o X۷i~2@s4ddlmZddlmZddlZddlZddlZddlmZddlm Z ddlm Z e e e e fZ e e e fZeeee e ffZee eee fZe e dfZejGdd d Zd@ddZdAddZdBddZdCddZdDd!d"ZdEd)d*ZdFd+d,ZdFd-d.ZdFd/d0ZdFd1d2Z dGdHd7d8Z dId>d?Z!dS)J) annotations)CallableN)_array)_chunk)_modes.c@s&eZdZUded<ded<ded<dS) _Blocking list[int] i_partitions j_partitions k_partitionsN)__name__ __module__ __qualname____annotations__rrU/home/ubuntu/vllm_env/lib/python3.10/site-packages/cupyx/distributed/array/_linalg.pyr!s   rlocation_map_a_BlockLocationMaplocation_map_breturnc Csg}g}g}dd}|D]\}}||||||q|D]\}}||||||q!dd} | |}| |}| |}dd} |D]\}}| ||| ||qH|D]\}}| ||| ||q[t|||S)NcSs2|\}}}|dkr td||||dS)Nz"Step other than 1 is not supported) RuntimeErrorappend)indices partitionsstartstopsteprrradd_to_partitions8s  z)_find_blocking..add_to_partitionscSsXt|dkr td||dg}t||ddD] \}}||kr)||q|S)NrzArray has no chunkr)lenrsortzipr)rresxyrrrto_unique_sortedIs   z(_find_blocking..to_unique_sortedcSs.|\}}}||d||krtddS)NrzInconsistent index mapping)indexr)rrrr_rrr check_indicesZs z%_find_blocking..check_indices)keysr) rrr r r r i_indices k_indices j_indicesr%r(rrr_find_blocking/s,          r-blocking_ExecutionPlancCs|j}|j}|j}g}t||ddD]V}t||ddD]J}t||ddD]>} |d| df} | d|df} t|| } t|| } | | @}|r_|}|| | |fq*td| d| qq|S)NrrzAThere is no device that can perform multiplication between block z and ) r r r r!setr)poprr)r.rrr r r plani_rangej_rangek_rangeblock_ablock_b devices_a devices_b intersectiondevrrr_make_execution_planjs2r=slicestuple[slice, ...]shapetuple[int, ...]tuple[_SliceIndices, ...]cCs,t|t|ks Jtddt||DS)Ncss|] \}}||VqdSN)r).0slengthrrr sz%_convert_to_tuples..)rtupler!)r>r@rrr_convert_to_tuplessrItuplescCstdd|DS)Ncss|]}t|VqdSrCslice)rDtrrrrGsz%_convert_to_slices..)rH)rJrrr_convert_to_slicessrN index_map"dict[int, list[tuple[slice, ...]]]"dict[_BatchIdx, _BlockLocationMap]c Cs|i}|D]5\}}t|D],\}}t||}|dd|dd}} tt| } ||i} | | i} || |<qq|S)N)items enumeraterItypingcast _BlockIdx setdefault) r@rO location_mapsr<idxschunk_iidx idx_tuples batch_idx block_idx location_maplocationrrr_group_by_batchs      rbarr_array.DistributedArrayf_shape,Callable[[tuple[int, ...]], tuple[int, ...]]f_idx0Callable[[tuple[slice, ...]], tuple[slice, ...]]cs`dfdd i}|jD]\}}fdd|D||<q|j}t||j||j|jS) Nchunk _chunk._ChunkrcsH|j|jj}|j}fdd|jD}t||j|||jS)Ncsg|] \}}||fqSrr)rDdatar\)rgrr sz>_reshape_array_with..reshape_chunk..) arrayreshaper@r&updatesr_Chunkready prevent_gc)rirkr&ro)rgrerr reshape_chunks  z*_reshape_array_with..reshape_chunkcsg|]}|qSrr)rDri)rsrrrlsz'_reshape_array_with..)rirjrrj) _chunks_maprSr@rDistributedArraydtype_mode_comms)rcrerg chunks_mapr<chunksr@r)rgrersr_reshape_array_withs r{cCt|ddddS)NcSsd|SNr0rr@rrrz'_prepend_one_to_shape..cSstdf|SrCrKr\rrrrr{rcrrr_prepend_one_to_shape rcCr|)NcSs|dSr}rr~rrrrrz&_append_one_to_shape..cSs|tdfSrCrKrrrrrrrrrrr_append_one_to_shaperrcC&|jddks Jt|ddddS)NrcS |ddSNrrr~rrrr z!_pop_from_shape..cSrrrrrrrrrr@r{rrrr_pop_from_shape rcCr)NrrcS |ddSNrrr~rrrrrz'_pop_front_from_shape..cSrrrrrrrrrrrrrr_pop_front_from_shaperrabout_array.DistributedArray | Nonec" Ks|durtddD]}||vrtd|dq t|tjr%t|tjs)td|tj}|tj}d}}|jdkrDd }t|}|jdkrOd }t |}|j d d\}}|j d d\} } || kss|j dd |j dd krwt d t |j |j } t |j |j } | | krtd d d|jD} d}| D]}| |}| |}t||}t|||}t|}|D]y\}}}||}||}|j|||}|j|||}|tj|tj|t|dt|df}|3}||jtjjj|j|jfi|}tj|| |||fd}| |!