ΑiE<SSKJr SSKJr SSKrSSKrSSKJr SSK J r J r SSK J r SSKJr SS KJr SS KJrJrJr SS jrSS jrS qSrSr0r"SS5rSrSSjrSSjrSq SSjr!SSjr"SSjr#g)) annotations)reduceN)core) ProcessMeshget_current_process_mesh)get_default_distributed_context)DistributedOperatorHelper)DistributedTensor)__no_shape_var_type__convert_to_dims_mappingverify_shard_specc Ub+[U[R5(d SUS35eO[5nUcS5e[U[5(d SUS35e[U[ 5(aG[ RR5R5RU5n[U5nO [U5nURnXRlUR[ ;a/nO UR"nUb[%US5(aUR'5O UR(n[+U5H9upxUcM [U[ 5(aX;dM&[-SUSUS US 35e [/X%U5(dS UR0S US USUS 3 5e[3X!5URlUbURR7S5 UbURR7S5 [95n U R;U5 U R=U5nU R?U5 U$)ap Shard a tensor on a process mesh according to the shard specification. Args: x (Tensor): the tensor to be sharded. process_mesh (ProcessMesh, optional): An instance of ProcessMesh describes a mesh topology of the used logical processes where the tensor is sharded. If it is None, the found current process mesh will be used. And an error will be raised if the current process mesh cannot be found. Default: None. shard_spec (list, optional): a list to describe the sharding mapping between `x` and `process_mesh`, which means the dimension `i` of `x` is split across the dimension `shard_spec[i]` of `process_mesh`, where `None` means that tensor dimension is not split. For example, given a tensor with the shape [6, 12] and a process mesh with the shape [2, 3] and the dimension names ["x", "y"]: If `shard_spec=["x", "y"]`, each shard of the tensor will have a shape [3, 4]; If `shard_spec=["y", "x"]`, each shard of the tensor will have a shape [2, 6]; If `shard_spec=["x", None]`, each shard of the tensor will have a shape [3, 12]; If `shard_spec=[None, "x"]`, each shard of the tensor will have a shape [6, 4]; If `shard_spec=["y", None]`, each shard of the tensor will have a shape [2, 12]; If `shard_spec=[None, "y"]`, each shard of the tensor will have a shape [6, 4]; If `shard_spec=[None, None]`, each shard of the tensor will have a shape [6, 12]; If the `shard_spec` is None, the tensor will be replicated across all the processes of `process_mesh`. In the above example, the `shard_spec=None` is same as 'shard_spec=[None, None]'. Defaults: None. Returns: Tensor: the tensor `x` annotated with sharding information. Examples: .. code-block:: python >>> # doctest: +REQUIRES(env:DISTRIBUTED) >>> import paddle >>> from paddle.distributed.fleet import auto >>> mesh = auto.ProcessMesh([[0, 1], [2, 3]], dim_names=["x", "y"]) >>> x = paddle.ones([4, 6]) >>> shard_spec = ["x", "y"] >>> auto.shard_tensor(x, mesh, shard_spec) Argument process_mesh " is not an instance of ProcessMeshKSpecify the process mesh argument or use ProcessMesh context manager first.zArgument shard_spec z is not an instance of list get_dim_nameszInvalid shard_spec at index z: 'z0' is not a valid dimension name in process_mesh .z For tensor z , shard_spec z is invalid with tensor_shape z and process_mesh process_mesh dims_mapping) isinstancerrrliststrpaddlestaticdefault_main_program global_block_var_recursiver serial_tensor dist_attrrtyper shapehasattrr dim_names enumerate ValueErrorrnamer rmark_annotatedr add_dist_tensor_for_programget_dist_tensor_for_programadd_process_mesh) xr shard_spec dist_tensorr tensor_shape valid_dimsidimdefault_dist_ctxs j/var/www/html/banglarbhumi/venv/lib/python3.13/site-packages/paddle/distributed/auto_parallel/interface.py shard_tensorr5"soR,(8(899 $\N2T U 901 ' Y ' j$ ' ' zl*EF '!S MM . . 