+ j6a Rt^RIHt^RIt^RIHt^RIHtHtHtH t H t H t H t H t HtHtHtHtHtHtHtRR.t!RR]4tRR ] R ] R ] R ] R ] R 2 ,]nR RltRRlt] !]R7RRRll4tR#)z'Implementation for the RAdam algorithm.)castN)Tensor)_capturable_doc_default_to_fused_or_foreach_differentiable_doc_disable_dynamo_if_unsupported _foreach_doc!_get_capturable_supported_devices_get_scalar_dtype _get_value _maximize_doc _params_doc _to_scalar_use_grad_for_differentiable _view_as_real OptimizerParamsTRAdamradamc |aa]tRt^toR RRRRRRRR/V3RlV3RlllltV3R ltR t]RR l4tR t Vt V;t #)rFforeachNmaximize capturabledifferentiablec<V^8dQhRS[RS[S[,RS[S[S[3,RS[RS[RS[RS[R,R S[R S[R S[R R/ #) paramslrbetaseps weight_decaydecoupled_weight_decayrNrrrreturn)rfloatrtuplebool)format __classdict__s"R/home/luca/screen-checking/.venv/lib/python3.14/site-packages/torch/optim/radam.py __annotate__RAdam.__annotate__ s&+&+&+ FN&+UE\" &+  &+  &+!%&+&+&+&+&+ &+c<\V\4'd!VP4^8wd \R4hRV8:g\RV 24hRV8:g\RV 24hRV^,u;8:dR8gM\RV^, 24hRV^,u;8:dR8gM\RV^, 24hRV8:g\RV 24hR VR VR VR VR VRVRV RVRV / p \S V`W4R#)zTensor lr must be 1-elementzInvalid learning rate: zInvalid epsilon value: ?z#Invalid beta parameter at index 0: z#Invalid beta parameter at index 1: zInvalid weight_decay value: rrrr rrrr!rN) isinstancernumel ValueErrorsuper__init__) selfrrrrr r!rrrrdefaults __class__s &&&&&&&$$$$ r(r4RAdam.__init__ s b& ! !bhhjAo:; ;by6rd;< <cz6se<= =eAh$$B58*MN NeAh$$B58*MN Nl";L>JK K " U 3 L  w * $&< n   *r+c<\SV`V4VPEF0pVPRR4VPRR4VPRR4VPRR4VPRR4VR,FpVPP V.4p\ V4^8wgK1\P!VR,4'dKV\VR,4pVR,'d,\P!V\4VPR 7M\P!V\4R 7VR&K EK3 R#) rNrFrr!rrstepdtypedevicer<) r3 __setstate__ param_groups setdefaultstategetlentorch is_tensorr#tensorr r=)r5rBgrouppp_statestep_valr7s&& r(r?RAdam.__setstate__Hs  U#&&E   Y -   Z /   -u 5   5u =   \5 18__**..B/w<1$U__WV_-M-M$WV_5H !.. $,=,?#\\(:K:MN FO % 'r+cbRpVR,EFpVPfKV\P!V4,pVPV4VPP'd \ R4hVPVP4VP V,p \V 4^8XdVR,'d,\P!R \4VPR7M\P!R\4R7V R&\P!V\PR7V R &\P!V\PR7V R &VPV R ,4VPV R ,4VPV R,4EK V#) Frz'RAdam does not support sparse gradientsrr;r.r>r:) memory_formatexp_avg exp_avg_sq)gradrE is_complexappend is_sparse RuntimeErrorrBrDzerosr r=rG zeros_likepreserve_format) r5rHparams_with_gradgradsexp_avgs exp_avg_sqs state_steps has_complexrIrBs &&&&&&& r( _init_groupRAdam._init_group\sP xAvv!u//22  ''*66###&'PQQ QVV$ 1 u:?!.. B.?.A!((S"\\#5F5HI&M (-'7'7)>)>(E)$+0*:*:)>)>+E,'i 01""5#67""5=17!:r+c VP4RpVe.\P!4;_uu_4V!4pRRR4VPFp.p.p.p.p.p\ \ \ \ 3,VR,4wrVPW4WVWx4p \VVVVVV V VR,VR,VR,VR,VR,VR,VR ,VR ,V R 7K V# +'giL;i) zPerform a single optimization step. Args: closure (Callable, optional): A closure that reevaluates the model and returns the loss. Nrrr rrrrrr!) beta1beta2rr rrrrrr!r_) '_accelerator_graph_capture_health_checkrE enable_gradr@rr$r#r`r) r5closurelossrHrZr[r\r]r^rcrdr_s && r(r: RAdam.steps 446  ""$$y%&&E-/ "$E%'H(*K(*KeUl 3U7^DLE**+K  ;">2%Lz*i( .