 |
gr.Markdown("run Wan 2.2 in just 4-8 steps, with [Lightning LoRA](https://huggingface.co/Kijai/WanVideo_comfy/tree/main/Wan22-Lightning), fp8 quantization & AoT compilation - compatible with 🧨 diffusers and ZeroGPU⚡️")
|
214 |
with gr.Row():
|
215 |
with gr.Column():
|
216 |
+
input_image_component = gr.Image(type="pil", label="Input Image")
|
217 |
prompt_input = gr.Textbox(label="Prompt", value=default_prompt_i2v)
|
218 |
duration_seconds_input = gr.Slider(minimum=MIN_DURATION, maximum=MAX_DURATION, step=0.1, value=3.5, label="Duration (seconds)", info=f"Clamped to model's {MIN_FRAMES_MODEL}-{MAX_FRAMES_MODEL} frames at {FIXED_FPS}fps.")
|
219 |
|
optimization.py
CHANGED
@@ -14,18 +14,21 @@ from torchao.quantization import Int8WeightOnlyConfig
|
|
14 |
|
15 |
from optimization_utils import capture_component_call
|
16 |
from optimization_utils import aoti_compile
|
17 |
-
from optimization_utils import ZeroGPUCompiledModel
|
18 |
from optimization_utils import drain_module_parameters
|
19 |
|
20 |
|
21 |
P = ParamSpec('P')
|
22 |
|
|
|
23 |
|
24 |
-
|
|
|
25 |
|
26 |
TRANSFORMER_DYNAMIC_SHAPES = {
|
27 |
'hidden_states': {
|
28 |
-
2:
|
|
|
|
|
29 |
},
|
30 |
}
|
31 |
|
@@ -44,6 +47,7 @@ def optimize_pipeline_(pipeline: Callable[P, Any], *args: P.args, **kwargs: P.kw
|
|
44 |
@spaces.GPU(duration=1500)
|
45 |
def compile_transformer():
|
46 |
|
|
|
47 |
pipeline.load_lora_weights(
|
48 |
"Kijai/WanVideo_comfy",
|
49 |
weight_name="Lightx2v/lightx2v_I2V_14B_480p_cfg_step_distill_rank128_bf16.safetensors",
|
@@ -70,61 +74,33 @@ def optimize_pipeline_(pipeline: Callable[P, Any], *args: P.args, **kwargs: P.kw
|
|
70 |
quantize_(pipeline.transformer, Float8DynamicActivationFloat8WeightConfig())
|
71 |
quantize_(pipeline.transformer_2, Float8DynamicActivationFloat8WeightConfig())
|
72 |
|
73 |
-
|
74 |
-
|
75 |
-
if hidden_states.shape[-1] > hidden_states.shape[-2]:
|
76 |
-
hidden_states_landscape = hidden_states
|
77 |
-
hidden_states_portrait = hidden_states_transposed
|
78 |
-
else:
|
79 |
-
hidden_states_landscape = hidden_states_transposed
|
80 |
-
hidden_states_portrait = hidden_states
|
81 |
-
|
82 |
-
exported_landscape_1 = torch.export.export(
|
83 |
mod=pipeline.transformer,
|
84 |
args=call.args,
|
85 |
-
kwargs=call.kwargs
|
86 |
dynamic_shapes=dynamic_shapes,
|
87 |
)
|
88 |
|
89 |
-
|
90 |
mod=pipeline.transformer_2,
|
91 |
args=call.args,
|
92 |
-
kwargs=call.kwargs
|
93 |
dynamic_shapes=dynamic_shapes,
|
94 |
)
|
95 |
|
96 |
-
|
97 |
-
|
|
|
|
|
98 |
|
99 |
-
compiled_landscape_2 = ZeroGPUCompiledModel(compiled_landscape_1.archive_file, compiled_portrait_2.weights)
|
100 |
-
compiled_portrait_1 = ZeroGPUCompiledModel(compiled_portrait_2.archive_file, compiled_landscape_1.weights)
|
101 |
-
|
102 |
-
return (
|
103 |
-
compiled_landscape_1,
|
104 |
-
compiled_landscape_2,
|
105 |
-
compiled_portrait_1,
|
106 |
-
compiled_portrait_2,
|
107 |
-
)
|
108 |
|
109 |
quantize_(pipeline.text_encoder, Int8WeightOnlyConfig())
|
110 |
-
|
111 |
-
|
112 |
-
|
