Spaces:
Running
on
Zero
Running
on
Zero
up
Browse files- optimization.py +45 -11
optimization.py
CHANGED
@@ -11,7 +11,7 @@ import torch
|
|
11 |
from torch.utils._pytree import tree_map_only
|
12 |
from torchao.quantization import quantize_
|
13 |
from torchao.quantization import Float8DynamicActivationFloat8WeightConfig
|
14 |
-
|
15 |
from optimization_utils import capture_component_call
|
16 |
from optimization_utils import aoti_compile
|
17 |
|
@@ -19,10 +19,12 @@ from optimization_utils import aoti_compile
|
|
19 |
P = ParamSpec('P')
|
20 |
|
21 |
|
22 |
-
|
|
|
|
|
23 |
|
24 |
TRANSFORMER_DYNAMIC_SHAPES = {
|
25 |
-
'hidden_states': {
|
26 |
}
|
27 |
|
28 |
INDUCTOR_CONFIGS = {
|
@@ -33,9 +35,18 @@ INDUCTOR_CONFIGS = {
|
|
33 |
'max_autotune': True,
|
34 |
'triton.cudagraphs': True,
|
35 |
}
|
36 |
-
|
|
|
|
|
|
|
37 |
|
38 |
def optimize_pipeline_(pipeline: Callable[P, Any], *args: P.args, **kwargs: P.kwargs):
|
|
|
|
|
|
|
|
|
|
|
|
|
39 |
|
40 |
@spaces.GPU(duration=1500)
|
41 |
def compile_transformer():
|
@@ -49,13 +60,28 @@ def optimize_pipeline_(pipeline: Callable[P, Any], *args: P.args, **kwargs: P.kw
|
|
49 |
quantize_(pipeline.transformer, Float8DynamicActivationFloat8WeightConfig())
|
50 |
|
51 |
hidden_states: torch.Tensor = call.kwargs['hidden_states']
|
52 |
-
|
53 |
-
|
54 |
-
|
55 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
56 |
else:
|
57 |
-
hidden_states_landscape =
|
58 |
-
hidden_states_portrait =
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
59 |
|
60 |
exported_landscape = torch.export.export(
|
61 |
mod=pipeline.transformer,
|
@@ -81,7 +107,15 @@ def optimize_pipeline_(pipeline: Callable[P, Any], *args: P.args, **kwargs: P.kw
|
|
81 |
|
82 |
def combined_transformer(*args, **kwargs):
|
83 |
hidden_states: torch.Tensor = kwargs['hidden_states']
|
84 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
85 |
return compiled_landscape(*args, **kwargs)
|
86 |
else:
|
87 |
return compiled_portrait(*args, **kwargs)
|
|
|
11 |
from torch.utils._pytree import tree_map_only
|
12 |
from torchao.quantization import quantize_
|
13 |
from torchao.quantization import Float8DynamicActivationFloat8WeightConfig
|
14 |
+
from diffusers import LTXConditionPipeline
|
15 |
from optimization_utils import capture_component_call
|
16 |
from optimization_utils import aoti_compile
|
17 |
|
|
|
19 |
P = ParamSpec('P')
