AVhG dZddlZddlZddlmZmZmZmZddlm Z ddlm Z ddlm Z ddl mZddlmZdd lmZdd lmZdd lmZdaej0Zed g dedefdee j4defdZedg ej8ddej:de j<fdZedg ej:de j<fdZ edg dee j<fdZ!edg de"fdZ#edg de$fdZ%edg  d?d ede j4d!ee j4dejLfd"Z'ed#g d$eede j4defd%Z(ed&g d edeefd'Z)ed(g d ejLde j4fd)Z*ed*g d ejLde j4ddfd+Z+ed,g  d?d ejLde j4d-ee"dejLfd.Z,ed/g  d?d ejLd0ejLd-ee"dejLfd1Z-ed2g d3e$ddfd4Z.de j^fd5Z0d@d6Z1ejdd7d8Z3ejdd9d:Z4ejdd;d<Z5ejdd=d>Z6y)AzCore DTensor Python API.N)AnyCallableOptionalSequence)dtensor_device)gen_dtensor_ops)layout)context)opstensor) deprecation) tf_exportz%experimental.dtensor.call_with_layout)v1fn.r returnc:|ptjrJt|j5t j |5||i|cdddcdddSt ||i||S||i|S#1swYnxYw ddd#1swY'xYw)aECalls a function in the DTensor device scope if `layout` is not None. If `layout` is not None, `fn` consumes DTensor(s) as input and produces a DTensor as output; a DTensor is a tf.Tensor with layout-related attributes. If `layout` is None, `fn` consumes and produces regular tf.Tensors. Args: fn: A supported TF API function such as tf.zeros. layout: Optional, the layout of the output DTensor. *args: Arguments given to `fn`. **kwargs: Keyword arguments given to `fn`. Returns: The return value of `fn` transformed to a DTensor if requested. N)r executing_eagerly default_meshmesh_dtensor_device_default_layoutrelayout)rr argskwargss M/home/dcms/DCMS/lib/python3.12/site-packages/tensorflow/dtensor/python/api.pycall_with_layoutr%s(   "  $%   . .v 6 %T$V$ % %%%b$)&)6 22 T V   % % %%%%s#BA; B;B BBzexperimental.dtensor.run_onz)Use `dtensor.default_mesh` scope instead.rc#TKt|5ddddy#1swYyxYww)a,Runs enclosed functions in the DTensor device scope. This function returns a scope. All the ops and tf.functions in this scope will run on the DTensor device using the mesh provided. This is useful for wrapping any tf.function that doesn't take a DTensor as input but would like to produce DTensor as result. The scope will also make sure all small constants be replicated as DTensor. Args: mesh: A Mesh instance to extract a default mesh from. Yields: A context in which all ops and tf.functions will run on the DTensor device. N)rrs rrun_onr Cs'$D     s ( (%(z!experimental.dtensor.default_meshc#6Kt|tjstdt |t j |5tjt5dddddddy#1swYxYw#1swYyxYww)aYSets the default DTensor device mesh to use for enclosed functions. This function returns a scope. All the ops and tf.functions in this scope will default to this DTensor mesh if a mesh cannot be inferred from any of the inputs This is useful for wrapping any tf.function that doesn't take a DTensor as input but would like to produce DTensor as result. The scope will also make sure all small constants are replicated as DTensors. Args: mesh: A Mesh instance to extract a default mesh from. Yields: A context in which all ops and tf.functions will run on the given mesh. z Expect `mesh` to be `Mesh`, got N) isinstance layout_libMesh ValueErrortyper_experimental_default_meshr device device_namers rrrYs}$ D*// * 7T |D EE33D9 KM "         s6A B B +B0B 8 BB B  BBz%experimental.dtensor.get_default_meshc0tytjS)zReturn the default mesh under the current dtensor device context. In the case that dtensor device system is not initialized, this function will return None. Returns: The current default mesh for the dtensor device context. N)_dtensor_singleton_current_default_meshrget_default_meshr/ss   3 33r.z experimental.dtensor.device_namec*tjS)zReturns the singleton DTensor device's name. This function