FLUX.2-dev / optimization.py
multimodalart's picture
Fix AoT (#1)
fc39399 verified
"""
"""
from typing import Any
from typing import Callable
from typing import ParamSpec
import spaces
import torch
from spaces.zero.torch.aoti import ZeroGPUCompiledModel
from spaces.zero.torch.aoti import ZeroGPUWeights
from torch.utils._pytree import tree_map
P = ParamSpec('P')
TRANSFORMER_IMAGE_DIM = torch.export.Dim('image_seq_length', min=4096, max=16384) # min: 0 images, max: 3 (1024x1024) images
TRANSFORMER_DYNAMIC_SHAPES = {
'double': {
'hidden_states': {
1: TRANSFORMER_IMAGE_DIM,
},
'image_rotary_emb': (
{0: TRANSFORMER_IMAGE_DIM + 512},
{0: TRANSFORMER_IMAGE_DIM + 512},
),
},
'single': {
'hidden_states': {
1: TRANSFORMER_IMAGE_DIM + 512,
},
'image_rotary_emb': (
{0: TRANSFORMER_IMAGE_DIM + 512},
{0: TRANSFORMER_IMAGE_DIM + 512},
),
},
}
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):
blocks = {
'double': pipeline.transformer.transformer_blocks,
'single': pipeline.transformer.single_transformer_blocks,
}
@spaces.GPU(duration=1200)
def compile_block(blocks_kind: str):
block = blocks[blocks_kind][0]
with spaces.aoti_capture(block) as call:
pipeline(*args, **kwargs)
dynamic_shapes = tree_map(lambda t: None, call.kwargs)
dynamic_shapes |= TRANSFORMER_DYNAMIC_SHAPES[blocks_kind]
with torch.no_grad():
exported = torch.export.export(
mod=block,
args=call.args,
kwargs=call.kwargs,
dynamic_shapes=dynamic_shapes,
)
return spaces.aoti_compile(exported, INDUCTOR_CONFIGS).archive_file
for blocks_kind in ('double', 'single'):
archive_file = compile_block(blocks_kind)
for block in blocks[blocks_kind]:
weights = ZeroGPUWeights(block.state_dict())
block.forward = ZeroGPUCompiledModel(archive_file, weights)