o H h2@sjUddlmZddlZddlmmZddlmZddlm Z gZ e e e d<ejjGdddZdS))OptionalN)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 r r rz%optimizer got an empty parameter listr )rdefaultsrrrrtorchjitannotatedictrstrstatelen ValueError param_group) selfr r r r r rrrrrr Z/var/www/vscode/kcb/lib/python3.10/site-packages/torch/distributed/optim/functional_sgd.py__init__s $z_FunctionalSGD.__init__paramgradc Cs|jd}|jd}|jd}|jd}|g}g}g} d} |durK| ||jr+d} ||jvr5i|j|<|j|} d| vrD|dn|| dttj|| ||||||j|j | |j |j ddd Wdn1sqwY|j|} |d } | dur| | d<dSdS) z[Similar to self.step, but operates on a single parameter and its gradient. r r r r FNTmomentum_buffer r r r r rrhas_sparse_gradrr grad_scale found_infr) rappend is_sparserrno_gradFsgdrrrr) rr#r$r r r r 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 r r r zEthe 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)rr2r params_with_gradr0r/r r r r r'r#gradientripr%r r r!stepmsn              z_FunctionalSGD.stepN) rrrrFFFFF) __name__ __module__ __qualname__listrfloatboolr"rr1r9r r r r!rsB   !2r)typingrrtorch.optim._functionaloptim _functionalr-r,torch.distributed.optim._deprecation_warningrrr=r__annotations__rscriptrr r r r!s