can be used in the following way: ```python import tensorflow as tf with tf.device(dtensor.device_name()): # ... ``` )rnamer-r.rr)r)s    r.zexperimental.dtensor.is_dtensorc4tj|S)a6Check whether the input tensor is a DTensor. In Python, a DTensor has the same type as a `tf.Tensor`. This method will let you check and handle the tensor differently if a tf.Tensor is a DTensor. Args: tensor: an object to be checked. Returns: bool, True if the given tensor is a DTensor. )r is_dtensorr s rr3r3s   % %f --r.z!experimental.dtensor.copy_to_meshr source_layoutc~t||S)a:Copies a tf.Tensor onto the DTensor device with the given layout. Copies a regular tf.Tensor onto the DTensor device. Use the mesh attached to `layout` as target mesh. This method currently only supports replicated layouts, or one-to-one copies for sharded layouts. Args: tensor: A regular tf.Tensor to be copied as a DTensor. layout: Target layout (and mesh) for the result DTensor. source_layout: Source layout of the tensor before copy. This argument is deprecated. Returns: A DTensor on the DTensor device with the given layout. )r)r r r4s r copy_to_meshr6s( && !!r.zexperimental.dtensor.packtensorsc6tj||S)aPacks `tf.Tensor` components into a DTensor. Packing and unpacking are inverse operations: ``` * unpack(pack(tensors)) == tensors * pack(unpack(dtensor)) == dtensor ``` 1. For any DTensor on the mesh, `unpack` returns the raw components placed on each underlying device. 2. Packing these raw components in the same order using `pack` returns a DTensor which should be identical to the original DTensor--both the content value and the layout. **Shape, Rank, and Scalars**: The rank of the DTensor is the same as the rank of its raw components, i.e., rank is preserved. This leads to a consistent interpretation for packing scalar values into a DTensor. The only valid layout for a scalar value is fully replicated, and the individual components must be identical scalars. Each input `tensors[i]` will be copied to `layout.mesh.local_device[i]` if not already on the local device. Non-local components should not be passed to `pack`; use `copy_to_mesh` and `relayout` to place tensors on all global devices on a mesh. It is the caller's responsibility to ensure that the underlying values for `pack` adhere to the specified layout, and that only as many values are specified as there are local devices. Pack does not move data between clients. See examples below for more detail about layouts. For example, assume we have a mesh `[X(2), Y(3)]`, which has in total 6 underlying devices. Futuremore, assume that the device location mapping is the following: ``` device_ID | location X, Y 0 0, 0 1 0, 1 2 0, 2 3 1, 0 4 1, 1 5 1, 2 ``` 1. For 1-D vector DTensor with shape `[128]` with layout `[mesh.X]` and value as `range(128)`, the raw components will have shape `[64]` each, and the raw components will be: ``` device_ID | raw component 0 range(0, 64) 1 range(0, 64) 2 range(0, 64) 3 range(64, 128) 4 range(64, 128) 5 range(64, 128) ``` This also means for a 1-D DTensor with shape `[2]` and layout `[mesh.X]`, the raw components have shape `[1]` rather than the shape for scalar values `[]`. 