o X۷i@sddlmZddlZddlZddlZddlmZddlmZddlm Z e Z da a addZedd Zd d Zd d ZddZdS)) annotationsN)_util) _accelerator)HandlecCshtdurdSz#ddlattdrddlmaddlmaWdSddlmataWdSty3YdSw)Nrbindings) cutensornet) tensornet) cuquantumhasattrcuquantum.bindingsrr ImportErrorr r N/home/ubuntu/vllm_env/lib/python3.10/site-packages/cupy/linalg/_einsum_cutn.py_maybe_lazy_load_cutensornets     rcCsBddtjdD}|ddks|ddkr|ddkrdSd S) NcSsg|]}t|qSr )int.0ir r r (sz-_is_cuqnt_22_11_or_higher...r TF)r __version__split)verr r r_is_cuqnt_22_11_or_higher&s$rcCstS)N)rr r r r_is_nonblocking_supported.srcCst|dkr tdt|dtr!|d}t|dd}||fSt|}g}g}d}t|dkrG||d||dt|dks1t|dkrR|d}|rVJ|||fS)znParse & retrieve einsum operands, assuming ``args`` is in either "subscript" or "interleaved" format. rzmust specify the einstein sum subscripts string and at least one operand, or at least one operand and its corresponding subscripts listrN)len ValueError isinstancestrlistappendpop)argsexproperandsinputsoutputr r r_get_einsum_operands2s(     r+c stjjjrdStjtvrdSttdur t j ddddS| dd}| dd}|dur2d}t |}d d |d D}t |d krGdStd d |DsYtdd |Dr[dS|durdtj|n|tjtjtjtjfvrtdSfdd |D}tjj}ttdsi}t_ntj}| |}|durt}t|tj||<n|j}||d}trd|d<d} |durdd tt |d dD}n5t |r|ddkr|d d}n$t |dkrt|d tt frd} |ddkrd} d}n|dkrd} d}| r t j ddd|rd|ind} t |dkr*t!j"|dg|R|| d} | St |dkrVdd t#||dD} |ddurK| $|dt!j"| || d} | SJ)Nzusing the cuTensorNet backend was requested but it cannot be imported -- maybe you forgot to install cuQuantum Python? Please do "pip install cuquantum-python" or "conda install -c conda-forge cuquantum-python" and retryr) stackleveldtypeoptimizeFTgreedycSsg|]}t|qSr )cupyasarrayropr r rrlsz(_try_use_cutensornet..rcss|]}|jdkVqdSrN)sizer2r r r tsz'_try_use_cutensornet..css|] }t|jdkVqdSr4)rshaper2r r rr6uscsg|] }|jddqS)F)copy)astyper2 result_dtyper rr~cutn_handle_cache) device_idhandleautoblockingcSsg|] }|d|dfqS)rrr rr r rrsr einsum_pathrzthe cuTensorNet backend ignores the "optimize" option except when an explicit contraction path is provided or when optimize=False (disable optimization); also, the maximum intermediate size, if set, is ignoredpath)optionsr.cSsg|] }|D]}|qqSr r )rpairrr r rrr<)%r0cudaruntimeis_hiprACCELERATOR_CUTENSORNETget_routine_acceleratorsrrwarningswarngetr+rany result_typefloat32float64 complex64 complex128 getDevicer _tlsr=createrdestroyr?rranger!rfloatrcontractzipr$) r&kwargsr-rDr(devicer=r? cutn_options raise_warningcutn_optimizeroutr)r r:r_try_use_cutensornetOs            rd) __future__r threadingrMr0r cupy._corercupy.cuda.devicerlocalrWr rrrmemoizerrr+rdr r r rs