o ߗi>@srUddlmZmZmZddlZddlmmZddlm Z ddl m Z gZ ee ed<ejjGdddZdS))DictListOptionalN)Tensor)2_scripted_functional_optimizer_deprecation_warning__all__c@seZdZ         ddeededededed ed ed ed ed efddZdedeefddZ deeefddZ dS)_FunctionalSGD{Gz?Fparamslrmomentum dampening weight_decaynesterovmaximizeforeachfused_allow_empty_param_listc Cs|tdd||||d|_||_||_||_| |_tjt tj t t tj ffi|_ t |dkr7| s7tdd|i|_dS)N) stacklevel)r r rrrz%optimizer got an empty parameter listr )rdefaultsrrrrtorchjitannotaterrstrstatelen ValueError param_group) selfr r r rrrrrrrr!d/home/ubuntu/transcripts/venv/lib/python3.10/site-packages/torch/distributed/optim/functional_sgd.py__init__s $z_FunctionalSGD.__init__paramgradc Cs|jd}|jd}|jd}|jd}|g}g}g} d} |durK| ||jr+d} ||jvr5i|j|<|j|} d| vrD|dn|| dttj|| ||||||j|j | |j |j ddd Wdn1sqwY|j|} |d } | dur| | d<dSdS) z[Similar to self.step, but operates on a single parameter and its gradient. rr rr FNTmomentum_buffer rr r rrrhas_sparse_gradrr grad_scale found_infr) rappend is_sparserrno_gradFsgdrrrr) r r$r%rr rr r momentum_buffer_listgradsr(rr&r!r!r" step_param;sR            z_FunctionalSGD.step_param gradientscCs|jd}g}g}g}|jd}|jd}|jd}|jd} t|t|kr:tddt|dd t|d } t||D]7\} } | durx|| || | jrXd } | |jvrbi|j| <|j| } d | vrq|dqA|| d qAt t j ||||||| |j |j | |j|jddd Wdn1swYt|D]\}}|j|} ||}|dur|| d <qdS)Nr r rr rzEthe gradients passed in does not equal to the size of the parameters!zParams length: z. zGradients length: FTr&r')rrrrzipr+r,rrr-r.r/rrrr enumerate)r r3r params_with_gradr1r0r rr rr(r$gradientripr&r!r!r"stepmsn              z_FunctionalSGD.stepN) r r r r FFFFF) __name__ __module__ __qualname__rrfloatboolr#rr2r:r!r!r!r"rsB   !2r)typingrrrrtorch.optim._functionaloptim _functionalr.r,torch.distributed.optim._deprecation_warningrrr__annotations__rscriptrr!r!r!r"s