2. For 2-D vector DTensor with shape `[2, 3]` with layout `[mesh.X, mesh.Y]` and value as `range(6)`, this is basically a fully-sharded DTensor. From global view, the content looks like ``` [ [0.0, 1.0, 2.0], [3.0, 4.0, 5.0], ] ``` The raw components will have shape `[1, 1]` each, and have the following content: ``` device_ID | raw component 0 [[0.0]] 1 [[1.0]] 2 [[2.0]] 3 [[3.0]] 4 [[4.0]] 5 [[5.0]] ``` 3. For a scalar value `123.0` DTensor, it can only have one legitimate layout `[]` (no dimension, but fully replicated). The raw components will have shape `[]` each, and have the following content: ``` device_ID | raw component 0 123.0 1 123.0 2 123.0 3 123.0 4 123.0 5 123.0 ``` Again, caller of `pack` is expected to provide 6 identical value raw components with scalar shapes. 4. For 3-D vector DTensor with shape `[2, 2, 3]` with layout `[X, unsharded, unsharded]` and value as `range(12)`, From global view, the content looks like: ``` [ [ [0.0, 1.0, 2.0], [3.0, 4.0, 5.0], ], [ [6.0, 7.0, 8.0], [9.0, 10., 11.], ], ] ``` The raw components will have shape `[1, 2, 3]` each, and have the following content: ``` device_ID | raw component 0 range(6).reshape([1, 2, 3]) 1 range(6).reshape([1, 2, 3]) 2 range(6).reshape([1, 2, 3]) 3 range(6, 12).reshape([1, 2, 3]) 4 range(6, 12).reshape([1, 2, 3]) 5 range(6, 12).reshape([1, 2, 3]) ``` Args: tensors: The list of local tensor components to pack into a DTensor. layout: The layout of the DTensor to be created. Returns: A DTensor created from the individual component tensors. Raises: RuntimeError: When `pack` is not called eagerly. )rpack)r7r s rr9r9sh     00r.zexperimental.dtensor.unpackc4tj|S)aXUnpacks a DTensor into `tf.Tensor` components. Packing and unpacking are inverse operations: ``` * unpack(pack(tensors)) == tensors * pack(unpack(dtensor)) == dtensor ``` 1. For any DTensor on the mesh, `unpack` returns the raw components placed on each underlying device. 2. Packing these raw components in the same order using `pack` returns a DTensor which should be identical to the original DTensor--both the content value and the layout. See the documentation for `pack` for more information about how packing and unpacking works. Args: tensor: The DTensor to unpack. Returns: The individual component tensors of the DTensor. This will include only the client-local components, i.e. the components placed on the local devices. Raises: RuntimeError: When `unpack` is not called eagerly. )runpackr s rr;r;Vs<   ! !& ))r.z!experimental.dtensor.fetch_layoutc4tj|S)zFetches the layout of a DTensor. Args: tensor: The DTensor whose layout is to be fetched. Returns: The `Layout` of this DTensor. Raises: RuntimeError: When not called eagerly. )r fetch_layoutr s rr=r={s   ' ' //r.z!experimental.dtensor.check_layoutc~t||k7r/tdtt|zdzt|zy)zAsserts that the layout of the DTensor is `layout`. Args: tensor: A DTensor whose layout is to be checked. layout: The `Layout` to compare against. Raises: ValueError: If the layout of `tensor` does not match the supplied `layout`. zLayout of tensor: z!, did not match expected layout: N)r=r%str)r r s r check_layoutr@sL&V# )C V0D,EE89;>v;G HH$r.zexperimental.dtensor.relayoutr1c|j}t|j5tj|||cdddS#1swYyxYw)a.Changes the layout of `tensor`. Changes the layout of `tensor` to `layout`. This is used to fine-tune the behavior of ops following/connected to `tensor`, such as choosing one SPMD expansion pattern over another. This works by forward propagating `layout` to connected TensorFlow computation graphs during layout propagation. Currently, only converting layouts from replicated to sharded or sharded to replicated per mesh dimension is supported. That is, "x, y" -> "unsharded, y" is