Update optimization.py
Browse files- optimization.py +53 -41
optimization.py
CHANGED
@@ -1,14 +1,32 @@
|
|
1 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
2 |
import spaces
|
3 |
import torch
|
4 |
from torch.utils._pytree import tree_map_only
|
5 |
-
from torchao.quantization import quantize_
|
6 |
-
from
|
|
|
|
|
|
|
|
|
|
|
|
|
7 |
|
8 |
P = ParamSpec('P')
|
9 |
|
|
|
10 |
TRANSFORMER_NUM_FRAMES_DIM = torch.export.Dim('num_frames', min=3, max=21)
|
11 |
-
|
|
|
|
|
|
|
|
|
|
|
12 |
|
13 |
INDUCTOR_CONFIGS = {
|
14 |
'conv_1x1_as_mm': True,
|
@@ -19,48 +37,43 @@ INDUCTOR_CONFIGS = {
|
|
19 |
'triton.cudagraphs': True,
|
20 |
}
|
21 |
|
22 |
-
def optimize_pipeline_(pipeline: Callable[P, Any], *args: P.args, **kwargs: P.kwargs):
|
23 |
-
print("[optimize_pipeline_] Starting pipeline optimization")
|
24 |
|
25 |
-
|
26 |
-
print("[optimize_pipeline_] Text encoder quantized")
|
27 |
|
28 |
@spaces.GPU(duration=1500)
|
29 |
def compile_transformer():
|
30 |
-
|
31 |
pipeline.load_lora_weights(
|
32 |
-
"
|
33 |
-
weight_name="
|
34 |
adapter_name="lightning"
|
35 |
)
|
|
|
|
|
36 |
pipeline.load_lora_weights(
|
37 |
-
"
|
38 |
-
|
39 |
-
|
40 |
-
|
41 |
)
|
42 |
-
pipeline.set_adapters(["lightning", "lightning_2"], adapter_weights=[1
|
43 |
|
44 |
-
print("[compile_transformer] Fusing LoRA weights")
|
45 |
-
pipeline.fuse_lora(adapter_names=["lightning"], lora_scale=3.0, components=["transformer"])
|
46 |
-
pipeline.fuse_lora(adapter_names=["lightning_2"], lora_scale=1.0, components=["transformer_2"])
|
47 |
-
pipeline.unload_lora_weights()
|
48 |
-
|
49 |
-
print("[compile_transformer] Running dummy forward pass to capture component call")
|
50 |
-
with torch.inference_mode():
|
51 |
-
with capture_component_call(pipeline, 'transformer') as call:
|
52 |
-
pipeline(*args, **kwargs)
|
53 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
54 |
dynamic_shapes = tree_map_only((torch.Tensor, bool), lambda t: None, call.kwargs)
|
55 |
dynamic_shapes |= TRANSFORMER_DYNAMIC_SHAPES
|
56 |
|
57 |
-
print("[compile_transformer] Quantizing transformers with Float8DynamicActivationFloat8WeightConfig")
|
58 |
quantize_(pipeline.transformer, Float8DynamicActivationFloat8WeightConfig())
|
59 |
quantize_(pipeline.transformer_2, Float8DynamicActivationFloat8WeightConfig())
|
60 |
-
|
61 |
hidden_states: torch.Tensor = call.kwargs['hidden_states']
|
62 |
hidden_states_transposed = hidden_states.transpose(-1, -2).contiguous()
|
63 |
-
|
64 |
if hidden_states.shape[-1] > hidden_states.shape[-2]:
|
65 |
hidden_states_landscape = hidden_states
|
66 |
hidden_states_portrait = hidden_states_transposed
|
@@ -68,34 +81,34 @@ def optimize_pipeline_(pipeline: Callable[P, Any], *args: P.args, **kwargs: P.kw
|
|
68 |
hidden_states_landscape = hidden_states_transposed
|
69 |
hidden_states_portrait = hidden_states
|
70 |
|
71 |
-
print("[compile_transformer] Exporting transformer landscape model")
|
72 |
exported_landscape_1 = torch.export.export(
|
73 |
mod=pipeline.transformer,
|
74 |
args=call.args,
|
75 |
-
kwargs=
|
76 |
