Q/jy*:ddlmZmZmZmZmZddlZddlmZddlm Z m Z m Z m Z m Z mZmZmZmZmZmZmZmZmZmZmZmZddgZGddeZd d ed ed e d e d ed ze_d eedeedeedeedeedeedeedeededededeeefdedededededef$dZd eedeedeedeedeedeedeedeededededeeefdedededededef$d Z d eedeedeedeedeedeedeedeededededeeefdedededededed!df&d"Z!e e# d'd eedeedeedeedeedeed$eededed%eedeedeedededededeeefdededef(d&Z"y)()castListOptionalTupleUnionN)Tensor)_capturable_doc_default_to_fused_or_foreach_device_dtype_check_for_fused_differentiable_doc_disable_dynamo_if_unsupported _foreach_doc _fused_doc!_get_capturable_supported_devices_get_scalar_dtype _get_value _maximize_doc_stack_if_compiling_use_grad_for_differentiable _view_as_real DeviceDict OptimizerParamsTAdamWadamwceZdZ ddddddddedeeefdeeefdeded ed ed e ed ed ede effdZ fdZ dZ e ddZxZS)rFN)maximizeforeach capturabledifferentiablefusedparamslrbetaseps weight_decayamsgradrrr r!r"c t|tr-|r | s td|jdk7r tdd|kstd|d|kstd|d|dcxkrdksntd |dd|dcxkrdksntd |dd|kstd |t |||||||| | | } t ||| | r"| r td d|_|r tdyy)NElr as a Tensor is not supported for capturable=False and foreach=Truer zTensor lr must be 1-elementzInvalid learning rate: zInvalid epsilon value: r?z#Invalid beta parameter at index 0: z#Invalid beta parameter at index 1: zInvalid weight_decay value: ) r$r%r&r'r(rrr r!r"z)`fused` does not support `differentiable`Tz0`fused` and `foreach` cannot be `True` together.) isinstancer ValueErrornumeldictsuper__init__ RuntimeError_step_supports_amp_scaling)selfr#r$r%r&r'r(rrr r!r"defaults __class__s T/root/repos/pocs/video-evals/.venv/lib/python3.12/site-packages/torch/optim/adamw.pyr2zAdamW.__init__!sJ b& !z [xxzQ !>??by6rd;< <cz6se<= =eAh$$B58*MN NeAh$$B58*MN Nl";L>JK K%!)   * "#NOO.2D +"#UVV ct|||jD]#}|jdd|jdd|jdd|jdd|jdd|jdd}|dD]}|jj |g}t |d k7s.tj|d rGt|d }|ds|dr,tj|t| |j ntj|t |d <&y)Nr(Frrr r!r"r#rstepis_fuseddtypedevicer?) r1 __setstate__ param_groups setdefaultstategetlentorch is_tensorfloattensorrr@)r5rEgroupr"pp_statestep_valr7s r8rBzAdamW.__setstate__Vs% U#&& E   Y .   Z /   Y -   \5 1   -u 5$$Wd3E8_ **..B/w<1$U__WV_-M$WV_5H!.%.  $"3U"C#$88 #\\(:K:MNFO  r9c Bd} |dD]} | j| tj| z} |j| | jjr t d|j| j|j | } t| dk(r|dr t| |ds|dr/tjdt|d| j ntjd t | d <tj| tj | d<tj| tj | d<|r(tj| tj | d<|j| d|j| d|dr|j| d|dr| d jr t d|dr#t!