AVh> dZddlZddlZddlZddlmZmZmZmZddl Z ddl m Z ddl mZddl mZddlmZddlmZdd lmZdd lmZdd lmZdd lmZdd lmZddlmZddlmZdaej@Z!Gdde"Z#y)z>Propagates information about tensor layouts across operations.N)AnyListSequenceSet)attr_value_pb2)config)layout)_pywrap_dtensor_device)context)core)device)dtypes)ops) sparse_tensor) tensor_util) _pywrap_utilsceZdZdZ ddeej fdZdZdej fdZ e de ej fdZ d e ed ejdefd Zd ede efd Zd edejfdZdedefdZdZdZdZdZdZdeejfdZej8dej fdZej8d ejfdZy) DTensorDevicezAWraps a custom device which attempts to propagate tensor layouts.meshescRtd|Drtdj|tj}t5dj|j t |_t dz adddtj|j||\}}tj||j|||_ d|_ d|_ t|_t!j"|_|D]}|j'|y#1swYxYw)aCreate a new DTensorDevice which executes ops on `underlying_device`. Args: meshes: A list of `Mesh` objects indicating groups of devices to execute on. These may also be registered lazily. is_async: Indicates whether DTensor operations on this client will return immediately (with "non-ready" handles) or block until executed. This is on by default and is exposed as an option for ease of debugging. in_flight_nodes_limit: Indicates the limit of in-flight nodes before enqueueing of async operations to DTensorDevice is blocked. This limit is per mesh. 0 for no limits from DTensor. Default is 8. c3RK|]}t|tj !ywN) isinstance layout_libMesh).0meshs X/home/dcms/DCMS/lib/python3.12/site-packages/tensorflow/dtensor/python/dtensor_device.py z)DTensorDevice.__init__..@s DTz$ 0 0 Ds%'z,Expected a flat list of Mesh objects, got {}z{}/device:CUSTOM:{}N)any TypeErrorformatr _next_device_number_lockhost_address_space_next_device_numbernamer Allocateregister_custom_device _device_info_current_output_layout_current_default_meshset_meshes threadingLock _mesh_lock_register_mesh)selfris_asyncin_flight_nodes_limitctxr device_infors r__init__zDTensorDevice.__init__0s  DV DD  8 ? ? G II // C !'..s/E/E/G/BDdiQ199 82FK ""6499kB#D"&D!%D5DLnn&DO  $ s 3DD&c0tj|}tj|j|}t j t jddg}t|}t|Dcgc] }|||zz }}|||fScc}w)zCReturns ID and device lists that can be used to create a host mesh.CPUr)job device_type device_index) npprodarangereshape tf_device DeviceSpecr full_job_namelenrange) r3shapehost_idnum_global_devicesglobal_device_idslocal_device_listnum_local_devicesxlocal_device_idss r_create_host_arrayz DTensorDevice._create_host_arrayVs "45==eD$$&E K-.167H1I,-G' '' .0A AAs=Brc|j5||jvrtj|j|j d|jj ||jjdk(rtjd|jj |j tj|j|jj d|jj |jdddy#1swYyxYw)z5Idempotently register `mesh` with the dtensor device.FTPUz7Registering virtual 1:1 mapped host mesh %s for mesh %sTN) r1r.r AddMeshr* to_stringaddr<upperlogginginfo host_mesh)r3rs rr2zDTensorDevice._register_meshds  - T\\ !&&   t~~/       # # % . ,,Gnn((*DNN,< > ! ( (!1!;!;!=t  ,,  4>>+ , - - -s D"D88Ereturnc|jSr)r.r3s rrzDTensorDevice.meshesus <<tensorsr c tjs td|j|jt j |j5tdDrtfdDs tdd}Dcgc]}|jc}Dcgc]}|jc}zDcgc]1}t j|jtj3c}znt!dDr tdd } t#j$tjj&|j)|j*|cd d d Scc}wcc}wcc}w#t,j.$r}t-j0|d d }~wwxYw#1swYy xYw) a Packs tensors into a DTensor handle on this DTensor device. Packing and unpacking are inverse operations: ``` * unpack(pack(tensors)) == tensors * pack(unpack(dtensor)) == dtensor ``` Refer to `dtensor.pack` for more information. Args: tensors: The list of tensors 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 not called eagerly. z`pack` must be called eagerly.c3PK|]}t|tj ywrrr SparseTensorrts rrz%DTensorDevice.pack..s H1Z=55 6 H$&c3VK|] }|jdjk("yw)rN)rG)rrcr]s rrz%DTensorDevice.pack..s$@1177gaj...