$%56',-E'F'! '> E%$s C33 D rQ)gMbP?)g?g+?g:0yE>FN) __name__ __module__ __qualname____firstlineno__r4r?r`rr:__static_attributes____classdictcell__ __classcell__)r7r's@@r(rrs\&+ $&+&+!&+ %&+&+P(!F"-"--r+aImplements RAdam algorithm. .. math:: \begin{aligned} &\rule{110mm}{0.4pt} \\ &\textbf{input} : \gamma \text{ (lr)}, \: \beta_1, \beta_2 \text{ (betas)}, \: \theta_0 \text{ (params)}, \:f(\theta) \text{ (objective)}, \: \lambda \text{ (weight decay)}, \:\textit{maximize} \\ &\hspace{13mm} \epsilon \text{ (epsilon)}, \textit{decoupled\_weight\_decay} \\ &\textbf{initialize} : m_0 \leftarrow 0 \text{ ( first moment)}, v_0 \leftarrow 0 \text{ ( second moment)}, \\ &\hspace{18mm} \rho_{\infty} \leftarrow 2/(1-\beta_2) -1 \\[-1.ex] &\rule{110mm}{0.4pt} \\ &\textbf{for} \: t=1 \: \textbf{to} \: \ldots \: \textbf{do} \\ &\hspace{6mm}\textbf{if} \: \textit{maximize}: \\ &\hspace{12mm}g_t \leftarrow -\nabla_{\theta} f_t (\theta_{t-1}) \\ &\hspace{6mm}\textbf{else} \\ &\hspace{12mm}g_t \leftarrow \nabla_{\theta} f_t (\theta_{t-1}) \\ &\hspace{6mm} \theta_t \leftarrow \theta_{t-1} \\ &\hspace{6mm} \textbf{if} \: \lambda \neq 0 \\ &\hspace{12mm}\textbf{if} \: \textit{decoupled\_weight\_decay} \\ &\hspace{18mm} \theta_t \leftarrow \theta_{t} - \gamma \lambda \theta_{t} \\ &\hspace{12mm}\textbf{else} \\ &\hspace{18mm} g_t \leftarrow g_t + \lambda \theta_{t} \\ &\hspace{6mm}m_t \leftarrow \beta_1 m_{t-1} + (1 - \beta_1) g_t \\ &\hspace{6mm}v_t \leftarrow \beta_2 v_{t-1} + (1-\beta_2) g^2_t \\ &\hspace{6mm}\widehat{m_t} \leftarrow m_t/\big(1-\beta_1^t \big) \\ &\hspace{6mm}\rho_t \leftarrow \rho_{\infty} - 2 t \beta^t_2 /\big(1-\beta_2^t \big) \\[0.1.ex] &\hspace{6mm}\textbf{if} \: \rho_t > 5 \\ &\hspace{12mm} l_t \leftarrow \frac{\sqrt{ (1-\beta^t_2) }}{ \sqrt{v_t} +\epsilon } \\ &\hspace{12mm} r_t \leftarrow \sqrt{\frac{(\rho_t-4)(\rho_t-2)\rho_{\infty}}{(\rho_{\infty}-4)(\rho_{\infty}-2) \rho_t}} \\ &\hspace{12mm}\theta_t \leftarrow \theta_t - \gamma \widehat{m_t} r_t l_t \\ &\hspace{6mm}\textbf{else} \\ &\hspace{12mm}\theta_t \leftarrow \theta_t - \gamma \widehat{m_t} \\ &\rule{110mm}{0.4pt} \\[-1.ex] &\bf{return} \: \theta_t \\[-1.ex] &\rule{110mm}{0.4pt} \\[-1.ex] \end{aligned} For further details regarding the algorithm we refer to `On the variance of the adaptive learning rate and beyond`_. This implementation provides an option to use either the original weight_decay implementation as in Adam (where the weight_decay is applied to the gradient) or the one from AdamW (where weight_decay is applied to the weight) through the decoupled_weight_decay option. When decoupled_weight_decay is set