0 \^ ^A  (* '* --M)5&22 $** |_55  & & (''   +FAsC((C,A 21#S>EEOLPQS ,!<HH -,,-]:,FdeqdrsEFRESST U H.E . *,,^<,,^<6800=">>qAK%%l3 Hc Ub![U[5(d SUS35eO[5nUcS5e/nUb[[SU55(d SUS35eUH3nUbUR [ Xa55 M"UR S5 M5 /nUb[[SU55(d S US35eUH3nUbUR [ Xa55 M"UR S5 M5 [ XXWU5nU$) a Shard an operation on a process mesh according to its input and output shard specification. Args: op (Callable): a callable operator or module to be sharded. process_mesh (ProcessMesh, optional): An instance of ProcessMesh describes a mesh topology of the used logical processes where the op is sharded. All of its inputs and outputs are sharded by this process mesh. If it is None, the found current process mesh will be used. And an error will be raised if the current process mesh cannot be found. Default: None. in_shard_specs (list of list, optional): a list of list to describe the sharding specifications for the inputs. Each item of `in_shard_specs` is a `shard_spec` between the corresponding input and `process_mesh`. If one item is None, the corresponding input is replicated across all processes If it is None, all inputs are replicated across all processes. Note that the length of the `in_shard_specs` should be equal to the actual number of inputs when calling this operation. Default: None. out_shard_specs (list of list, optional): a list of list to describe the sharding specifications for the outputs. Each item of `out_shard_specs` is a `shard_spec` between the corresponding output and `process_mesh`. If one item is None, the corresponding output is replicated across all processes If it is None, all outputs are replicated across all processes. Note that the length of the `in_shard_specs` should be equal to the actual number of inputs when calling this operation. Default: None. Default: None. Returns: Outputs of `op`, each of which is annotated with sharding information. Examples: .. code-block:: python >>> import paddle >>> from paddle.distributed.fleet import auto >>> x = paddle.ones([4, 6]) >>> y = paddle.zeros([4, 6]) >>> mesh = auto.ProcessMesh([[0, 1], [2, 3]], dim_names=["x", "y"]) >>> dist_add = auto.shard_op(paddle.add, ... mesh, ... in_shard_specs=[["x", "y"], ["y", None]], ... out_shard_specs=[[None, "x"]]) >>> dist_add(x, y) Nrrrc3X# UH n[U[5=(d USLv M" g7fNrr.0r-s r4 shard_op..s+ ,  D ) ?Z4-? ?,(*zin_shard_spec z is not a list of list or Nonec3X# UH n[U[5=(d USLv M" g7fr9r:r;s r4r=r>s+ -  D ) ?Z4-? ?-r?zout_shard_spec )rrrallappendr r )oprin_shard_specsout_shard_specskwargsin_dims_mappingsr-out_dims_mappingss r4shard_oprIso\, 44 $\N2T U 401 ' Y '! ,    KN++I J K )J% ''+JE!''- )" -    M_--K L M *J%!((+JE"((. * # *v B Ir6c8[S- q"SS5nU"U5$)Nrc \rSrSrSrSrSrg)$recompute..RecomputeOperatorcXlgr9_op)selfrCs r4__init__-recompute..RecomputeOperator.__init__sHr6cX[RR5R5n[ UR 5n[RR S[35 [RRRR5(aD[RRRRUR5"U0UD6nOUR"U0UD6nSSS5 [R R#5(ax[RR5R5n[ UR 5n[%XF5H(nUR UnUR'S[5 M* W$!,(df  N=f)Nz/auto_parallel/rc_fwd_recompute_id)rrrrlenops name_scope_g_recompute_idxbasedygraphin_to_static_modejit dy2staticconvert_call_func convert_callrQ framework in_pir_moderange set_int_attr) rRargsrFblock rc_begin_idoutput rc_end_ididxrc_ops r4__call__-recompute..RecomputeOperator.