撰文|郑建华


【资料图】

更新|赵露阳

tensor和op是神经网络模型最基本的组件:op是模型的节点,tensor是连接节点的边。然而,构建一个tensor并不仅仅是构造一个对象那么简单,至少要考虑以下问题:

要支持节点本地的local tensor,以及分布式的global tensor;

要支持eager和lazy执行模式;

要支持不同的数据类型,包括float、double、int等;

要支持不同设备。

创建tensor的方法

与PyTorch类似,在OneFlow中也可以通过两种主要的方式来创建tensor:Tensor和tensor。这两种方式最终都会创建出OneFlow内部的C++ Tensor对象,即对应Python层的flow.Tensor类型。

1.1 Tensor

Python层的Tensor是在tensor.py(https://github.com/Oneflow-Inc/oneflow/blob/2e6a72c8734b9929191306df35b4284e9caa8126/python/oneflow/framework/tensor.py#L23)中引入的,通过python c api注册的Tensor类型对象,此对象在MakeTensorType

https://github.com/Oneflow-Inc/oneflow/blob/2e6a72c8734b9929191306df35b4284e9caa8126/oneflow/api/python/framework/tensor.cpp#L623)中被定义和返回。

在MakeTensorType中主要通过PyTensorObject_init创建了Tensor对象:

static int PyTensorObject_init(PyObject* self, PyObject* args, PyObject* kwargs) { HANDLE_ERRORS auto* temp = functional::_legacy_tensor_ctor(NULL, args, kwargs); if (PyErr_Occurred()) { throw py::error_already_set(); } auto* _self = (PyTensorObject*)self; _self->data = PyTensor_Unpack(temp); _self->data->set_pyobject(self); // reset temp data to prevent clearing the pyobject // when the temp is deallocated ((PyTensorObject*)temp)->data.reset(); Py_XDECREF(temp); return 0; END_HANDLE_ERRORS_RET(-1)}

通过functional::_legacy_tensor_ctor函数创建了OneFlow内部的c++ Tensor对象:oneflow::one::Tensor,并作为data绑定至Python的Tensor类型。在MakeTensorType中,还通过PyMethodDef(https://github.com/Oneflow-Inc/oneflow/blob/2e6a72c8734b9929191306df35b4284e9caa8126/oneflow/api/python/framework/tensor.cpp#L639-L641)为Tensor注册了很多C++方法,如:

static PyMethodDef PyTensorObject_methods[] = { {"storage_offset", PyTensorObject_storage_offset, METH_NOARGS, NULL}, {"stride", PyTensorObject_stride, METH_NOARGS, NULL}, {"is_contiguous", PyTensorObject_is_contiguous, METH_NOARGS, NULL}, {"contiguous", PyTensorObject_contiguous, METH_NOARGS, NULL}, {"contiguous_", PyTensorObject_contiguous_, METH_NOARGS, NULL}, {"pin_memory", PyTensorObject_pin_memory, METH_NOARGS, NULL}, {"is_pinned", PyTensorObject_is_pinned, METH_NOARGS, NULL}, {"requires_grad_", (PyCFunction)PyTensorObject_requires_grad_, METH_VARARGS | METH_KEYWORDS, NULL}, {"retain_grad", PyTensorObject_retain_grad, METH_NOARGS, NULL}, {"detach", PyTensorObject_detach, METH_NOARGS, NULL}, {"clone", PyTensorObject_clone, METH_NOARGS, NULL}, {"zero_", PyTensorObject_zero_, METH_NOARGS, NULL}, {"register_hook", PyTensorObject_register_hook, METH_O, NULL}, {"_register_post_grad_accumulation_hook", PyTensorObject__register_post_grad_accumulation_hook, METH_O, NULL}, {"global_id", PyTensorObject_global_id, METH_NOARGS, NULL}, {"check_meta_consistency", PyTensorObject_check_meta_consistency, METH_NOARGS, NULL}, {"to_numpy", PyTensorObject_to_numpy, METH_NOARGS, NULL}, {"type", (PyCFunction)PyTensorObject_type, METH_VARARGS | METH_KEYWORDS, NULL},

此外,在Python层通过RegisterMethods(https://github.com/Oneflow-Inc/oneflow/blob/2e6a72c8734b9929191306df35b4284e9caa8126/python/oneflow/framework/tensor.py#L502)也为Tensor注册了一些Python实现的Tensor方法或属性(如tensor.numpy),在OneFlow包初始化时会通过RegisterMethod4Class

https://github.com/Oneflow-Inc/oneflow/blob/2e6a72c8734b9929191306df35b4284e9caa8126/python/oneflow/framework/register_class_method_util.py#L23)完成这些Python方法和属性的注册。RegisterMethod4Class的调用流程如下:

相比于Python实现来说,Tensor的++实现的方法/属性通常具有较高的性能。

1.2 tensor函数

Tensor是类型,而tensor则是函数,flow.tensor函数在oneflow/api/python/functional/tensor_api.yaml中被定义:

- name: "tensor" signature: [ "Tensor (PyObject* data, *, DataType dtype=None, Device device=None, Bool requires_grad=False, Bool pin_memory=False) => TensorWithData", "Tensor (PyObject* data, *, DataType dtype=None, Placement placement, SbpList sbp, Bool requires_grad=False) => GlobalTensorWithData", ] bind_python: True

其C++实现位于tensor_api.yaml.pybind.cpp中,这是构建阶段自动生成的文件。

通过函数签名可以看到,flow.tensor()有两种重载的方法:

TensorWithData

GlobalTensorWithData

它们分别用于构造local tensor和global tensor的构造。和上面的Tensor类似,flow.tensor返回的也是OneFlow内部的oneflow::one::Tensor对象(绑定至Python的Tensor对象)。

1.3 手动构建tensor的两种方式

和PyTorch类似,在OneFlow中常用创建tensor的方式也分为两种:

flow.Tensor

flow.tensor

创建方式示例:

import oneflowimport numpy as nponeflow.tensor([[1., -1.], [1., -1.]])# tensor([[ 1., -1.],# [ 1., -1.]], dtype=oneflow.float32)oneflow.tensor(np.array([[1, 2, 3], [4, 5, 6]]))# tensor([[ 1, 2, 3],# [ 4, 5, 6]], dtype=oneflow.int64)flow.Tensor([[1,2,3],[4,5,6]])

大多数情况下(和PyTorch类似的eager模式),可以通过指定device、dtype、shape等参数创建普通tensor(local tensor);

少数情况下(如OneFlow特有的eager global、lazy模式),需要global tensor时,可以通过指定sbp和placement的方式直接创建global tensor,也可通过tensor.to_global的方式将普通tensor转换为global tensor,可参考:

oneflow.tensor

https://oneflow.readthedocs.io/en/master/generated/oneflow.tensor.html#

global tensor

https://docs.oneflow.org/master/parallelism/03_consistent_tensor.html

2

OneFlow的tensor类型体系

上述内容中介绍的oneflow内部的C++ Tensor对象,实际上其定义位于:oneflow/core/framework/tensor.h,是一个抽象的Tensor类型。

其中LocalTensor即为普通的单卡视角下的Tensor(和PyTorch的Tensor类似);GlobalTensor则为OneFlow所特有的全局视角下的Tensor(通常用于eager global模式或lazy模式下)。Tensor使用了Bridge模式,每个Tensor子类内部有一个TensorImpl字段,负责抽象Tensor的实际实现:

3

local tensor的构造

我们以flow.tensor([[1,2,3],[4,5,6]])为例,看一下tensor构造的过程。主要的流程如下:

在这个例子中,由于使用的是flow.tensor方法创建tensor(且为普通的local tensor)所以会用到在oneflow/api/python/functional/tensor_api.yaml中定义的TensorWithData方法,其实现,是位于oneflow/api/python/functional/tensor_api.cpp的TensorWithDataFunctor:

class TensorWithDataFunctor { public: Maybe operator()(PyObject* data, const Optional>& dtype, const Optional>& device, const bool requires_grad, const bool pin_memory) const { ... if (PyTensor_Check(data)) { // Throw warnings like pytorch. auto ret = PyErr_WarnEx( PyExc_UserWarning, "To copy construct from a tensor, it is recommended to use sourceTensor.clone().detach() " "or sourceTensor.clone().detach().requires_grad_(True), rather than " "oneflow.tensor(sourceTensor).", 1); if (ret != 0) { return Error::RuntimeError(); } const auto& other = PyTensor_Unpack(data); return MakeTensorFromOtherTensor(other, dtype, device, requires_grad, pin_memory); } else { // Make tensor from python sequence or numpy array. return MakeLocalTensorFromData(data, dtype, device, requires_grad, pin_memory); } }};

由于这里传入的data是一个Python的list对象,所以最终会调用MakeLocalTensorFromData方法,创建tensor主要的逻辑都在这个函数中。其中大量调用Python和Numpy的接口,检查PyObject的数据类型,获取Shape

https://github.com/Oneflow-Inc/oneflow/blob/2e6a72c8734b9929191306df35b4284e9caa8126/oneflow/api/python/utils/tensor_utils.cpp#L184)和DataType(https://github.com/Oneflow-Inc/oneflow/blob/2e6a72c8734b9929191306df35b4284e9caa8126/oneflow/api/python/utils/tensor_utils.cpp#L185),如果用户没有制定device,默认会设置为CPU设备(https://github.com/Oneflow-Inc/oneflow/blob/2e6a72c8734b9929191306df35b4284e9caa8126/oneflow/api/python/utils/tensor_utils.cpp#L191)。