to False (default), it uses the original Adam style weight decay, otherwise, it uses the AdamW style which corresponds more closely to the `author's implementation`_ in the RAdam paper. Further information about decoupled weight decay can be found in `Decoupled Weight Decay Regularization`_. z Args: a lr (float, Tensor, optional): learning rate (default: 1e-3) betas (Tuple[float, float], optional): coefficients used for computing running averages of gradient and its square (default: (0.9, 0.999)) eps (float, optional): term added to the denominator to improve numerical stability (default: 1e-8) weight_decay (float, optional): weight decay (L2 penalty) (default: 0) decoupled_weight_decay (bool, optional): whether to decouple the weight decay as in AdamW to obtain RAdamW. If True, the algorithm does not accumulate weight decay in the momentum nor variance. (default: False) z a .. _On the variance of the adaptive learning rate and beyond: https://arxiv.org/abs/1908.03265 .. _author's implementation: https://github.com/LiyuanLucasLiu/RAdam .. _Decoupled Weight Decay Regularization: https://arxiv.org/abs/1711.05101 c >V^8dQhR\\,R\\,R\\,R\\,R\\,R\R\R\R \R \R \R \R \R\R\RR/#rrr[r\r]r^rcrdrr rr!rrrr_r"Nlistrr#r%)r&s"r(r)r)sfDfD LfD <fD6lfDf fD f fD fD fD fDfD fD!fDfDfDfD !fD" #fDr+c a a aaaa\PP4'g \V4p\ V4EFwppV 'g W,MW,)pW/,pW?,oWO,p\P P 4'glV 'dd\4pVPPVPP8XdVPPV9g\RV R24h\P!V4'dY\P!V4p\P!V4p\P!V4p\P!S4oV^, pV 'dTM \V4pV^8wd;V 'd VP^Wx,, 4MVPVVR7pVP!V^V, 4SPV4P#VV^V, R7^VV,, p^VV,, oVV, p^^V, , ^, oS^V,VV,,S, , oVV3RlpVV V V3RlpV 'dQ\P$!SR8V!4V!4,R4pVP'VV,V,R R7EKSR8d6VP'VV,V!4,V!4,R R7EKVP'VV,R R7EK R#) IIf capturable=True, params and state_steps must be on supported devices: .alpha)valuec<S^, S^, ,S,S^, S^, ,S,, R,#)?rQ)rho_infrho_tsr( _compute_rect+_single_tensor_radam.._compute_rectEsI19aKGaK058:  r+c<SP4pS'dVPS4pMVPS4pSR,V, #)r)sqrtaddadd_)exp_avg_sq_sqrtbias_correction2rrrPs r(_compute_adaptive_lr2_single_tensor_radam.._compute_adaptive_lrMsB(oo/O"1"5"5c":"1"6"6s";$c)_< V^8dQhR\\,R\\,R\\,R\\,R\\,R\R\R\R \R \R \R \R \R\R\RR/#rtru)r&s"r(r)r)isJJJJ LJJ <JJ6lJJf JJ f JJ JJ JJ JJJJ JJ!JJJJJJJJ !JJ" #JJr+c ja*\V4^8XdR#V 'd \R4h\PP 4'gzV 'dr\ RR7o*\ ;QJd*V*3Rl\WRR74F 'dK RM RM!V*3Rl\WRR744'g\RS* R 24h\V4p\P!WW#V.4pVP4EFwwpppppp\\\,V4p\\\,V4p\\\,V4p\\\,V4p\\\,V4p\PP 4'gJV^,P'd1\P !V\P"!R R R 7R R 7M\P !V^4V'd\%VVVV4V 'd\P&!V4p^^V, , ^, pV 'd\P(!VV4p\P*!V4\P !V^4\P(!VV4p\P,!VV4\P,!V^4\P.!VV4\P*!V4\P !VV4TpM]VUu.uFQpV^\1V4,V\1V4,,^V\1V4,, , , NKS ppV^8wdiV 'd&\P,!V^Wx,, 4M;V 'd\P !VVVR 7M\P2!VVVR 7p\P4!VV^V, 4\P,!VV4\P6!VVV^V, 4?V 'Edb\P8!V^4p \P8!V^4p!\P,!V V!4?!\P,!V V4V^, V^, ,p\P:!VV4p"\P.!V V"4?"