JTh_"SrSSKJrJrJr SSKrSSKJr SSKJrJ r J r J r J r J r JrJrJrJrJrJrJrJr SS/r"S S\5rS S \S \ S \S \S \ S3 -\lS\\S\\S\\S\\S\\S\S\S\S\S\S\S\S\S\S\4SjrS\\S\\S\\S\\S\\S\S\S\S\S\S\S\S\S\S\4Sjr\ "\S 9S#S\\S\\S\\S\\S\\S\S!\\S\S\S\S\S\S\S\S\S\4 S"jj5rg)$z'Implementation for the RAdam algorithm.)castOptionalUnionN)Tensor)_capturable_doc_default_to_fused_or_foreach_differentiable_doc_disable_dynamo_if_unsupported _foreach_doc!_get_capturable_supported_devices_get_scalar_dtype _get_value _maximize_doc _params_doc_use_grad_for_differentiable _view_as_real OptimizerParamsTRAdamradamc^\rSrSrSSSSSS.S\S\\\4S\\\4S\S \S \ S \ \ S \ S \ S\ 4U4Sjjjjr U4Sjr Sr \SSj5rSrU=r$)rFN)foreachmaximize capturabledifferentiableparamslrbetaseps weight_decaydecoupled_weight_decayrrrrc >[U[5(aUR5S:wa [S5eSU::d[SU35eSU::d[SU35eSUSs=::aS:dO [SUS35eSUSs=::aS:dO [S US35eSU::d[S U35e[ UUUUUUU UU S 9 n [ T U]X5 g) NrzTensor lr must be 1-elementzInvalid learning rate: zInvalid epsilon value: r?z#Invalid beta parameter at index 0: z#Invalid beta parameter at index 1: zInvalid weight_decay value: ) rr r!r"rrrr#r) isinstancernumel ValueErrordictsuper__init__) selfrrr r!r"r#rrrrdefaults __class__s I/var/www/auris/envauris/lib/python3.13/site-packages/torch/optim/radam.pyr,RAdam.__init__s b& ! !bhhjAo:; ;by6rd;< <cz6se<= =eAh$$B58*MN NeAh$$B58*MN Nl";L>JK K%!#9)   *ct>[TU]U5 URGHnURSS5 URSS5 URSS5 URSS5 URSS5 USHnURR U/5n[ U5S:wdM0[R"US 5(aMP[US 5nUS(a([R"U[5URS 9O[R"U[5S 9US 'M GM g) NrrFrr#rrrstepdtypedevicer6) r+ __setstate__ param_groups setdefaultstategetlentorch is_tensorfloattensorrr7)r-r<grouppp_statestep_valr/s r0r9RAdam.__setstate__Fs U#&&E   Y -   Z /   -u 5   5u =   \5 18_**..B/w<1$U__WV_-M-M$WV_5H !. $,=,?#\\(:K:MN FO % 'r2c SnUSGHvnURcMU[R"U5-nURU5 URR(a [ S5eURUR5 UR Un [U 5S:XaUS(a([R"S[5URS9O[R"S[5S 9U S '[R"U[RS 9U S '[R"U[RS 9U S 'URU S 5 URU S 5 URU S 5 GMy U$)NFrz'RAdam does not support sparse gradientsrrr5r%r8r4) memory_formatexp_avg exp_avg_sq)gradr? is_complexappend is_sparse RuntimeErrorr<r>zerosrr7rB zeros_likepreserve_format) r-rCparams_with_gradgradsexp_avgs exp_avg_sqs state_steps has_complexrDr<s r0 _init_groupRAdam._init_groupZsH xAvv!u//22  ''*66##&'PQQ QVV$ 1 u:?!. B.?.A!