Overmind：将机器学习模型加载时间从15秒缩短到0.2秒

# 摘自 torch.multiprocessing.reductions，省略了大部分代码
def reduce_tensor(tensor):
    ...
    storage = tensor._typed_storage()
    ...
    metadata = (
        tensor.storage_offset(),
        tensor.size(),
        tensor.stride(),
        tensor.requires_grad,
    )
    return (rebuild_tensor, (type(tensor), storage, metadata))

// 摘自 torch/csrc/Storage.cpp
static PyObject* THPStorage_get(THPStorage* self, PyObject* index) {
    // ...省略无关代码...

    auto new_storage_impl = make_storage_impl(
        c10::StorageImpl::use_byte_size_t(),
        slicelength,
        at::DataPtr(
            static_cast<void*>(data + start),
            old_storage_impl,
            [](void* s) {
              c10::raw::intrusive_ptr::decref(static_cast<at::StorageImpl*>(s));
            },
            old_storage_impl->device()),
        old_storage_impl->allocator(),
        /* resizable */ false,
        device_opt);

    PyObject* _ret =
        THPStorage_NewWithStorage(Py_TYPE(self), std::move(new_storage_impl));

    return _ret;
}

void initOvermindHelpers(py::module m) {
    // ...
    m.def("_make_untyped_storage", [](py::buffer b) {
        auto info = new py::buffer_info(b.request());

        return pybind11::reinterpret_steal<py::object>(THPStorage_NewWithStorage(
            THPStorageClass,
            c10::make_intrusive<at::StorageImpl>(
                c10::StorageImpl::use_byte_size_t(),
                info->size,
                at::DataPtr(
                    info->ptr, info,
                    [](void* ptr) {
                        py::gil_scoped_acquire gil;
                        auto b = static_cast<py::buffer_info*>(ptr);
                        delete b;
                    },
                    at::DeviceType::CPU
                ),
                /*allocator=*/nullptr,
                /*resizable=*/false,
            )
        ));
    });
}

class SharedMemory:
    @classmethod
    def create(cls, shift):
        # 在服务端调用
        libc = ctypes.CDLL(None)
        name = _make_filename(shift).encode('utf-8')
        fd = libc.memfd_create(name, os.O_RDWR)
        os.ftruncate(fd, 1 << shift)
        mem_id = (os.getpid(), fd)
        return cls(fd=fd, mem_id=mem_id)

    @classmethod
    def rebuild(cls, mem_id):
        # 在客户端调用
        pid, fd = mem_id
        local_fd = os.open(f'/proc/{pid}/fd/{fd}', os.O_RDWR)
        return cls(fd=local_fd, mem_id=mem_id)

    def get_buffer(self):
        # 两端都会调用
        self._mmap = mmap.mmap(self._fd, size)
        self._buf = memoryview(self._mmap)
        return self._buf

# Hoarder 和 borrower 是 SharedMemory 的封装，包含内存池等
# 不太有趣的实现细节。
def _reduce_storage(storage):
    # 服务端调用
    device = storage.device
    storage = storage.cpu()

    # 将内容存入共享内存
    # `frag` 包含定位内容所需的全部信息
    frag = hoarder.put(storage)

    return (_rebuild_storage_on_client, (frag, device))

def _rebuild_storage_on_client(frag, device):
    # 客户端调用
    mv = borrower.borrow(frag)  # 从共享内存获取 memoryview
    storage = _make_untyped_storage(mv)  # 零拷贝！
    if device.type == 'cuda':
        return storage.cuda(device.index)
    return storage

class OvermindPickler(dill.Pickler):
    ...