\P<!V 4\V VRR7U#U$u.uF!wp#p$\P>!V$R8V#R4NK# p%p#p$? ?V%U%u.uFp%\P>!V%^8RR 4NK p&p%\P,!V&V4\P(!VV4p\P*!V4\P !V^4\P.!V&V4\P*!V&4\P(!VV4p\P*!V4\P !V^4\P<!V4\P,!VV4\P,!VX%4?%\P*!V4\P.!VV4?EM9VU$u.uFTp$V$^8dIV$^, V$^, ,V,V^, V^, ,V$,, R,M^NKV p%p$V%U%u.uFp%V%^8d^MR NK p'p%VUu.uFp^V\1V4,, NK pp\V'VRR7U%U(u.uFwp%p(VV%,V(, R,NK p&p%p(\VX%VRR7UU%U(u.uFCwpp%p(^V\1V4,, R,VV%,V(, ,R,NKE pp%pp(\P@!V4p)\P !V)V 4\P.!V)V4\PB!V)4\P !V)V&4\P6!VVV)4EK R#uupiuup$p#iuup%iuup$iuup%iuupiuup(p%iuup(p%pi)rjNz#_foreach ops don't support autogradF) supports_xlac3<"TFUwrVPPVPP8H;'dVPPS9xKW R#5irk)r=r).0rIr:rs& r( &_multi_tensor_radam..sT A HHMMT[[-- - > > !== >@s :A "A T)strictrxryr/cpu)r=rzrr.r)"rDrrErrr allziprr"_group_tensors_by_device_and_dtypevaluesrrvris_cpu _foreach_add_rGr _foreach_neg _foreach_pow _foreach_neg_ _foreach_mul_ _foreach_div_r _foreach_add_foreach_lerp__foreach_addcmul_ _foreach_sub _foreach_mul_foreach_sqrt_r _foreach_sqrt_foreach_reciprocal_)+rr[r\r]r^rcrdrr rr!rrrr_grouped_tensorsgrouped_params_grouped_grads_grouped_exp_avgs_grouped_exp_avg_sqs_grouped_state_steps__grouped_params grouped_gradsgrouped_exp_avgsgrouped_exp_avg_sqsgrouped_state_stepsrrr rho_t_listr:numsub2denomnrrectunrect_step_size unrectifiedbcbufferrs+&&&&&$$$$$$$$$$ @r(_multi_tensor_radamris$ 6{aBCC >> & & ( (Z'H( $s v4@ sss v4@   ![\x[yyz{  BBBB {;O  " " $   d6lO<T&\>: V .?@"4<1EF"4<1EF ~~**,,1DQ1G1N1N1N   #U\\#e%DC     3Q 7   /?AT  !..}=Mq5y/A% $11%9LM     0 1    0! 4$11%9LM     02E F    0! 4    02B C    0 1    0' :)J0  0D T"#Jt,,.u 4 000222 0  1 %##NA8I4IJ''%~\%*$6$6%~\%M -}a%iH /7   q5y  :$$Z3C%%j!4D   T *   W -{w{3G&&z7;E   U +   % !$CD A AHAu ECKC0 A LPQDD D1Hc3 ?D Q    0" 5$11%9LM     0 1    0! 4    02B C    0 1$11%9LM     0 1    0! 4  !1 2    0" 5    0$ 7    0 1    02B C ( (E19 QYqy" ! ! 4u<>  (  ?CCddq1c1dKC;N :M$EZ----:M   !$K1A$ O  OHD"dR2%% O   '*'/?' 'ND$ez$///C7BINKbPP'  $$%89 FC( F$45 ""6* F$45 0@&Is %X^ R* D   s37Ab )'b$bAb$b>"b#3#b(,A b. )single_tensor_fnc"XV^8dQhR\\,R\\,R\\,R\\,R\\,R\R\R,R \R \R \R \R \R\R\R\R\RR/#)rrr[r\r]r^r!rNrrr_rrcrdrr rr")rvrr%r#)r&s"r(r)r)7s;; L; <;6l;f ; f ;!;D[;;;;; ; !;" #;$%;& ';( );r+c \;QJdRV4F 'dK RM RM !RV44'g \R4hVf\WRR7wppV'd0\PP 4'd \R4hV'd,\PP 4'g\ pM\pV!VVVVVV V V VVV VVVV R7R#) zhFunctional API that performs RAdam algorithm computation. See :class:`~torch.optim.RAdam` for details. c3V"TFp\V\P4xK! R#5irk)r0rEr)rts& r(rradam..Ps@Kqz!U\\**Ks')FTzPAPI has changed, `state_steps` argument must contain a list of singleton tensorsN) use_fusedz6torch.jit.script not supported with foreach optimizers) rcrdrr rrr!rrr_)rrVrrErrrr)rr[r\r]r^r!rrrr_rrcrdrr rrfuncs&&&&&&&&&&&$$$$$ r(rr6s4 3@K@333@K@ @ @ ^  1 e 7599))++STTuyy--//"#  ! 5%r+)FNFFFF)__doc__typingrrEr optimizerrrrrrr r r r r rrrrr__all__rrrrrQr+r(rs. & G NINd2f        gK `fDRJJZ 1EF;G;r+