((S"\\#5F5HI&M (-'7'7)>)>(E)$+0*:*:)>)>+E,'i 01""5#67""5=17!:r2cUR5 SnUb%[R"5 U"5nSSS5 URHrn/n/n/n/n/n[ [ [ [ 4US5upURX4XVXx5n [UUUUUU U USUSUSUSUSUSUS US U S 9 Mt U$!,(df  N=f) zPerform a single optimization step. Args: closure (Callable, optional): A closure that reevaluates the model and returns the loss. Nr rr"r!rrrrr#) beta1beta2rr"r!rrrrr#rZ) _cuda_graph_capture_health_checkr? enable_gradr:rtuplerAr[r) r-closurelossrCrUrVrWrXrYr^r_rZs r0r4 RAdam.step}s --/  ""$y%&&E-/ "$E%'H(*K(*KeUl 3U7^DLE**+K  ;">2%Lz*i( .$%56',-E'F'! '> E%$s B?? C rI)gMbP?)g?g+?g:0yE>rFN)__name__ __module__ __qualname____firstlineno__rrrArrbboolrr,r9r[rr4__static_attributes__ __classcell__)r/s@r0rrs$(%1',&+#' $&+&+ %- &+UE\" &+  &+  &+!%&+$&+&+&+&+&+P(!F"-"-r2aImplements 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{ (weightdecay)}, \:\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 rrVrWrXrYr^r_rr"r!r#rrrrZc ^ ^ ^^^^[U5GHjunnU (dXOX*nX/nX?mXOn[RR5(ddU (a][ 5nUR R UR R :XaUR R U;d SUS35e[R"U5(aX[R"U5n[R"U5n[R"U5n[R"T5mUS- nU (aUO [U5nUS:wa.U (aURSXx-- 5 OURUUS9nURUSU- 5 TRU5RUUSU- S9 SUU-- nSUU-- mUU- nSSU- - S- mTSU-UU--T- - mUU4SjnUU U U4S jnU (aA[R"TS :U"5U"5-S 5nURUU-U-S S9 GM+TS :a&URUU-U"5-U"5-S S9 GMWURUU-S S9 GMm g) NIIf capturable=True, params and state_steps must be on supported devices: .rralpha)valuecD>TS- TS- -T-TS- TS- -T-- S-$)Nrt?rI)rho_infrho_tsr0 _compute_rect+_single_tensor_radam.._compute_rect=sI19aKGaK058:  r2c>TR5nT(aURT5nOURT5nTS-U- $)Nrw)sqrtaddadd_)exp_avg_sq_sqrtbias_correction2rr!rLs r0_compute_adaptive_lr2_single_tensor_radam.._compute_adaptive_lrEsB(oo/O"1"5"5c":"1"6"6s";$c)_< U (a7[ SS9m*[ U*4Sj[ X555(d ST*S35e[R"XX#U/5nUR5GHuunnnnnn[[[U5n[[[U5n[[[U5n[[[U5n[[[U5n[RR5(d>USR(a*[R"U[R"SS S 9SS 9 O[R"US 5 U(a[!UUUU5 U (a[R""U5nS S U- - S - nU (a[R$"UU5n[R&"U5 [R"US 5 [R$"UU5n[R("UU5 [R("US 5 [R*"UU5 [R&"U5 [R"UU5 UnOBUVs/sH5nUS [-U5-U[-U5--S U[-U5-- - - PM7 nnUS:waXU (a[R("US Xx-- 5 O4U (a[R"UUUS 9 O[R."UUUS 9n[R0"UUS U- 5 [R("UU5 [R2"UUUS U- 5 AU (GaT[R4"US5n [R4"US 5n![R("U U!5 A![R("U U5 US- US - -n[R6"UU5n"[R*"U U"5 A"[R8"U 5 [ U U5V#V$s/sH!un#n$[R:"U$S:U#S5PM# n%n#n$A AU%V%s/sHn%[R:"U%S:SS5PM n&n%[R("U&U5 [R$"UU5n[R&"U5 [R"US 5 [R*"U&U5 [R&"U&5 [R$"UU5n[R&"U5 [R"US 5 [R8"U5 [R("UU5 [R("UW%5 A%[R&"U5 [R*"UU5 AOUV$s/sH+n$U$S:a U$S- U$S - -U-US- US - -U$-- S-OSPM- n%n$U%V%s/sH n%U%S:aSOSPM n'n%UVs/sHnS U[-U5-- PM nn[ U'U5V%V(s/sHun%n(UU%-U(- S-PM n&n%n([ UW%U5VV%V(s/sH'unn%n(S U[-U5-- S-UU%-U(- -S-PM) nn%nn([R<"U5n)[R"U)U 5 [R*"U)U5 [R>"U)5 [R"U)U&5 [R2"UUU)5 GM gs snfs sn$n#fs sn%fs sn$fs sn%fs snfs sn(n%fs sn(n%nf)Nrz#_foreach ops don't support autogradF) supports_xlac3># UHTupURRURR:H=(a URRT;v MV g7frf)r7r).0rDr4rs r0 &_multi_tensor_radam..