后面主要是调用EmptyFunctor

https://github.com/Oneflow-Inc/oneflow/blob/2e6a72c8734b9929191306df35b4284e9caa8126/oneflow/api/python/utils/tensor_utils.cpp#L194)和SwitchCopyLocalTensorFromUntypedArray(https://github.com/Oneflow-Inc/oneflow/blob/2e6a72c8734b9929191306df35b4284e9caa8126/oneflow/api/python/utils/tensor_utils.cpp#L195)。前者为tensor分配内存,后者进行数据拷贝,两个步骤都会通过虚拟机指令完成。其中EmptyFunctor会走普通的OpCall指令、而CopyLocalTensorFromUntypedArray会根据是否需要同步copy走到AccessBlobByCallback/SyncAccessBlobByCallback指令。

为什么要通过虚拟机指令完成呢?无论是内存资源的分配,还是数据拷贝,CPU和CUDA等不同设备上的操作都不一样。之前讨论Op/Kernel时已经看到,在OneFlow中所有动静态图任务执行、eager模式下op/kernel执行、内存/显存的分配和释放、device、stream等统一由虚拟机进行管理。

3.1 分配内存:EmptyFunctor

matmul和relu(inplace=false时)等操作在执行过程中也会创建output tensor。之前讨论relu时重点关注了op和kernel的计算逻辑,而忽略了tensor相关的内容。

而这里只需要先构造一个空tensor对象,不需要其它计算,所以是一个Empty操作,Empty op对应的kernel——EmptyKernel(https://github.com/Oneflow-Inc/oneflow/blob/2e6a72c8734b9929191306df35b4284e9caa8126/oneflow/user/kernels/empty_kernel.cpp#L30)没有实质性的计算逻辑,只是先根据shape、dtype、device信息创建一个空tensor,等待后续将实际的数据从内存中copy至此空tensor,从而完成整个tensor的创建过程。

EmptyFunctor同样和其他functor一样,最终会被Dispacth至对应的interpreter被解释执行,这里由于是eager模式下的local tensor,EmptyFunctor最终会进入eager local interpreter,交给NaiveInterpret(https://github.com/Oneflow-Inc/oneflow/blob/2e6a72c8734b9929191306df35b4284e9caa8126/oneflow/core/framework/op_interpreter/eager_local_op_interpreter.cpp#L74)方法处理。流程如下:

1. 在构造EagerLocalTensorImpl(https://github.com/Oneflow-Inc/oneflow/blob/2e6a72c8734b9929191306df35b4284e9caa8126/oneflow/core/framework/op_interpreter/eager_local_op_interpreter.cpp#L110)对象,用于存放tensor结果。但这只是一个壳子,还没有为tensor的数据分配存储空间。

2. 之后会初始化EagerBlobObject(https://github.com/Oneflow-Inc/oneflow/blob/2e6a72c8734b9929191306df35b4284e9caa8126/oneflow/core/framework/op_interpreter/eager_local_op_interpreter.cpp#L114)、TensorStorage(https://github.com/Oneflow-Inc/oneflow/blob/2e6a72c8734b9929191306df35b4284e9caa8126/oneflow/core/framework/tensor_impl.cpp#L120),这样tensor主要的字段基本构建完毕

3. 然后构造OpCall指令、提交虚拟机PhysicalRun(https://github.com/Oneflow-Inc/oneflow/blob/2e6a72c8734b9929191306df35b4284e9caa8126/oneflow/core/framework/op_interpreter/eager_local_op_interpreter.cpp#L134-L136),等待vm的调度执行。

OpCall对应的指令策略最终会进入oneflow/core/vm/op_call_instruction_policy.cpp,并在Prepare方法中通过AllocateOutputBlobsMemory方法对TensorStorage完成实际的内存分配;在Compute方法中启动(empty op对应的)实际的kernel执行。

3.2 拷贝数据:SwitchCopyLocalTensorFromUntypedArray

SwitchCopyMirroredTensorFromUntypedArray其实是MAKE_SWITCH_ENTRY(https://github.com/Oneflow-Inc/oneflow/blob/2e6a72c8734b9929191306df35b4284e9caa8126/oneflow/api/python/utils/tensor_utils.cpp#L150)宏展开后的函数名。宏展开后的代码如下。实际会调用CopyLocalTensorFromUntypedArray(https://github.com/Oneflow-Inc/oneflow/blob/2e6a72c8734b9929191306df35b4284e9caa8126/oneflow/api/python/utils/tensor_utils.cpp#L68)。

templatestatic Maybe SwitchCopyLocalTensorFromUntypedArray( const std::tuple& switch_tuple, Args&& ... args) { static const std::map, std::function(Args && ...)>> case_handlers { {SwitchCase(DataType::kFloat), [](Args&&... args) { return CopyLocalTensorFromUntypedArray(std::forward(args)...); }}, // ... }; return case_handlers.at(switch_tuple)(std::forward(args)...);};

