Spaces:
Running
on
Zero
Running
on
Zero
Create optimization.py
Browse files- optimization.py +128 -0
optimization.py
ADDED
@@ -0,0 +1,128 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
|
2 |
+
from typing import Any
|
3 |
+
from typing import Callable
|
4 |
+
from typing import ParamSpec
|
5 |
+
|
6 |
+
import spaces
|
7 |
+
import torch
|
8 |
+
from torch.utils._pytree import tree_map_only
|
9 |
+
from torchao.quantization import quantize_
|
10 |
+
from torchao.quantization import Float8DynamicActivationFloat8WeightConfig
|
11 |
+
from torchao.quantization import Int8WeightOnlyConfig
|
12 |
+
|
13 |
+
from optimization_utils import capture_component_call
|
14 |
+
from optimization_utils import aoti_compile
|
15 |
+
from optimization_utils import ZeroGPUCompiledModel
|
16 |
+
from optimization_utils import drain_module_parameters
|
17 |
+
|
18 |
+
|
19 |
+
P = ParamSpec('P')
|
20 |
+
|
21 |
+
|
22 |
+
TRANSFORMER_NUM_FRAMES_DIM = torch.export.Dim('num_frames', min=3, max=21)
|
23 |
+
|
24 |
+
TRANSFORMER_DYNAMIC_SHAPES = {
|
25 |
+
'hidden_states': {
|
26 |
+
2: TRANSFORMER_NUM_FRAMES_DIM,
|
27 |
+
},
|
28 |
+
}
|
29 |
+
|
30 |
+
INDUCTOR_CONFIGS = {
|
31 |
+
'conv_1x1_as_mm': True,
|
32 |
+
'epilogue_fusion': False,
|
33 |
+
'coordinate_descent_tuning': True,
|
34 |
+
'coordinate_descent_check_all_directions': True,
|
35 |
+
'max_autotune': True,
|
36 |
+
'triton.cudagraphs': True,
|
37 |
+
}
|
38 |
+
|
39 |
+
|
40 |
+
def optimize_pipeline_(pipeline: Callable[P, Any], *args: P.args, **kwargs: P.kwargs):
|
41 |
+
|
42 |
+
@spaces.GPU(duration=1500)
|
43 |
+
def compile_transformer():
|
44 |
+
|
45 |
+
pipeline.load_lora_weights(
|
46 |
+
"Kijai/WanVideo_comfy",
|
47 |
+
weight_name="Lightx2v/lightx2v_I2V_14B_480p_cfg_step_distill_rank128_bf16.safetensors",
|
48 |
+
adapter_name="lightx2v"
|
49 |
+
)
|
50 |
+
kwargs_lora = {}
|
51 |
+
kwargs_lora["load_into_transformer_2"] = True
|
52 |
+
pipeline.load_lora_weights(
|
53 |
+
"Kijai/WanVideo_comfy",
|
54 |
+
weight_name="Lightx2v/lightx2v_I2V_14B_480p_cfg_step_distill_rank128_bf16.safetensors",
|
55 |
+
adapter_name="lightx2v_2", **kwargs_lora
|
56 |
+
)
|
57 |
+
pipeline.set_adapters(["lightx2v", "lightx2v_2"], adapter_weights=[1., 1.])