}sN 4 HHMMT[[-- - > !== >3sAArorpr&cpu)r7rqrrtrvrr%rw) r>r?rrr allzipr"_group_tensors_by_device_and_dtypevaluesrlistris_cpu _foreach_add_rBr _foreach_neg _foreach_pow _foreach_neg_ _foreach_mul_ _foreach_div_r _foreach_add_foreach_lerp__foreach_addcmul_ _foreach_sub _foreach_mul_foreach_sqrt_r _foreach_sqrt_foreach_reciprocal_)+rrVrWrXrYr^r_rr"r!r#rrrrZgrouped_tensorsgrouped_params_grouped_grads_grouped_exp_avgs_grouped_exp_avg_sqs_grouped_state_steps__grouped_params grouped_gradsgrouped_exp_avgsgrouped_exp_avg_sqsgrouped_state_stepsrxrr rho_t_listr4numsub2denomnryrectunrect_step_size unrectifiedbcbufferrs+ @r0_multi_tensor_radamrasM$ 6{aDDD  >> & & ( (Z'H( $ v3    wWWsVttu v  w  BB {;O  " " $   d6lO<T&\>: V .?@"4<1EF"4<1EF ~~**,,1DQ1G1N1N   #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 00022 0  1 %##NA8I4IJ''%~\%*$6$6%~\%M -}a%iH /7   q5y  $$Z3C%%j!4D   T *   W -{w{3G&&z7;E   U +   %BES*AUAUXQ ECKC0AU LPQDD D1Hc3 ?D Q    0" 5$11%9LM     0 1    0! 4    02B C    0 1$11%9LM     0 1    0! 4  !1 2    0" 5    0$ 7    0 1    02B C ( (E19 QYqy" ! ! 4u<>  (  ?CCddq1c1dKC;N :M$EZ---:M  7:+GW6X 6X($dR2%6X   '**=tEU&V &VND$ez$///C7BINKbP&V   $$%89 FC( F$45 ""6* F$45 0@&Ik %X^ R* D   s0:<]0)(]5%];2^=^^ ^/.^ )single_tensor_fnrc h[SU55(d [S5eUc [XSS9unnU(a.[RR 5(a [S5eU(a*[RR 5(d[ nO[nU"UUUUUU U U UUU UUUU S9 g)zhFunctional API that performs RAdam algorithm computation. See :class:`~torch.optim.RAdam` for details. c3V# UHn[U[R5v M! g7frf)r'r?r)rts r0rradam..>s@Kqz!U\\**Ks')zPAPI has changed, `state_steps` argument must contain a list of singleton tensorsNF) use_fusedz6torch.jit.script not supported with foreach optimizers) r^r_rr"r!rr#rrrZ)rrQr r?jit is_scriptingrr)rrVrWrXrYr#rrrrZrr^r_rr"r!rfuncs r0rr$s4 @K@ @ @ ^  1 e 7599))++STTuyy--//"#  ! 5%r2)FNFFFF)__doc__typingrrrr?r optimizerrr r r r r rrrrrrrr__all__rrrArkrrrrIr2r0rs.(( $ G NINd2f        gK ``D L`D <`D6l`Df `D f `D `D `D `D`D `D!`D`D`D`D !`DF@J L@J <@J6l@Jf @J f @J @J @J @J@J @J!@J@J@J@J !@JF 1EF$)" ; L; <;6l;f ; f ;!;d^;;;;; ; !;" #;$%;& ';G;r2