@s&)z3All input SparseTensors to Pack must be same shape.T)dtypec3PK|]}t|tj ywrr`rbs rrz%DTensorDevice.pack..sJQz!]778Jrdz'Cannot Pack SparseTensors with Tensors.FN)r executing_eagerly RuntimeErrorr2rrr r'allr"indicesvaluesconvert_to_tensorrGrint64r!r Pack_handlerSr*r _NotOkStatusException_status_to_exception)r3r]r is_sparserces ` rpackzDTensorDevice.packysu,  $ $ & 9 :: $ DII 5 H H H@@@OP P &-.199.G1Lq!((1LLHOP CDC ! !!'' >P   J'J JABB 5%** OO  % %          55 /1LP  ' '5''*45'55sU8F9E9&F9,E>?F96F=%F9#A F9F9F6F11F66F99Gdtensorc tjs td tjtjj ||j }tjtjj ||j }|rpg}tt|dzD]R}|jtj||||t|dzz||dt|zdzzT|S|S#tj$r}tj|dd}~wwxYw)aUnpacks a DTensor handle on this DTensor device. Packing and unpacking are inverse operations: ``` * unpack(pack(tensors)) == tensors * pack(unpack(dtensor)) == dtensor ``` Refer to `dtensor.unpack` for more information. Args: dtensor: The DTensor to unpack. Returns: The raw underlying tensor components of the DTensor. Raises: RuntimeError: When not called eagerly. z `unpack` must be called eagerly.N)r rhrir Unpackrpr*r rqrrIsSparseDTensorrFrEappendrra)r3rvr]rtrsresultis runpackzDTensorDevice.unpacks.*  $ $ & ; <<3&-- //  # #    g'66!! IfS\Q&'L!  & &wqz'.q3w<13D/D'E'.q1s7|3Cq3H/H'I K LL m n!  % %3  % %a (d23s=DD="D88D=ctjs tdtj|r|j } t jtjj||j}|ytjj|S#tj$r}tj|dd}~wwxYw)zFetches the layout of the DTensor. Args: dtensor: The DTensor whose layout is to be fetched. Returns: The `Layout` of this DTensor. Raises: RuntimeError: When not called eagerly. z&`fetch_layout` must be called eagerly.N)r rhrir IsVariable read_valuer FetchLayoutrpr*r rqrrrLayout from_string)r3rv layout_stringrts r fetch_layoutzDTensorDevice.fetch_layouts  $ $ & A BB(""$g3,88 //  # #    m     ( ( 77  % %3  % %a (d23s=B%%C8CCtensorc(tjs tdtj|syt j |r |j}tjtjj||jS)a~Check whether the input tensor is a DTensor. In Python, a DTensor has the same type as a `tf.Tensor`. This method will let you check and handle the tensor differently if a tf.Tensor is a DTensor. Args: tensor: an object to be checked. Returns: bool, True if the given tensor is a DTensor. Raises: RuntimeError: When not called eagerly. z$`is_dtensor` must be called eagerly.F) r rhrir is_tensorrrrpr IsDTensorr*)r3rs r is_dtensorzDTensorDevice.is_dtensorss  $ $ & ? @@   ( '~~f ! + +!!  r\cFtj|j||y)zSets the singleton global device ID-to-physical core ID map. Args: mesh_name: The name of a mesh. If empty, set the default mapping. tpu_core_ids: TPU core IDs sorted by TF task/device ordinal. N)r SetTPUCoreIDsr*)r3 mesh_name tpu_core_idss rset_tpu_core_idszDTensorDevice.set_tpu_core_ids s(():):I)57r\cBtj|jyr)r ClearTPUCoreIDsr*r[s rclear_tpu_core_idsz DTensorDevice.clear_tpu_core_idss**4+<+<=r\c|tjtjj|j|S)zTranslates TPU core IDs to TPU core locations. Args: tpu_core_ids: A list of TPU core IDs. Each one is an unsigned integer. Returns: A list of corresponding TPU core locations. )r TPUCoreIDsToLocationsr rpr*)r3rs rtpu_core_ids_to_locationsz'DTensorDevice.tpu_core_ids_to_locationss3 " 7 7!!  