113 |
-
|
114 |
-
if hidden_states.shape[-1] > hidden_states.shape[-2]:
|
115 |
-
return cl1(*args, **kwargs)
|
116 |
-
else:
|
117 |
-
return cp1(*args, **kwargs)
|
118 |
-
|
119 |
-
def combined_transformer_2(*args, **kwargs):
|
120 |
-
hidden_states: torch.Tensor = kwargs['hidden_states']
|
121 |
-
if hidden_states.shape[-1] > hidden_states.shape[-2]:
|
122 |
-
return cl2(*args, **kwargs)
|
123 |
-
else:
|
124 |
-
return cp2(*args, **kwargs)
|
125 |
-
|
126 |
-
pipeline.transformer.forward = combined_transformer_1
|
127 |
drain_module_parameters(pipeline.transformer)
|
128 |
|
129 |
-
pipeline.transformer_2.forward =
|
130 |
drain_module_parameters(pipeline.transformer_2)
|
|
|
14 |
|
15 |
from optimization_utils import capture_component_call
|
16 |
from optimization_utils import aoti_compile
|
|
|
17 |
from optimization_utils import drain_module_parameters
|
18 |
|
19 |
|
20 |
P = ParamSpec('P')
|
21 |
|
22 |
+
LATENT_FRAMES_DIM = torch.export.Dim('num_latent_frames', min=8, max=81)
|
23 |
|
24 |
+
LATENT_PATCHED_HEIGHT_DIM = torch.export.Dim('latent_patched_height', min=30, max=52)
|
25 |
+
LATENT_PATCHED_WIDTH_DIM = torch.export.Dim('latent_patched_width', min=30, max=52)
|
26 |
|
27 |
TRANSFORMER_DYNAMIC_SHAPES = {
|
28 |
'hidden_states': {
|
29 |
+
2: LATENT_FRAMES_DIM,
|
30 |
+
3: 2 * LATENT_PATCHED_HEIGHT_DIM,
|
31 |
+
4: 2 * LATENT_PATCHED_WIDTH_DIM,
|
32 |
},
|
33 |
}
|
34 |
|
|
|
47 |
@spaces.GPU(duration=1500)
|
48 |
def compile_transformer():
|
49 |
|
50 |
+
# This LoRA fusion part remains the same
|
51 |
pipeline.load_lora_weights(
|
52 |
"Kijai/WanVideo_comfy",
|
53 |
weight_name="Lightx2v/lightx2v_I2V_14B_480p_cfg_step_distill_rank128_bf16.safetensors",
|
|
|
74 |
quantize_(pipeline.transformer, Float8DynamicActivationFloat8WeightConfig())
|
75 |
quantize_(pipeline.transformer_2, Float8DynamicActivationFloat8WeightConfig())
|
76 |
|
77 |
+
|
78 |
+
exported_1 = torch.export.export(
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
79 |
mod=pipeline.transformer,
|
80 |
args=call.args,
|
81 |
+
kwargs=call.kwargs,
|
82 |
dynamic_shapes=dynamic_shapes,
|
83 |
)
|
84 |
|
85 |
+
exported_2 = torch.export.export(
|
86 |
mod=pipeline.transformer_2,
|
87 |
args=call.args,
|
88 |
+
kwargs=call.kwargs,
|
89 |
dynamic_shapes=dynamic_shapes,
|
90 |
)
|
91 |
|
92 |
+
compiled_1 = aoti_compile(exported_1, INDUCTOR_CONFIGS)
|
93 |
+
compiled_2 = aoti_compile(exported_2, INDUCTOR_CONFIGS)
|
94 |
+
|
95 |
+
return compiled_1, compiled_2
|
96 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
97 |
|
98 |
quantize_(pipeline.text_encoder, Int8WeightOnlyConfig())
|
99 |
+
|
100 |
+
compiled_transformer_1, compiled_transformer_2 = compile_transformer()
|
101 |
+
|
102 |
+
pipeline.transformer.forward = compiled_transformer_1
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
103 |
drain_module_parameters(pipeline.transformer)
|
104 |
|
105 |
+
pipeline.transformer_2.forward = compiled_transformer_2
|
106 |
drain_module_parameters(pipeline.transformer_2)
|