OvermindPickler.register(torch.storage.UntypedStorage, _reduce_storage)

def diffusers_dyn_module_workaround():
    from diffusers.utils.constants import HF_MODULES_CACHE
    modpath = Path(HF_MODULES_CACHE) / "diffusers_modules/__init__.py"
    spec = importlib.util.spec_from_file_location("diffusers_modules", modpath)
    sys.modules["diffusers_modules"] = importlib.util.module_from_spec(spec)

def _reduce_bnb_param(p):
    dev = p._prev_device
    assert p.quant_state
    return (_rebuild_bnb_param, (type(p), p.data, p.quant_state.as_dict(packed=True), dev))


def _rebuild_bnb_param(typ, data, qs_dict, dev):
    return typ.from_prequantized(data, qs_dict, device=dev)


def bitsandbytes_quirks():
    try:
        import bitsandbytes
    except ImportError as e:
        return

    ForkingPickler.register(bitsandbytes.nn.modules.Params4bit, _reduce_bnb_param)
    ForkingPickler.register(bitsandbytes.nn.modules.Int8Params, _reduce_bnb_param)

from accelerate.hooks import remove_hook_from_module
remove_hook_from_module(model, True)
model.__dict__.pop('to', None)  # 移除告警用的 monkeypatch
model.__dict__.pop('cuda', None)

def stable_fast_quirks():
    ...

    # 将 dataclass 类型 pickle 存储，而不是仅仅放入容器中（那样在 torch.jit.save 后就丢了）
    def flatten_dataclass(obj):
        from sfast.utils.flat_tensors import flatten_bytes, flatten_dict
        import dataclasses
        d = dict((field.name, getattr(obj, field.name))
                for field in dataclasses.fields(obj))
        import pickle
        pickled = pickle.dumps(obj.__class__)
        return flatten_bytes(pickled) + flatten_dict(d)

    def unflatten_dataclass(tensors, start):
        from sfast.utils.flat_tensors import unflatten_bytes, unflatten_dict
        import pickle
        pickled, start = unflatten_bytes(tensors, start)
        clz = pickle.loads(pickled)
        content, start = unflatten_dict(tensors, start)
        return clz(**content), start

    sfast.utils.flat_tensors.flatten_dataclass = flatten_dataclass
    sfast.utils.flat_tensors.unflatten_dataclass = unflatten_dataclass

void initOvermindHelpers(py::module m) {
    // ...
    m.def("import_ir_module_from_buffer_0copy",
        [](std::shared_ptr<torch::jit::CompilationUnit> cu, py::buffer buffer) {
            auto info = buffer.request();
            imemstream in((char*)info.ptr, info.size);  // 零拷贝！
            return import_ir_module(std::move(cu), in, ...);
        }
    );
}

import overmind.api
overmind.api.monkey_patch_all()

import torch
from diffusers.models import AutoencoderKL

from diffusers import (
    ControlNetModel,
    StableDiffusionControlNetPipeline,
)

vae = AutoencoderKL.from_pretrained(
    "lemon2431/ChineseInkComicStrip_v10",
    subfolder="vae",
    torch_dtype=torch.float16,
)
controlnet_depth = ControlNetModel.from_pretrained(
    "lllyasviel/control_v11f1p_sd15_depth",
    torch_dtype=torch.float16,
    variant="fp16",
)
controlnet_edge = ControlNetModel.from_pretrained(
    "lllyasviel/control_v11p_sd15_softedge",
    torch_dtype=torch.float16,
    variant="fp16",
)

pipeline = StableDiffusionControlNetPipeline.from_pretrained(
    "lemon2431/ChineseInkComicStrip_v10",
    vae=vae,  # 注意这里！
    controlnet=[controlnet_edge, controlnet_depth],  # 还有这里！
    torch_dtype=torch.float16,
    safety_checker=None,
)

pipeline.to('cuda')

In [1]: t = torch.ones((5808, 1024, 1024), dtype=torch.uint8)

In [2]: %time a = t.cuda()
CPU times: user 976 ms, sys: 874 μs, total: 977 ms
Wall time: 976 ms

In [3]: %timeit a = t.cuda()
1.01 s ± 56.7 ms per loop (mean ± std. dev. of 7 runs, 1 loop each)