|
58 |
+
pipeline.fuse_lora(adapter_names=["lightx2v"], lora_scale=3., components=["transformer"])
|
59 |
+
pipeline.fuse_lora(adapter_names=["lightx2v_2"], lora_scale=1., components=["transformer_2"])
|
60 |
+
pipeline.unload_lora_weights()
|
61 |
+
|
62 |
+
with capture_component_call(pipeline, 'transformer') as call:
|
63 |
+
pipeline(*args, **kwargs)
|
64 |
+
|
65 |
+
dynamic_shapes = tree_map_only((torch.Tensor, bool), lambda t: None, call.kwargs)
|
66 |
+
dynamic_shapes |= TRANSFORMER_DYNAMIC_SHAPES
|
67 |
+
|
68 |
+
quantize_(pipeline.transformer, Float8DynamicActivationFloat8WeightConfig())
|
69 |
+
quantize_(pipeline.transformer_2, Float8DynamicActivationFloat8WeightConfig())
|
70 |
+
|
71 |
+
hidden_states: torch.Tensor = call.kwargs['hidden_states']
|
72 |
+
hidden_states_transposed = hidden_states.transpose(-1, -2).contiguous()
|
73 |
+
if hidden_states.shape[-1] > hidden_states.shape[-2]:
|
74 |
+
hidden_states_landscape = hidden_states
|
75 |
+
hidden_states_portrait = hidden_states_transposed
|
76 |
+
else:
|
77 |
+
hidden_states_landscape = hidden_states_transposed
|
78 |
+
hidden_states_portrait = hidden_states
|
79 |
+
|
80 |
+
exported_landscape_1 = torch.export.export(
|
81 |
+
mod=pipeline.transformer,
|
82 |
+
args=call.args,
|
83 |
+
kwargs=call.kwargs | {'hidden_states': hidden_states_landscape},
|
84 |
+
dynamic_shapes=dynamic_shapes,
|
85 |
+
)
|
86 |
+
|
87 |
+
exported_portrait_2 = torch.export.export(
|
88 |
+
mod=pipeline.transformer_2,
|
89 |
+
args=call.args,
|
90 |
+
kwargs=call.kwargs | {'hidden_states': hidden_states_portrait},
|
91 |
+
dynamic_shapes=dynamic_shapes,
|
92 |
+
)
|
93 |
+
|
94 |
+
compiled_landscape_1 = aoti_compile(exported_landscape_1, INDUCTOR_CONFIGS)
|
95 |
+
compiled_portrait_2 = aoti_compile(exported_portrait_2, INDUCTOR_CONFIGS)
|
96 |
+
|
97 |
+
compiled_landscape_2 = ZeroGPUCompiledModel(compiled_landscape_1.archive_file, compiled_portrait_2.weights)
|
98 |
+
compiled_portrait_1 = ZeroGPUCompiledModel(compiled_portrait_2.archive_file, compiled_landscape_1.weights)
|
99 |
+
|
100 |
+
return (
|
101 |
+
compiled_landscape_1,
|
102 |
+
compiled_landscape_2,
|
103 |
+
compiled_portrait_1,
|
104 |
+
compiled_portrait_2,
|
105 |
+
)
|
106 |
+
|
107 |
+
quantize_(pipeline.text_encoder, Int8WeightOnlyConfig())
|
108 |
+
cl1, cl2, cp1, cp2 = compile_transformer()
|
109 |
+
|
110 |
+
def combined_transformer_1(*args, **kwargs):
|
111 |
+
hidden_states: torch.Tensor = kwargs['hidden_states']
|
112 |
+
if hidden_states.shape[-1] > hidden_states.shape[-2]:
|
113 |
+
return cl1(*args, **kwargs)
|
114 |
+
else:
|
115 |
+
return cp1(*args, **kwargs)
|
116 |
+
|
117 |
+
def combined_transformer_2(*args, **kwargs):
|
118 |
+
hidden_states: torch.Tensor = kwargs['hidden_states']
|
119 |
+
if hidden_states.shape[-1] > hidden_states.shape[-2]:
|
120 |
+
return cl2(*args, **kwargs)
|
121 |
+
else:
|
122 |
+
return cp2(*args, **kwargs)
|
123 |
+
|
124 |
+
pipeline.transformer.forward = combined_transformer_1
|
125 |
+
drain_module_parameters(pipeline.transformer)
|
126 |
+
|
127 |
+
pipeline.transformer_2.forward = combined_transformer_2
|
128 |
+
drain_module_parameters(pipeline.transformer_2)
|