Pi#ddlZddlmcmZddlZddlmZddlm Z m Z m Z ddl mZ ejdddZejddZejd dZejd Zejd d ZdZejdddZejddZejjejd ddedejde de ejedfdede j dejfdZ!ee!_dS) N)partial)CallableUnionAny PartitionSpecPz jax.sharding)from_ NamedShardingMeshjaxnumpynp)as_cl|}|/t|tjjjrdS|j}|/dS)NTF) isinstanceeinxexprstage3 Compositionparent)rnodes n/home/jaya/work/projects/VOICE-AGENT/VIET/agent-env/lib/python3.11/site-packages/einx/experimental/op/shard.py _is_composedrsB D   dDI,8 9 9 4{   5cdfd S)Nc0|||SN)expr_in tensor_inexpr_outbackendcts rz..s&!! ) KKrrrr$r#s``rr%r%s$r)tracecddl}||fD]}|D]}t|tjjjrtdt|tjjjrw|}|j nt|j tjjj r7t|j r#||j j durtd|j }|j nьtj ||j|}||g\}t#j|g|g|\\}\}t'd|D} |t(t,t'd| D} t3| t'd| D}nt||jjrbt9d | D} t9d |jD} | | std | d | nlt(|t'd | D} t3| t'd| D}t'd|D} d| D}t?|}tA||}t,!||}t#j"|g|g|g|\}||fS)NrzConcatenation not allowedzhIf device axes are used within a composition they must appear as the left-most member of the composition)r"c3K|]N}t|tjjj&tjj|J|VOdSr)rrrrAxis is_marked.0axiss r zshard_stage3..:sl  dDI,1 2 28.Bs$9W9Wa!'9W9W9W9W9W9Wrc3$K|] }|jV dSrnamer3s rr/zshard_stage3..C$.K.K!qv.K.K.K.K.K.Kr) axis_namesc3$K|] }|jV dSrr6r3s rr/zshard_stage3..Fs$77a16777777rc34K|]}t|VdSr)strr3s rr/zshard_stage3..Gs(99AQ999999rz9Marked axes must be subset of mesh axes. Got marked axes z and mesh axes c3$K|] }|jV dSrr1r3s rr/zshard_stage3..Ns$/M/MA/M/M/M/M/M/Mrc3$K|] }|jV dSrr6r3s rr/zshard_stage3..Or8rc3bK|]*}t|tjjj&|V+dSr)rrrrr*r,s rr/zshard_stage3..Rs9UU$Zdi>N>S-T-TUUUUUUUrchg|]/}tjj|r|jnd0Sr)rrrr+r7r,s r z shard_stage3..Ss7___RV49#3#=#=d#C#CMdii___r)#r allrrrr Concatenation ValueErrorMarkerrListrchildrentracer call_factoryshape all_to_tensorutilflattentupletnparraytjaxdevicesreshapetMeshshardingr setr9issubsettPtNamedSharding device_put unflatten)rr r!meshr"r rootrchild marked_axesrR marked_names mesh_namesaxespartition_specrUs r shard_stage3rdstJJJ(#))HHJJ ) )D$ 0 >?? > !<===$ 0 788 )l."5<1A1FGG(66")>q)AAA(U"LEl.  )$ ((GM7(SSI(()55LY $|WI {GTTTJW K  |))DLLNN++33E9W9W;9W9W9W4W4WXXW.K.K{.K.K.K)K)KLLL D#,+ , , M77;77777 9999999 $$Z00 uLuuisuu   ))D//))%/M/M/M/M/M*M*MNNW.K.K{.K.K.K)K)KLLL UU'UUU U UD__Z^___N(NdN33H 844I>7)i[8*gVVVLY g rTc  ddl}tjj|\}tjj| t dkrtdt  fd}| |g}n/#tjj $r} tjj tjj dd} tj | t|g} || g}n#tjj $r| wxYwYd} ~ n[d} ~ wwxYwt||jjrtjjtjjjd|jD} tj | |jj} || g}nt|t.t0frtjj tjj dd} tj | t|g} || g}|} || fS)Nrz%Expected exactly one expression, got ctjtjddg|zdDzdS)Nrcg|]L\}}tj|tj|dtjfddMS).N)depth1depth2)rrEquationrasarraynewaxis)r-kvs rrAz(parse..solve..qsYAq ""1bjmmCO&DTZ^"__r)cse)rrsolverkitems)eqsrpop parameters tensor_shapes rrqzparse..solvemsy Y  1a, 7 7 8 &,,..       rcVg|]&}tjj|'Sr)rrstage1 NamedAxis)r-r7s rrAzparse..s8)))59  **400)))r)r rrtrL!_clean_description_and_parametersrrxparse_oplenrDSolveExceptionr get_markedrkrRrrUr rErFmayber9rJlistrN __deepcopy__) descriptionrvrpr\ jax_devicesrur rqre expr_meshmesh_eqr!rts `` ` @rparser`sJJJ"glLLZK   " "; / /B 2ww!||JRJJKKK          | eBiiGGy'    (44TY5E5P5PQSTUQVWXQY5Z5Z[[Ii((S5G5G4HIIG % **9+       D#,+ , ,#I$++ I  ! ' '))=A_)))    )$$Y 0BCC% "" D4- ( (#I$001A1L1LRPQUSTX1V1VWW )$$YT <<% ""##%%H H s1 BE83BE3 EE3E**E33E8cdfd S)Nc.