Spaces:
Runtime error
Runtime error
File size: 3,324 Bytes
22ac777 425ba26 22ac777 356294b 22ac777 356294b 22ac777 356294b 22ac777 356294b 1017ac0 e9125ed 1017ac0 356294b 22ac777 |
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 |
"""
"""
from typing import Any
from typing import Callable
from typing import ParamSpec
import spaces
import torch
from torch.utils._pytree import tree_map_only
from torchao.quantization import quantize_
from torchao.quantization import Float8DynamicActivationFloat8WeightConfig
from optimization_utils import capture_component_call
from optimization_utils import aoti_compile
P = ParamSpec('P')
TRANSFORMER_NUM_FRAMES_DIM = torch.export.Dim('num_frames', min=3, max=21)
TRANSFORMER_DYNAMIC_SHAPES = {
'hidden_states': {
2: TRANSFORMER_NUM_FRAMES_DIM,
},
}
INDUCTOR_CONFIGS = {
'conv_1x1_as_mm': True,
'epilogue_fusion': False,
'coordinate_descent_tuning': True,
'coordinate_descent_check_all_directions': True,
'max_autotune': True,
'triton.cudagraphs': True,
}
def optimize_pipeline_(pipeline: Callable[P, Any], *args: P.args, **kwargs: P.kwargs):
@spaces.GPU(duration=1500)
def compile_transformer():
with capture_component_call(pipeline, 'transformer') as call:
pipeline(*args, **kwargs)
dynamic_shapes = tree_map_only((torch.Tensor, bool), lambda t: None, call.kwargs)
dynamic_shapes |= TRANSFORMER_DYNAMIC_SHAPES
quantize_(pipeline.transformer, Float8DynamicActivationFloat8WeightConfig())
hidden_states: torch.Tensor = call.kwargs['hidden_states']
hidden_states_transposed = hidden_states.transpose(-1, -2).contiguous()
if hidden_states.shape[-1] > hidden_states.shape[-2]:
hidden_states_landscape = hidden_states
hidden_states_portrait = hidden_states_transposed
else:
hidden_states_landscape = hidden_states_transposed
hidden_states_portrait = hidden_states
exported_landscape = torch.export.export(
mod=pipeline.transformer,
args=call.args,
kwargs=call.kwargs | {'hidden_states': hidden_states_landscape},
dynamic_shapes=dynamic_shapes,
)
exported_portrait = torch.export.export(
mod=pipeline.transformer,
args=call.args,
kwargs=call.kwargs | {'hidden_states': hidden_states_portrait},
dynamic_shapes=dynamic_shapes,
)
compiled_landscape = aoti_compile(exported_landscape, INDUCTOR_CONFIGS)
compiled_portrait = aoti_compile(exported_portrait, INDUCTOR_CONFIGS)
compiled_portrait.weights = compiled_landscape.weights # Avoid weights duplication when serializing back to main process
return compiled_landscape, compiled_portrait
compiled_landscape, compiled_portrait = compile_transformer()
def combined_transformer(*args, **kwargs):
hidden_states: torch.Tensor = kwargs['hidden_states']
if hidden_states.shape[-1] > hidden_states.shape[-2]:
return compiled_landscape(*args, **kwargs)
else:
return compiled_portrait(*args, **kwargs)
transformer_config = pipeline.transformer.config
transformer_dtype = pipeline.transformer.dtype
pipeline.transformer = combined_transformer
pipeline.transformer.config = transformer_config # pyright: ignore[reportAttributeAccessIssue]
pipeline.transformer.dtype = transformer_dtype # pyright: ignore[reportAttributeAccessIssue]
|