supported, while "x, y" -> "z, y" is not supported. We also support a special "match" sharding spec, which instructs the relayout to act as an identity operation with respect to any sharding on these mesh dimensions. Relayout is internally lowered to a set of Split and/or AllToAll ops. When tensor layouts are converted from replicated to sharded, the cost is comparatively low because we only insert Split ops and no cross-device communication is needed. However, when tensor layouts are converted from sharded to replicated, cross-device communication may occur, causing potential performance impact. Args: tensor: A DTensor to specify a new layout for. layout: A Layout object specifying a new sharding spec. name: name of the Op. Returns: A DTensor output from the Relayout op. )r1N) to_stringrrrr)r r r1 layout_strs rrrsJH!*FKK C  # #FJT BCCCs AAz"experimental.dtensor.relayout_like layout_tensorc2tj|||S)auChanges the layout of `tensor` to the same as `layout_tensor`. `relayout_like` is often used inside a `tf.function`, to ensure a tensor is placed to the same mesh and with the same layout as another tensor. The backward gradient of a `relayout` is a `relayout_like` operation, to ensure the backward tensor has the same layout as the forward input tensor: ``` @ops.RegisterGradient("Relayout") def _relayout_gradient(op, grad): return relayout_like(grad, layout_input=op.inputs[0]) ``` Here is another illustrative example: ``` @tf.function def func(x): z = tf.ones(x.shape) z = dtensor.relayout_like(z, x) return x + z with dtensor.default_mesh(cpu_mesh): x = tf.ones((4, 4)) with dtensor.default_mesh(gpu_mesh): y = func(x) # y would be on the cpu mesh, following the mesh of x. ``` Args: tensor: A DTensor to specify a new layout for. layout_tensor: A Tensor object whose layout will be used for the layout of result. The shape and type of layout_tensor are irrelevant. name: name of the Op. Returns: A DTensor output from the RelayoutLike op. )input layout_inputr1)r relayout_like)r rDr1s rrHrHs`  & & T r.z*experimental.dtensor._reset_dtensor_deviceis_asyncc6tjg|}|ay)zResets the Eager execution device for DTensor. This function is only intended for testing and diagnostics. Args: is_async: If True, the device uses async execution. )meshesrIN)r DTensorDevicer+)rIr(s rreset_dtensor_devicerMs  ' 'rH E&r.cjt5t tdtcdddS#1swYyxYw)NT)rI)_dtensor_singleton_lockr+rMr-r.rrrs+!D) s)2crt5ttjdadddy#1swYyxYwN)rOr+clear_tpu_core_idsr-r.r_resetrSs2%++-s-6RelayoutcNtj||jd}|SNr)rGrrHinputsopgrads r_relayout_gradientr\s!  & &t"))A, G$ +r. RelayoutLikecRtj||jd}|dfSrVrWrYs r_relayout_grad_gradientr_ s*  & &t"))A, G$ tr. CopyToMeshcNtj||jd}|SNr) forward_inputrcopy_to_mesh_gradrXrYs r_copy_to_mesh_gradientrf)s'  * * IIaL $ +r.CopyToMeshGradcRtj||jd}|dfSrbrdrYs r_copy_to_mesh_grad_gradientri2s,  * * IIaL $ tr.rQ)rN)7__doc__ contextlib threadingtypingrrrrtensorflow.dtensor.pythonrrr r#tensorflow.python.eagerr tensorflow.python.frameworkr r tensor_libtensorflow.python.utilr tensorflow.python.util.tf_exportrr+LockrOLayoutr deprecatedcontextmanagerr$r rr/r?r)boolr3Tensorr6r9r;r=r@rrHrMrLrrSRegisterGradientr\r_rfrir-r.rr{sO4445:++<.6()..*  2r:#"%#&'089J9J0K),;: (R0IJ   K1 & .26  z 7 0 2r: 4(:??3 4; 4 -"5  S  6   ,4 .$ .5 .& .2626" "   "J--.";E:K:K"7". &2.S1(3-S1):):S1sS1/S1l (R0*3*8C=*1*H .26 0** 0z/@/@ 07 0 .26 H** HJ4E4E H$ H7 H *r2%C   %C   %C 3-%C %C3%CP /B71   1$$1 3-1 181h 7B? 4 D @ 55j!" n%&l#$&'(r.