"""
"""

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')


# Mark only the plain-tensor args dynamic, with Dim.AUTO (dynamic where the
# model allows, static where it specializes). image_rotary_emb is left static
# (via tree_map) — its nested structure drifts across diffusers versions and
# hardcoding it breaks torch.export's tree matching.
_AUTO = torch.export.Dim.AUTO
TRANSFORMER_DYNAMIC_SHAPES = {
    'hidden_states': {1: _AUTO},
    'encoder_hidden_states': {1: _AUTO},
    'encoder_hidden_states_mask': {1: _AUTO},
}


INDUCTOR_CONFIGS = {
    'conv_1x1_as_mm': True,
    'epilogue_fusion': False,
    'coordinate_descent_tuning': True,
    'coordinate_descent_check_all_directions': True,
    'max_autotune': False,
    'triton.cudagraphs': False,
}


def optimize_pipeline_(pipeline: Callable[P, Any], *args: P.args, **kwargs: P.kwargs):

    @spaces.GPU(duration=750, size="xlarge")
    def compile_transformer():

        with spaces.aoti_capture(pipeline.transformer) as call:
            pipeline(*args, **kwargs)

        dynamic_shapes = tree_map(lambda t: None, call.kwargs)
        # Only mark an arg dynamic if it's actually a Tensor in this call — arg
        # names AND presence (e.g. encoder_hidden_states_mask=None) drift across
        # diffusers versions; marking a None arg dynamic errors.
        for _k, _v in TRANSFORMER_DYNAMIC_SHAPES.items():
            if isinstance(call.kwargs.get(_k), torch.Tensor):
                dynamic_shapes[_k] = _v

        # 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)