__call__s<MM668EEGEeii.K))$%5$67;;&&++==?? ,,>>KK HH+#)+ "XXt6v6F++-- ::<IIK N  8C!IIcNE&&'9;KL9M's (BF F)rPN__name__ __module__ __qualname____firstlineno__rSrm__static_attributes__r6r4RecomputeOperatorrMs   r6rv)rZ)rCrvs r4 recomputerws$: R  r6c&"SS5nU"U5$)z Exclude some operators in recompute segments. Args: run_function (callable): The callable function to be excluded. Returns: ExcludeOperator: The callable object. c \rSrSrSrSrSrg)1exclude_ops_in_recompute..ExcludeOperatoricXlgr9 _run_function)rR run_functions r4rS:exclude_ops_in_recompute..ExcludeOperator.__init__s!- r6c[RRS5 [RRRR 5(aD[R RRRUR5"U0UD6nOUR"U0UD6nSSS5 U$!,(df  W$=f)Nz /exclude_rc) rrrYr[r\r]r^r_r`rar})rRrfrFris r4rm:exclude_ops_in_recompute..ExcludeOperator.__call__s))-8;;&&++==?? ,,>>KK ..+#)+ "//@@F9M98Ms BB88 Cr|Nrorur6r4ExcludeOperatorrzs  . r6rru)r~rs r4exclude_ops_in_recomputers" < ((r6c\rSrSrSrSrSrg)CollectionNamesi%fetchesloggingruN)rprqrrrsFETCHESLOGGINGrtrur6r4rr%s GGr6rc\[RUS5nUc /nU[U'[U$r9)_g_collectionsget)r' collections r4get_collectionr*s4##D$/J )t $ r6cU[;a /[U'Ub3[UH up4XA:XdM g [URX!45 g[UH up4XA:XdM g [URSU45 gr9)rrB)collection_namevaluer'_vs r4add_to_collectionr2s|n,*,' "?3DAz4 '..}="?3DAz4 '..e}=r6cH[U[RR5(a URnO0[U[ 5(aUnO[ S[U5S35e[[RX5 U(a[[RX5 gg)NzDOnly support fetch `Variable` or `str`[`Variable`'s name], but got ``) rrrVariabler'r TypeErrorr!rrrr)tensorr'rs r4fetchrAs&&--0011 FC RSWX^S_R``a b  o--v</116@r6c[$)ar Get the global mesh set by set_mesh. Returns: mesh (paddle.distributed.ProcessMesh): the global mesh. Examples: .. code-block:: python >>> import paddle >>> import paddle.distributed as dist >>> mesh = dist.ProcessMesh([[[0, 1], [2, 3]], [[4, 5], [6, 7]]], dim_names=["dp", "mp", "pp"]) >>> # doctest: +REQUIRES(env:DISTRIBUTED) >>> dist.auto_parallel.set_mesh(mesh) >>> mesh = dist.auto_parallel.get_mesh() >>> # This case need to be executed in multi-card environment >>> # python -m paddle.distributed.launch --gpus=0,1,2,3,4,5,6,7 {test_case}.py _g_meshrur6r4get_meshrRs ( Nr6cUqg)aG Set the global mesh. Args: mesh (paddle.distributed.ProcessMesh): global mesh to be set. Returns: None Examples: .. code-block:: python >>> import paddle >>> import paddle.distributed as dist >>> mesh = dist.ProcessMesh([[[0, 1], [2, 3]], [[4, 5], [6, 7]]], dim_names=["dp", "mp", "pp"]) >>> # doctest: +REQUIRES(env:DISTRIBUTED) >>> dist.auto_parallel.set_mesh(mesh) >>> # This case need to be executed in multi-card environment >>> # python -m paddle.distributed.launch --gpus=0,1,2,3,4,5,6,7 {test_case}.py Nr)meshs r4set_meshris ,Gr6cUVs/sHoSPM nnUVs/sHoSPM nn[R"S[SUS55RU5n[ XB5q[ $s snfs snf)z Create a global process_mesh for auto parallel. Args: mesh_dims (list[tuple[str, int]]): A list of tuple, each element is (dim_name, dim_degree). rrc X-$r9ru)r,ys r4create_mesh..sr6)nparangerreshaperr) mesh_dimsmesh_dimr$ mesh_shapemesh_arrs r4 create_meshrsq.77Y!YI7.78i(1+iJ8yyF#5z1EFNNH(.G N 88s A,A1)NN)NNNr9)NF)returnpaddle.distributed.ProcessMesh)rrrNone)rzlist[tuple[str, int]])$ __future__r functoolsrnumpyrrpaddle.frameworkrrrrstatic.dist_contextr static.dist_opr static.dist_tensorr static.utilsr r rr5rIrZrwrrrrrrrrrrrur6r4rs# !?@51` HAETn!!H)>   > A .2r6