CopyLocalTensorFromUntypedArray方法如下:

templateMaybe CopyLocalTensorFromUntypedArray(const std::shared_ptr& tensor, PyObject* array) { return CopyBetweenLocalTensorAndNumpy(tensor, array, CopyFromNumpyArray, "mut", /*block_host_until_done=*/false);}

其内部实际调用了CopyBetweenLocalTensorAndNumpy方法。

CopyBetweenLocalTensorAndNumpy

顾名思义,这个方法主要是用在numpy和tensor之间进行数据copy的。其中第3个参数:CopyFromNumpyArray实际是一个函数回调的callback方法,其主要通过SyncAutoMemcpy进行array和tensor(blob)之间的内存拷贝:

void CopyFromNumpyArray(ep::Stream* stream, const std::shared_ptr& eager_blob_object, const NumPyArrayPtr& array_ptr) { SyncAutoMemcpy(stream, eager_blob_object->mut_dptr(), array_ptr.data(), eager_blob_object->ByteSizeOfBlobBody(), eager_blob_object->mem_case(), memory::MakeHostMemCase());}

继续看CopyBetweenLocalTensorAndNumpy(https://github.com/Oneflow-Inc/oneflow/blob/2e6a72c8734b9929191306df35b4284e9caa8126/oneflow/api/python/utils/tensor_utils.h#L93)方法,其中最关键的是:

JUST(PhysicalRun([&](InstructionsBuilder* builder) -> Maybe { return builder->AccessBlobByCallback( tensor, [array_ptr, Copy](ep::Stream* stream, const std::shared_ptr& eager_blob_object) { Copy(stream, eager_blob_object, array_ptr); }, modifier); }));

通过InstructionsBuilder构建了AccessBlobByCallback指令,参数为上面通过EmptyFuncor创建的空tensor、callback的函数指针及参数、以及modifier(string "mut"表示可动态修改)。

AccessBlobByCallback

和OpCall类似,InstructionsBuilder调用AccessBlobByCallback时,也会实际构造对应的vm指令策略——AccessBlobArgCbInstructionPolicy并派发至vm,等待被调度和实际执行:

templateMaybe InstructionsBuilder::AccessBlobByCallback( const T tensor, const std::function&)>& callback, const std::string& modifier) { const std::shared_ptr& eager_blob_object = JUST(tensor->eager_blob_object()); Symbol device = JUST(GetDevice(tensor)); ... Symbol stream = JUST(GetDefaultStreamByDevice(device)); JUST(SoftSyncStream({eager_blob_object}, stream)); auto instruction = intrusive::make_shared( // Never replace `stream` with producer_stream or last_used_stream. JUST(Singleton::Get()->GetVmStream(stream)), std::make_shared(eager_blob_object, callback, modifier)); instruction_list_->EmplaceBack(std::move(instruction)); return Maybe::Ok();}

等该条AccessBlobArgCbInstructionPolicy指令实际执行时,会在指令的Compute(https://github.com/Oneflow-Inc/oneflow/blob/2e6a72c8734b9929191306df35b4284e9caa8126/oneflow/core/vm/access_blob_arg_cb_instruction_policy.h#L79)方法中调用callback完成从tensor的blob <-> numpy的ndarray之间的数据copy,至此拷贝过程结束,flow.tensor的创建全部完成。

(本文经授权后发布。原文:

https://segmentfault.com/a/1190000041989895)

参考资料

On‍eFlow源码:https://github.com/Oneflow-Inc/oneflow

‍OneFlow源码解析:Op、Kernel与解释器

OneFlow源码解析:算子指令在虚拟机中的执行

其他人都在看

OneFlow v0.8.0正式发布

9篇分布式机器学习系统经典论文

深度学习硬件的过去、现在和未来

从Core Dump中提取CUDA的报错信息

分布式深度学习编程新范式:Global Tensor

OneEmbedding:单卡训练TB级推荐模型不是梦

大模型训练难?效率超群、易用的“李白”模型库来了

欢迎体验OneFlow v0.8.0:GitHub - Oneflow-Inc/oneflow: OneFlow is a deep learning framework designed to be user-friendly, scalable and efficient.OneFlow is a deep learning framework designed to be user-friendly, scalable and efficient. - GitHub - Oneflow-Inc/oneflow: OneFlow is a deep learning framework designed to be user-friendly, scalable and efficient.https://github.com/Oneflow-Inc/oneflow

推荐内容