optimization_utils.py

"""
Taken from https://huggingface.co/spaces/cbensimon/wan2-1-fast/
"""

import contextlib
from contextvars import ContextVar
from io import BytesIO
from typing import Any, cast
from unittest.mock import patch

import torch
from torch.utils._pytree import tree_map_only
from torch._inductor.package.package import package_aoti
from torch._inductor.package import load_package
from torch.export.pt2_archive._package import AOTICompiledModel
from torch.export.pt2_archive._package_weights import Weights


INDUCTOR_CONFIGS_OVERRIDES = {
    "aot_inductor.package_constants_in_so": False,
    "aot_inductor.package_constants_on_disk": True,
    "aot_inductor.package": True,
}


class ZeroGPUWeights:
    def __init__(self, constants_map: dict[str, torch.Tensor], to_cuda: bool = False):
        if to_cuda:
            self.constants_map = {name: tensor.to("cuda") for name, tensor in constants_map.items()}
        else:
            self.constants_map = constants_map

    def __reduce__(self):
        constants_map: dict[str, torch.Tensor] = {}
        for name, tensor in self.constants_map.items():
            tensor_ = torch.empty_like(tensor, device="cpu").pin_memory()
            constants_map[name] = tensor_.copy_(tensor).detach().share_memory_()
        return ZeroGPUWeights, (constants_map, True)


class ZeroGPUCompiledModel:
    def __init__(self, archive_file: torch.types.FileLike, weights: ZeroGPUWeights):
        self.archive_file = archive_file
        self.weights = weights
        self.compiled_model: ContextVar[AOTICompiledModel | None] = ContextVar("compiled_model", default=None)

    def __call__(self, *args, **kwargs):
        if (compiled_model := self.compiled_model.get()) is None:
            # compiled_model = cast(AOTICompiledModel, torch._inductor.aoti_load_package(self.archive_file))
            # compiled_model = torch._inductor.aoti_load_package(self.archive_file, run_single_threaded=True)
            compiled_model = load_package(self.archive_file, run_single_threaded=True)
            compiled_model.load_constants(self.weights.constants_map, check_full_update=True, user_managed=True)
            self.compiled_model.set(compiled_model)
        return compiled_model(*args, **kwargs)

    def __reduce__(self):
        return ZeroGPUCompiledModel, (self.archive_file, self.weights)


def aoti_compile(
    exported_program: torch.export.ExportedProgram,
    inductor_configs: dict[str, Any] | None = None,
):
    inductor_configs = (inductor_configs or {}) | INDUCTOR_CONFIGS_OVERRIDES
    gm = cast(torch.fx.GraphModule, exported_program.module())
    assert exported_program.example_inputs is not None
    args, kwargs = exported_program.example_inputs
    artifacts = torch._inductor.aot_compile(gm, args, kwargs, options=inductor_configs)
    archive_file = BytesIO()
    files: list[str | Weights] = [file for file in artifacts if isinstance(file, str)]
    package_aoti(archive_file, files)
    (weights,) = (artifact for artifact in artifacts if isinstance(artifact, Weights))
    zerogpu_weights = ZeroGPUWeights({name: weights.get_weight(name)[0] for name in weights}, to_cuda=True)
    return ZeroGPUCompiledModel(archive_file, zerogpu_weights)


@contextlib.contextmanager
def capture_component_call(
    pipeline: Any,
    component_name: str,
    component_method="forward",
):
    class CapturedCallException(Exception):
        def __init__(self, *args, **kwargs):
            super().__init__()
            self.args = args
            self.kwargs = kwargs

    class CapturedCall:
        def __init__(self):
            self.args: tuple[Any, ...] = ()
            self.kwargs: dict[str, Any] = {}

    component = getattr(pipeline, component_name)
    captured_call = CapturedCall()

    def capture_call(*args, **kwargs):
        raise CapturedCallException(*args, **kwargs)

    with patch.object(component, component_method, new=capture_call):
        try:
            yield captured_call
        except CapturedCallException as e:
            captured_call.args = e.args
            captured_call.kwargs = e.kwargs


# Taken from
# https://github.com/huggingface/flux-fast/blob/5027798d7f69a8e0e478df92f48663c40727f8ea/utils/pipeline_utils.py#L198C1-L231C14
def cudagraph(f):
    _graphs = {}
    def f_(*args, **kwargs):
        key = hash(tuple(tuple(kwargs[a].shape) for a in sorted(kwargs.keys())
                         if isinstance(kwargs[a], torch.Tensor)))
        if key in _graphs:
            # use the cached wrapper if one exists. this will perform CUDAGraph replay
            wrapped, *_ = _graphs[key]
            return wrapped(*args, **kwargs)

        # record a new CUDAGraph and cache it for future use
        g = torch.cuda.CUDAGraph()
        in_args, in_kwargs = tree_map_only(torch.Tensor, lambda t: t.clone(), (args, kwargs))
        f(*in_args, **in_kwargs) # stream warmup
        with torch.cuda.graph(g):
            out_tensors = f(*in_args, **in_kwargs)
        def wrapped(*args, **kwargs):
            # note that CUDAGraphs require inputs / outputs to be in fixed memory locations.
            # inputs must be copied into the fixed input memory locations.
            [a.copy_(b) for a, b in zip(in_args, args) if isinstance(a, torch.Tensor)]
            for key in kwargs:
                if isinstance(kwargs[key], torch.Tensor):
                    in_kwargs[key].copy_(kwargs[key])
            g.replay()
            # clone() outputs on the way out to disconnect them from the fixed output memory
            # locations. this allows for CUDAGraph reuse without accidentally overwriting memory
            return [o.clone() for o in out_tensors]

        # cache function that does CUDAGraph replay
        _graphs[key] = (wrapped, g, in_args, in_kwargs, out_tensors)
        return wrapped(*args, **kwargs)
    return f_