dynamic_shapes=dynamic_shapes,
|
77 |
)
|
78 |
-
|
79 |
-
|
80 |
-
print("[compile_transformer] Exporting transformer portrait model")
|
81 |
exported_portrait_2 = torch.export.export(
|
82 |
mod=pipeline.transformer_2,
|
83 |
args=call.args,
|
84 |
-
kwargs=
|
85 |
dynamic_shapes=dynamic_shapes,
|
86 |
)
|
87 |
-
torch.cuda.synchronize()
|
88 |
|
89 |
-
print("[compile_transformer] Compiling models with AoT compilation")
|
90 |
compiled_landscape_1 = aoti_compile(exported_landscape_1, INDUCTOR_CONFIGS)
|
91 |
compiled_portrait_2 = aoti_compile(exported_portrait_2, INDUCTOR_CONFIGS)
|
92 |
|
93 |
compiled_landscape_2 = ZeroGPUCompiledModel(compiled_landscape_1.archive_file, compiled_portrait_2.weights)
|
94 |
compiled_portrait_1 = ZeroGPUCompiledModel(compiled_portrait_2.archive_file, compiled_landscape_1.weights)
|
95 |
|
96 |
-
|
97 |
-
|
|
|
|
|
|
|
|
|
98 |
|
|
|
99 |
cl1, cl2, cp1, cp2 = compile_transformer()
|
100 |
|
101 |
def combined_transformer_1(*args, **kwargs):
|
@@ -114,7 +127,6 @@ def optimize_pipeline_(pipeline: Callable[P, Any], *args: P.args, **kwargs: P.kw
|
|
114 |
|
115 |
pipeline.transformer.forward = combined_transformer_1
|
116 |
drain_module_parameters(pipeline.transformer)
|
117 |
-
pipeline.transformer_2.forward = combined_transformer_2
|
118 |
-
drain_module_parameters(pipeline.transformer_2)
|
119 |
|
120 |
-
|
|
|
|
1 |
+
"""
|
2 |
+
"""
|
3 |
+
|
4 |
+
from typing import Any
|
5 |
+
from typing import Callable
|
6 |
+
from typing import ParamSpec
|
7 |
+
|
8 |
import spaces
|
9 |
import torch
|
10 |
from torch.utils._pytree import tree_map_only
|
11 |
+
from torchao.quantization import quantize_
|
12 |
+
from torchao.quantization import Float8DynamicActivationFloat8WeightConfig
|
13 |
+
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 |
P = ParamSpec('P')
|
21 |
|
22 |
+
|
23 |
TRANSFORMER_NUM_FRAMES_DIM = torch.export.Dim('num_frames', min=3, max=21)
|
24 |
+
|
25 |
+
TRANSFORMER_DYNAMIC_SHAPES = {
|
26 |
+
'hidden_states': {
|
27 |
+
2: TRANSFORMER_NUM_FRAMES_DIM,
|
28 |
+
},
|
29 |
+
}
|
30 |
|
31 |
INDUCTOR_CONFIGS = {
|
32 |
'conv_1x1_as_mm': True,
|
|
|
37 |
'triton.cudagraphs': True,
|
38 |
}
|
39 |
|
|
|
|
|
40 |
|
41 |
+
def optimize_pipeline_(pipeline: Callable[P, Any], *args: P.args, **kwargs: P.kwargs):
|
|
|
42 |
|
43 |
@spaces.GPU(duration=1500)
|
44 |
def compile_transformer():
|
45 |
+
|
46 |
pipeline.load_lora_weights(
|
47 |
+
"Kijai/WanVideo_comfy",
|
48 |
+
weight_name="Lightx2v/lightx2v_T2V_14B_cfg_step_distill_v2_lora_rank128_bf16.safetensors",
|
49 |
adapter_name="lightning"
|
50 |
)
|
51 |
+
kwargs_lora = {}
|
52 |
+
kwargs_lora["load_into_transformer_2"] = True
|
53 |
pipeline.load_lora_weights(
|
54 |
+
"Kijai/WanVideo_comfy",
|
55 |
+
weight_name="Lightx2v/lightx2v_T2V_14B_cfg_step_distill_v2_lora_rank128_bf16.safetensors",
|
56 |
+
#weight_name="Wan22-Lightning/Wan2.2-Lightning_T2V-A14B-4steps-lora_LOW_fp16.safetensors",
|
57 |
+
adapter_name="lightning_2", **kwargs_lora
|
58 |
)
|
59 |
+
pipeline.set_adapters(["lightning", "lightning_2"], adapter_weights=[1., 1.])