|
20 |
|
21 |
|
22 |
+
# Sequence packing in LTX is a bit of a pain.
|
23 |
+
# See: https://github.com/huggingface/diffusers/blob/c052791b5fe29ce8a308bf63dda97aa205b729be/src/diffusers/pipelines/ltx/pipeline_ltx.py#L420
|
24 |
+
TRANSFORMER_NUM_FRAMES_DIM = torch.export.Dim('seq_len', min=4680, max=6000)
|
25 |
|
26 |
TRANSFORMER_DYNAMIC_SHAPES = {
|
27 |
+
'hidden_states': {1: TRANSFORMER_NUM_FRAMES_DIM},
|
28 |
}
|
29 |
|
30 |
INDUCTOR_CONFIGS = {
|
|
|
35 |
'max_autotune': True,
|
36 |
'triton.cudagraphs': True,
|
37 |
}
|
38 |
+
TRANSFORMER_SPATIAL_PATCH_SIZE = 1
|
39 |
+
TRANSFORMER_TEMPORAL_PATCH_SIZE = 1
|
40 |
+
VAE_SPATIAL_COMPRESSION_RATIO = 32
|
41 |
+
VAE_TEMPORAL_COMPRESSION_RATIO = 8
|
42 |
|
43 |
def optimize_pipeline_(pipeline: Callable[P, Any], *args: P.args, **kwargs: P.kwargs):
|
44 |
+
num_frames = kwargs.get("num_frames")
|
45 |
+
height = kwargs.get("height")
|
46 |
+
width = kwargs.get("width")
|
47 |
+
latent_num_frames = (num_frames - 1) // VAE_TEMPORAL_COMPRESSION_RATIO + 1
|
48 |
+
latent_height = height // VAE_SPATIAL_COMPRESSION_RATIO
|
49 |
+
latent_width = width //VAE_SPATIAL_COMPRESSION_RATIO
|
50 |
|
51 |
@spaces.GPU(duration=1500)
|
52 |
def compile_transformer():
|
|
|
60 |
quantize_(pipeline.transformer, Float8DynamicActivationFloat8WeightConfig())
|
61 |
|
62 |
hidden_states: torch.Tensor = call.kwargs['hidden_states']
|
63 |
+
unpacked_hidden_states = LTXConditionPipeline._unpack_latents(
|
64 |
+
hidden_states,
|
65 |
+
latent_num_frames,
|
66 |
+
latent_height,
|
67 |
+
latent_width,
|
68 |
+
TRANSFORMER_SPATIAL_PATCH_SIZE,
|
69 |
+
TRANSFORMER_TEMPORAL_PATCH_SIZE,
|
70 |
+
)
|
71 |
+
unpacked_hidden_states_transposed = hidden_states.transpose(-1, -2).contiguous()
|
72 |
+
if unpacked_hidden_states.shape[-1] > hidden_states.shape[-2]:
|
73 |
+
hidden_states_landscape = unpacked_hidden_states
|
74 |
+
hidden_states_portrait = unpacked_hidden_states_transposed
|
75 |
else:
|
76 |
+
hidden_states_landscape = unpacked_hidden_states_transposed
|
77 |
+
hidden_states_portrait = unpacked_hidden_states
|
78 |
+
|
79 |
+
hidden_states_landscape = LTXConditionPipeline._pack_latents(
|
80 |
+
hidden_states_landscape, TRANSFORMER_SPATIAL_PATCH_SIZE, TRANSFORMER_TEMPORAL_PATCH_SIZE
|
81 |
+
)
|
82 |
+
hidden_states_portrait = LTXConditionPipeline._pack_latents(
|
83 |
+
hidden_states_portrait, TRANSFORMER_SPATIAL_PATCH_SIZE, TRANSFORMER_TEMPORAL_PATCH_SIZE
|
84 |
+
)
|
85 |
|
86 |
exported_landscape = torch.export.export(
|
87 |
mod=pipeline.transformer,
|
|
|
107 |
|
108 |
def combined_transformer(*args, **kwargs):
|
109 |
hidden_states: torch.Tensor = kwargs['hidden_states']
|
110 |
+
unpacked_hidden_states = LTXConditionPipeline._unpack_latents(
|
111 |
+
hidden_states,
|
112 |
+
latent_num_frames,
|
113 |
+
latent_height,
|
114 |
+
latent_width,
|
115 |
+
TRANSFORMER_SPATIAL_PATCH_SIZE,
|
116 |
+
TRANSFORMER_TEMPORAL_PATCH_SIZE,
|
117 |
+
)
|
118 |
+
if unpacked_hidden_states.shape[-1] > unpacked_hidden_states.shape[-2]:
|
119 |
return compiled_landscape(*args, **kwargs)
|
120 |
else:
|
121 |
return compiled_portrait(*args, **kwargs)
|