||fd|i|S)Nr\r)rtensorr\r"kwargsr#r$s rr%z..s<VWVWQQvYYWW%)W-3WWrNNrr&s``rr%r%s$rrrr\r"rprureturnc |jdkrtdt|tj|f||d|\}}t |||||\}}|S)a Shards a tensor over a mesh of devices. *This function is currently experimental and will likely change in future versions.* *This function is currently only supported for Jax: A sharding is created based on the given expression, and applied to the tensor using* ``jax.device_put``. The tensor is sharded across the marked axes in the input expression. The marked axes match the axis names and shape of the mesh: >>> x = jnp.ones((2, 4, 128)) >>> x = einx.experimental.shard("[d1 d2] c") >>> x.sharding NamedSharding(mesh=Mesh('d1': 2, 'd2': 4), spec=PartitionSpec('d1', 'd2', None)) Axis compositions can be used to apply the `sharding rules of Jax `_, where tensor axes are evenly divided by the number of shards: >>> x = jnp.ones((128, 640, 480, 3)) >>> x = einx.experimental.shard("([batch] _) ...", x) >>> x.sharding NamedSharding(mesh=Mesh('batch': 8), spec=PartitionSpec('batch',)) If possible, the sharding is created over all devices. ``_`` is a regular axis name, and its value is determined by :doc:`einx's expression solver `. Optionally, an existing mesh can be passed: >>> from jax.sharding import Mesh >>> devices = np.asarray(jax.devices()).reshape(4, 2) >>> mesh = Mesh(devices, axis_names=("d1", "d2")) >>> x = jnp.ones((4, 1024, 1024)) >>> x = einx.experimental.shard("a ([d2] b) ([d1] c)", x, mesh=mesh) >>> x.sharding NamedSharding(mesh=Mesh('d1': 4, 'd2': 2), spec=PartitionSpec(None, 'd2', 'd1')) The array is replicated over all mesh axes that are not part of the expression: >>> x = jnp.ones((1024, 1024)) >>> x = einx.experimental.shard("a ([d1] b)", x, mesh=mesh) >>> x.sharding NamedSharding(mesh=Mesh('d1': 4, 'd2': 2), spec=PartitionSpec(None, 'd1',)) Args: description: Description string in Einstein notation (see above). tensor: Input tensor or tensor factory matching the description string. mesh: Mesh or list of devices to shard the tensor over. If not given, a new mesh over all available devices will be created matching the axes in the given expression. Defaults to ``None``. cse: Whether to apply common subexpression elimination to the expressions. Defaults to True. graph: Whether to return the graph representation of the operation instead of computing the result. Defaults to False. **parameters: Additional parameters that specify values for single axes, e.g. ``a=4``. Returns: The sharded tensor if ``graph=False``, otherwise the graph representation of the operation. r z;einx.experimental.shard is currently only supported for Jax)r\rp)r\r")r7NotImplementedErrorrrrH get_shaperd)rrr\r"rprurr!s rshardrsT|u!"_```T[**6229=3JTGX$GVXDRYZZZFH Mrr)TNN)Nr T)"r einx.op.utilrtrLr r functoolsrtypingrrr numpy.typingnptrHimport_rXrYrTrQrOrjitrd lru_cachertraceback_utilfilterr<TensorBackendbool ArrayLikerrrrrs: '''''''''' [#^DD$$_N$KK F.99 {5!! k't,,   >>>>B6666r   .3 JJJ KJ J4<d* + J  J - J [JJJ  JZ r