Spaces:
Configuration error
Configuration error
Timing
Browse files- optimization.py +20 -4
optimization.py
CHANGED
|
@@ -1,6 +1,7 @@
|
|
| 1 |
"""
|
| 2 |
"""
|
| 3 |
|
|
|
|
| 4 |
from typing import Any
|
| 5 |
from typing import Callable
|
| 6 |
from typing import ParamSpec
|
|
@@ -38,17 +39,25 @@ INDUCTOR_CONFIGS = {
|
|
| 38 |
|
| 39 |
def optimize_pipeline_(pipeline: Callable[P, Any], *args: P.args, **kwargs: P.kwargs):
|
| 40 |
|
|
|
|
|
|
|
| 41 |
@spaces.GPU(duration=1500)
|
| 42 |
def compile_transformer():
|
| 43 |
|
|
|
|
|
|
|
| 44 |
with capture_component_call(pipeline, 'transformer') as call:
|
| 45 |
pipeline(*args, **kwargs)
|
| 46 |
|
|
|
|
|
|
|
| 47 |
dynamic_shapes = tree_map_only((torch.Tensor, bool), lambda t: None, call.kwargs)
|
| 48 |
dynamic_shapes |= TRANSFORMER_DYNAMIC_SHAPES
|
| 49 |
|
| 50 |
quantize_(pipeline.transformer, Float8DynamicActivationFloat8WeightConfig())
|
| 51 |
|
|
|
|
|
|
|
| 52 |
hidden_states: torch.Tensor = call.kwargs['hidden_states']
|
| 53 |
hidden_states_transposed = hidden_states.transpose(-1, -2).contiguous()
|
| 54 |
if hidden_states.shape[-1] > hidden_states.shape[-2]:
|
|
@@ -65,6 +74,8 @@ def optimize_pipeline_(pipeline: Callable[P, Any], *args: P.args, **kwargs: P.kw
|
|
| 65 |
dynamic_shapes=dynamic_shapes,
|
| 66 |
)
|
| 67 |
|
|
|
|
|
|
|
| 68 |
exported_portrait = torch.export.export(
|
| 69 |
mod=pipeline.transformer,
|
| 70 |
args=call.args,
|
|
@@ -72,10 +83,15 @@ def optimize_pipeline_(pipeline: Callable[P, Any], *args: P.args, **kwargs: P.kw
|
|
| 72 |
dynamic_shapes=dynamic_shapes,
|
| 73 |
)
|
| 74 |
|
| 75 |
-
|
| 76 |
-
|
| 77 |
-
|
| 78 |
-
)
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 79 |
|
| 80 |
compiled_landscape, compiled_portrait = compile_transformer()
|
| 81 |
|
|
|
|
| 1 |
"""
|
| 2 |
"""
|
| 3 |
|
| 4 |
+
from datetime import datetime
|
| 5 |
from typing import Any
|
| 6 |
from typing import Callable
|
| 7 |
from typing import ParamSpec
|
|
|
|
| 39 |
|
| 40 |
def optimize_pipeline_(pipeline: Callable[P, Any], *args: P.args, **kwargs: P.kwargs):
|
| 41 |
|
| 42 |
+
t0 = datetime.now()
|
| 43 |
+
|
| 44 |
@spaces.GPU(duration=1500)
|
| 45 |
def compile_transformer():
|
| 46 |
|
| 47 |
+
print('compile_transformer', -(t0 - (t0 := datetime.now())))
|
| 48 |
+
|
| 49 |
with capture_component_call(pipeline, 'transformer') as call:
|
| 50 |
pipeline(*args, **kwargs)
|
| 51 |
|
| 52 |
+
print('capture_component_call', -(t0 - (t0 := datetime.now())))
|
| 53 |
+
|
| 54 |
dynamic_shapes = tree_map_only((torch.Tensor, bool), lambda t: None, call.kwargs)
|
| 55 |
dynamic_shapes |= TRANSFORMER_DYNAMIC_SHAPES
|
| 56 |
|
| 57 |
quantize_(pipeline.transformer, Float8DynamicActivationFloat8WeightConfig())
|
| 58 |
|
| 59 |
+
print('quantize_', -(t0 - (t0 := datetime.now())))
|
| 60 |
+
|
| 61 |
hidden_states: torch.Tensor = call.kwargs['hidden_states']
|
| 62 |
hidden_states_transposed = hidden_states.transpose(-1, -2).contiguous()
|
| 63 |
if hidden_states.shape[-1] > hidden_states.shape[-2]:
|
|
|
|
| 74 |
dynamic_shapes=dynamic_shapes,
|
| 75 |
)
|
| 76 |
|
| 77 |
+
print('exported_landscape', -(t0 - (t0 := datetime.now())))
|
| 78 |
+
|
| 79 |
exported_portrait = torch.export.export(
|
| 80 |
mod=pipeline.transformer,
|
| 81 |
args=call.args,
|
|
|
|
| 83 |
dynamic_shapes=dynamic_shapes,
|
| 84 |
)
|
| 85 |
|
| 86 |
+
print('exported_portrait', -(t0 - (t0 := datetime.now())))
|
| 87 |
+
|
| 88 |
+
compiled_landscape = aoti_compile(exported_landscape, INDUCTOR_CONFIGS)
|
| 89 |
+
print('compiled_landscape', -(t0 - (t0 := datetime.now())))
|
| 90 |
+
|
| 91 |
+
compiled_portrait = aoti_compile(exported_portrait, INDUCTOR_CONFIGS)
|
| 92 |
+
print('compiled_portrait', -(t0 - (t0 := datetime.now())))
|
| 93 |
+
|
| 94 |
+
return compiled_landscape, compiled_portrait
|
| 95 |
|
| 96 |
compiled_landscape, compiled_portrait = compile_transformer()
|
| 97 |
|