||j"}Wdn 1s.wYqq|j dd || f} t| || tj#|j$}!|rRt%|!}!|rYt&|!}!|!S)akMatrix multiplication between distributed arrays. The arguments must have compatible :attr:`~DistributedArray.shape` and :attr:`~DistributedArray.index_map`. This operation converts its operands into the replica mode, and compute their product in the sum mode. Args: a, b: Input distributed arrays. out (optional): A location into which the result is stored. This option is currently not supported. Returns: The matrix product of the inputs. Example: >>> A = distributed_array( ... cupy.arange(6).reshape(2, 3), ... make_2d_index_map([0, 2], [0, 1, 3], ... [[{0}, {1, 2}]])) >>> B = distributed_array( ... cupy.arange(12).reshape(3, 4), ... make_2d_index_map([0, 1, 3], [0, 2, 4], ... [[{0}, {0}], ... [{1}, {2}]])) >>> C = A @ B >>> C.mode 'sum' >>> C.all_chunks() {0: [array([[0, 0], [0, 3]]), array([[0, 0], [6, 9]])], 1: [array([[20, 23], [56, 65]])], 2: [array([[26, 29], [74, 83]])]} >>> C array([[20, 23, 26, 29], [56, 68, 80, 92]]) .. seealso:: :obj:`numpy.matmul` NzArgument `out` is not supported)subokaxesaxisz Argument `z` is not supportedzHMixing a distributed array with a non-distributed array is not supportedFrTrRzShapes are incompatiblezMismatched batch shapescSsi|]}|gqSrr)rDr<rrr .szmatmul..r)rr)'r isinstancerru _to_op_moderREPLICAndimrrr@ ValueErrorrbrOr)devicesr-r=rNrtflushrLon_ready wait_eventrqcupylinalg_productmatmulrmrrprecordrrvSUMrxrr)"rrrkwargsparam one_prepended one_appendednmm2plocation_maps_alocation_maps_bryrvr^rrr.r3 index_prefixr7r8r<loc_aloc_bchunk_achunk_br&streamchunk_ab_arraychunk_abr@r"rrrrs/      $        rr rr rlist[list[set[int]]]c Cs |ddksJtt||ksJ|ddksJtt||ks$Ji}t|t|dks2Jtt|D]K}t||t|dksHJtt||D]2}||}||d}||}||d} t||t|| f} |||D] } || g| qvqPq8|S)aCreate an ``index_map`` for a 2D matrix with a specified blocking. Args: i_partitions (list of ints): boundaries of blocks on the `i` axis j_partitions (list of ints): boundaries of blocks on the `j` axis devices (2D list of sets of ints): devices owning each block Returns: dict from int to array indices: index_map Indices for the chunks that devices with designated IDs are going to own. Example: >>> index_map = make_2d_index_map( ... [0, 2, 4], [0, 3, 5], ... [[{0}, {1}], ... [{2}, {0, 1}]]) >>> pprint(index_map) {0: [(slice(0, 2, None), slice(0, 3, None)), (slice(2, 4, None), slice(3, 5, None))], 1: [(slice(0, 2, None), slice(3, 5, None)), (slice(2, 4, None), slice(3, 5, None))], 2: [(slice(2, 4, None), slice(0, 3, None))]} rr)sortedr1rrangerLrXr) r r rrOiji_starti_stopj_startj_stopr\r<rrrmake_2d_index_mapZs&   r)rrrrrr)r.rrrrrrr/)r>r?r@rArrB)rJrBrr?)r@rArOrPrrQ)rcrdrerfrgrhrrd)rrdrC)rrdrrdrrrrd)r rr rrrrrP)" __future__rcollections.abcr dataclassesrUrcupyx.distributed.arrayrrrrHint _SliceIndicesrWdictrlistr/ _BatchIdx dataclassrr-r=rIrNrbr{rrrrrrrrrrs8       ; #        {