撰文|郑建华
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。
1
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T ensor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:
创建方式示例:import oneflowimport numpy as np
oneflow.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:
( 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<Tensor> operator()(PyObject* data, const Optional<Symbol<DType>>& dtype, const Optional<Symbol<Device>>& 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 ) 。template<typename... Args>static Maybe<void> SwitchCopyLocalTensorFromUntypedArray( const std::tuple<DataType>& switch_tuple, Args&& ... args) { static const std::map<std::tuple<DataType>, std::function<Maybe<void>(Args && ...)>> case_handlers { {SwitchCase(DataType::kFloat), [](Args&&... args) { return CopyLocalTensorFromUntypedArray<float>(std::forward<Args>(args)...); }}, // ... }; return case_handlers.at(switch_tuple)(std::forward<Args>(args)...);}; CopyLocalTensorFromUntypedArray 方法如下:template<typename T>Maybe<void> CopyLocalTensorFromUntypedArray(const std::shared_ptr<Tensor>& tensor, PyObject* array) { return CopyBetweenLocalTensorAndNumpy<T>(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<vm::EagerBlobObject>& 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<void> { return builder->AccessBlobByCallback( tensor, [array_ptr, Copy](ep::Stream* stream, const std::shared_ptr<vm::EagerBlobObject>& 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:template<typename T>Maybe<void> InstructionsBuilder::AccessBlobByCallback( const T tensor, const std::function<void(ep::Stream*, const std::shared_ptr<vm::EagerBlobObject>&)>& callback, const std::string& modifier) { const std::shared_ptr<vm::EagerBlobObject>& eager_blob_object = JUST(tensor->eager_blob_object()); Symbol<Device> device = JUST(GetDevice(tensor)); ... Symbol<Stream> stream = JUST(GetDefaultStreamByDevice(device)); JUST(SoftSyncStream({eager_blob_object}, stream)); auto instruction = intrusive::make_shared<vm::Instruction>( // Never replace `stream` with producer_stream or last_used_stream. JUST(Singleton<VirtualMachine>::Get()->GetVmStream(stream)), std::make_shared<vm::AccessBlobArgCbInstructionPolicy>(eager_blob_object, callback, modifier)); instruction_list_->EmplaceBack(std::move(instruction)); return Maybe<void>::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
- On eFlow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:https://github.com/Oneflow-Inc/oneflow/
本文分享自微信公众号 - OneFlow(OneFlowTechnology)。
如有侵权,请联系 support@oschina.cn 删除。
本文参与“OSC源创计划”,欢迎正在阅读的你也加入,一起分享。
|