|
60 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
61 |
|
62 |
+
pipeline.fuse_lora(adapter_names=["lightning"], lora_scale=3., components=["transformer"])
|
63 |
+
pipeline.fuse_lora(adapter_names=["lightning_2"], lora_scale=1., components=["transformer_2"])
|
64 |
+
pipeline.unload_lora_weights()
|
65 |
+
|
66 |
+
with capture_component_call(pipeline, 'transformer') as call:
|
67 |
+
pipeline(*args, **kwargs)
|
68 |
+
|
69 |
dynamic_shapes = tree_map_only((torch.Tensor, bool), lambda t: None, call.kwargs)
|
70 |
dynamic_shapes |= TRANSFORMER_DYNAMIC_SHAPES
|
71 |
|
|
|
72 |
quantize_(pipeline.transformer, Float8DynamicActivationFloat8WeightConfig())
|
73 |
quantize_(pipeline.transformer_2, Float8DynamicActivationFloat8WeightConfig())
|
74 |
+
|
75 |
hidden_states: torch.Tensor = call.kwargs['hidden_states']
|
76 |
hidden_states_transposed = hidden_states.transpose(-1, -2).contiguous()
|
|
|
77 |
if hidden_states.shape[-1] > hidden_states.shape[-2]:
|
78 |
hidden_states_landscape = hidden_states
|
79 |
hidden_states_portrait = hidden_states_transposed
|
|
|
81 |
hidden_states_landscape = hidden_states_transposed
|
82 |
hidden_states_portrait = hidden_states
|
83 |
|
|
|
84 |
exported_landscape_1 = torch.export.export(
|
85 |
mod=pipeline.transformer,
|
86 |
args=call.args,
|
87 |
+
kwargs=call.kwargs | {'hidden_states': hidden_states_landscape},
|
88 |
dynamic_shapes=dynamic_shapes,
|
89 |
)
|
90 |
+
|
|
|
|
|
91 |
exported_portrait_2 = torch.export.export(
|
92 |
mod=pipeline.transformer_2,
|
93 |
args=call.args,
|
94 |
+
kwargs=call.kwargs | {'hidden_states': hidden_states_portrait},
|
95 |
dynamic_shapes=dynamic_shapes,
|
96 |
)
|
|
|
97 |
|
|
|
98 |
compiled_landscape_1 = aoti_compile(exported_landscape_1, INDUCTOR_CONFIGS)
|
99 |
compiled_portrait_2 = aoti_compile(exported_portrait_2, INDUCTOR_CONFIGS)
|
100 |
|
101 |
compiled_landscape_2 = ZeroGPUCompiledModel(compiled_landscape_1.archive_file, compiled_portrait_2.weights)
|
102 |
compiled_portrait_1 = ZeroGPUCompiledModel(compiled_portrait_2.archive_file, compiled_landscape_1.weights)
|
103 |
|
104 |
+
return (
|
105 |
+
compiled_landscape_1,
|
106 |
+
compiled_landscape_2,
|
107 |
+
compiled_portrait_1,
|
108 |
+
compiled_portrait_2,
|
109 |
+
)
|
110 |
|
111 |
+
quantize_(pipeline.text_encoder, Int8WeightOnlyConfig())
|
112 |
cl1, cl2, cp1, cp2 = compile_transformer()
|
113 |
|
114 |
def combined_transformer_1(*args, **kwargs):
|
|
|
127 |
|
128 |
pipeline.transformer.forward = combined_transformer_1
|
129 |
drain_module_parameters(pipeline.transformer)
|
|
|
|
|
130 |
|
131 |
+
pipeline.transformer_2.forward = combined_transformer_2
|
132 |
+
drain_module_parameters(pipeline.transformer_2)
|