optimization.py

import spaces
from typing import Any
from typing import Callable
from typing import ParamSpec
import torch
from torch.utils._pytree import tree_map

P = ParamSpec("P")

TRANSFORMER_HIDDEN_DIM = torch.export.Dim("hidden", min=4096, max=8212)

# Specific to Flux. More about this is available in
# https://huggingface.co/blog/zerogpu-aoti
TRANSFORMER_DYNAMIC_SHAPES = {
    "hidden_states": {1: TRANSFORMER_HIDDEN_DIM},
    "img_ids": {0: TRANSFORMER_HIDDEN_DIM},
}

INDUCTOR_CONFIGS = {
    "conv_1x1_as_mm": True,
    "epilogue_fusion": False,
    "coordinate_descent_tuning": True,
    "coordinate_descent_check_all_directions": True,
    # "max_autotune": True, # not very helpful.
    "triton.cudagraphs": True,
}


def compile_transformer(pipeline: Callable[P, Any], *args: P.args, **kwargs: P.kwargs):
    def f():
        with spaces.aoti_capture(pipeline.transformer) as call:
            pipeline(*args, **kwargs)

        print("Inputs captured.")
        dynamic_shapes = tree_map(lambda v: None, call.kwargs)
        dynamic_shapes |= TRANSFORMER_DYNAMIC_SHAPES

        exported = torch.export.export(
            mod=pipeline.transformer, args=call.args, kwargs=call.kwargs, dynamic_shapes=dynamic_shapes
        )
        print("Export done.")
        return spaces.aoti_compile(exported, INDUCTOR_CONFIGS)

    print(f"{pipeline.transformer.device=}")
    compiled_transformer = f()
    print("Compilation done.")
    return compiled_transformer