Spaces:
Running
on
Zero
Running
on
Zero
File size: 1,688 Bytes
96a0788 777d439 f2e1d62 96a0788 7601fad 96a0788 7601fad 580dbc5 7601fad 580dbc5 96a0788 59021ac 96a0788 d7dfea3 96a0788 f2e1d62 d90daf8 96a0788 7601fad 96a0788 59021ac 96a0788 59021ac |
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 |
"""
"""
from typing import Any
from typing import Callable
from typing import ParamSpec
from torchao.quantization import quantize_
from torchao.quantization import Float8DynamicActivationFloat8WeightConfig
import spaces
import torch
from torch.utils._pytree import tree_map
P = ParamSpec('P')
TRANSFORMER_IMAGE_SEQ_LENGTH_DIM = torch.export.Dim.AUTO(min=3584, max=9727)
TRANSFORMER_IMAGE_HEIGHT_DIM = torch.export.Dim.DYNAMIC
TRANSFORMER_IMAGE_WIDTH_DIM = torch.export.Dim.DYNAMIC
TRANSFORMER_DYNAMIC_SHAPES = {
'hidden_states': {1: TRANSFORMER_IMAGE_SEQ_LENGTH_DIM},
'img_shapes': [(None, TRANSFORMER_IMAGE_HEIGHT_DIM, TRANSFORMER_IMAGE_WIDTH_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 spaces.aoti_capture(pipeline.transformer) as call:
pipeline(*args, **kwargs)
dynamic_shapes = tree_map(lambda t: None, call.kwargs)
dynamic_shapes |= TRANSFORMER_DYNAMIC_SHAPES
quantize_(pipeline.transformer, Float8DynamicActivationFloat8WeightConfig())
exported = torch.export.export(
mod=pipeline.transformer,
args=call.args,
kwargs=call.kwargs,
dynamic_shapes=dynamic_shapes,
)
return spaces.aoti_compile(exported, INDUCTOR_CONFIGS)
spaces.aoti_apply(compile_transformer(), pipeline.transformer)
|