r\c|tjtjj|j|S)zTranslates TPU core locations to TPU core IDs. Args: tpu_core_locations: A list of TPU core locations. Each one is a list of four unsigned integers, [x, y, z, core]. Returns: A list of corresponding TPU core IDs. )r TPUCoreLocationsToIDsr rpr*)r3tpu_core_locationss rtpu_core_locations_to_idsz'DTensorDevice.tpu_core_locations_to_ids(s3 " 7 7!!  r\cztjtjj|jS)zReturns the number of cache hit and miss for function compilation. Returns: A dictionary. 'miss': number of cache misses; 'hit': number of cache hits; and 'size': size of cache; miss count. )r GetStatsr rpr*r[s r _get_statszDTensorDevice._get_stats7s0 " * *!!  r\layoutsc tjtjj||Dcgc]}|j c}|j ycc}w)a6Sets the element layouts on an iterator resource tensor. Args: iterator_resource_dtensor: a DTensor created by packing the individiual iterator resource tensors. layouts: the flattened list of layouts to be applied to the elements emitted by the iterator resource DTensor. N)r SetIteratorElementLayoutsr rprSr*)r3iterator_resource_dtensorrr s rset_iterator_element_layoutsz*DTensorDevice.set_iterator_element_layoutsFsJ44!!!*12   2   3sA c#K|j}|j|tj|j|j j d||_dtj|j|r=tj|j|j j d||_yw)aDSets a default mesh for all ops in the scope. Note: This is an internal helper method, which is not user facing api. Useful for requesting a specific mesh for ops which would have no inferred layout, e.g. tf.zeros. Args: mesh: A Mesh to be used for ops without Mesh. Yields: Nothing. utf-8N)r,r2r ExperimentalSetDefaultMeshr*rSencodeExperimentalClearDefaultMesh)r3rprevious_defaults r_experimental_default_meshz(DTensorDevice._experimental_default_meshVs1155  (*"&D 778I8IJ77     $ $ & - -g 68"2DsC Cc #Kd}d}d}|j|j |j}|jj d|_t j |j|jtjr-tj|j5ddddn1tj}t|j}d||j|dD]}|j!dt#j$t"j$j'|jg|j!dt#j$|jjj d||_|j t j(|jyt j |j|jj+dy#1swY-xYw#||j|dD]}|j!dt#j$t"j$j'|jg|j!dt#j$|jjj d||_|j t j(|jwt j |j|jj+dwxYww)aSets a default output layout for all ops in the scope. Note: This is an internal helper method, which is not user facing api. Useful for requesting a specific layout for ops which would have no inferred layout, e.g. tf.zeros. Caveats: - Currently only affects the first output of an op. For Op with multiple outputs, this does not support yet. - All Ops in the scope will be attached with the same layout. This might not be valid as the rank is different. The current suggestion is: Try to wrap the raw op wheneven possible. Args: layout: A Layout for the outputs of all operations in this scope. Yields: Nothing. Nr_layout)s)list_mesh)r2rr+rSrr ExperimentalSetDefaultLayoutr*r rhrr r'get_default_graphrEget_operations _set_attrr AttrValue ListValueExperimentalClearDefaultLayoutdecode)r3r rprevious_graph_sizegraph operations r_default_layoutzDTensorDevice._default_layoutss0 E $%L44$*$4$4$6$=$=g$Fd!99   T88:  " " $ ZZ "    %%'!%"6"6"89  --/0C0DE >I   &&%//99445:78 9   &&KK))+227;= > >%5d! $ $ ,==d>O>OP;;   t::AA'J L=    --/0C0DE >I   &&%//99445:78 9   &&KK))+227;= > >%5d! $ $ ,==d>O>OP;;   t::AA'J Ls7"MB H!2H7:H!1D#MHH!!D$MMN)T)__name__ __module__ __qualname____doc__rrrr8rOr2propertyrrrrrrurrboolrrrrrrr contextlibcontextmanagerrrr\rrr-sJI%&$ JOO,$ L B--" c*//* -5(3--51B1B-5s-5^,C,HSM,\8#8**;*;88st67>   ,01B1B,C 2Z__228ALJ$5$5ALALr\r)$rrrVr/typingrrrrnumpyr>tensorflow.core.frameworkrtensorflow.dtensor.pythonrr rtensorflow.pythonr tensorflow.python.eagerr r tensorflow.python.frameworkr rBrrrrtensorflow.python.utilrr&r0r$objectrrr\rrshE++4,:4+(;.+530 )9>>+HLFHLr\