|dt"r|ds t d|j| d | S)NFr#z'AdamW does not support sparse gradientsrr"r r<r>r+rAr;) memory_formatexp_avg exp_avg_sqmax_exp_avg_sqr(r!zB`requires_grad` is not supported for `step` in differentiable moderr$r*)gradrH is_complexappend is_sparser3rErGr zerosrr@rK zeros_likepreserve_format requires_gradr-r) r5rLparams_with_gradgradsr(exp_avgs exp_avg_sqsmax_exp_avg_sqs state_steps has_complexrMrEs r8 _init_groupzAdamW._init_groupms  x< .Avv~ 5++A. .K  # #A &vv"#LMM LL JJqME5zQ>1!4\*eGn KK/wH xx c1B1DEf $)#3#3U%:%:$i ',&6&6U%:%:'l#.3.>.>)>)>/E*+ OOE), -   u\2 3Y&&u-='>?%&5=+F+F"X i uT{F3l+"[   uV} -y< .zr9c|jd}|$tj5|}ddd|jD]}g}g}g}g}g}g} |d} t t t t f|d\} } |j|||| |||| } t|||||| f| | | |d|d|d|d|d|d |d |d t|d dt|d d| d|S#1swYxYw)zPerform a single optimization step. Args: closure (Callable, optional): A closure that reevaluates the model and returns the loss. Nr(r%r$r'r&rrr r!r" grad_scale found_inf)r(beta1beta2r$r'r&rrr r!r"rgrhrd) _cuda_graph_capture_health_checkrH enable_gradrCrrrJrergetattr)r5closurelossrLr^r_r`rarbrcr(rirjrds r8r;z AdamW.stepsV --/  ""$ !y !&&* E-/ "$E%'H(*K,.O(*K!),GeUl 3U7^DLE5**  K     ;">2%Lz*i( .$%56Gn"4t<!$ T:') +* X _ ! !s C$$C-)gMbP?)g?g+?g:0yE>g{Gz?FN)__name__ __module__ __qualname__rrrJrrboolrr2rBrerr; __classcell__)r7s@r8rr s$(%1"3W"& $ $3W3W %- 3WUE\" 3W  3W  3W3W3W$3W3W3W~3Wj.IV":":r9a Implements AdamW 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)}, \: \epsilon \text{ (epsilon)} \\ &\hspace{13mm} \lambda \text{(weight decay)}, \: \textit{amsgrad}, \: \textit{maximize} \\ &\textbf{initialize} : m_0 \leftarrow 0 \text{ (first moment)}, v_0 \leftarrow 0 \text{ ( second moment)}, \: \widehat{v_0}^{max}\leftarrow 0 \\[-1.ex] &\rule{110mm}{0.4pt} \\ &\textbf{for} \: t=1 \: \textbf{to} \: \ldots \: \textbf{do} \\ &\hspace{5mm}\textbf{if} \: \textit{maximize}: \\ &\hspace{10mm}g_t \leftarrow -\nabla_{\theta} f_t (\theta_{t-1}) \\ &\hspace{5mm}\textbf{else} \\ &\hspace{10mm}g_t \leftarrow \nabla_{\theta} f_t (\theta_{t-1}) \\ &\hspace{5mm} \theta_t \leftarrow \theta_{t-1} - \gamma \lambda \theta_{t-1} \\ &\hspace{5mm}m_t \leftarrow \beta_1 m_{t-1} + (1 - \beta_1) g_t \\ &\hspace{5mm}v_t \leftarrow \beta_2 v_{t-1} + (1-\beta_2) g^2_t \\ &\hspace{5mm}\widehat{m_t} \leftarrow m_t/\big(1-\beta_1^t \big) \\ &\hspace{5mm}\widehat{v_t} \leftarrow v_t/\big(1-\beta_2^t \big) \\ &\hspace{5mm}\textbf{if} \: amsgrad \\ &\hspace{10mm}\widehat{v_t}^{max} \leftarrow \mathrm{max}(\widehat{v_t}^{max}, \widehat{v_t}) \\ &\hspace{10mm}\theta_t \leftarrow \theta_t - \gamma \widehat{m_t}/ \big(\sqrt{\widehat{v_t}^{max}} + \epsilon \big) \\ &\hspace{5mm}\textbf{else} \\ &\hspace{10mm}\theta_t \leftarrow \theta_t - \gamma \widehat{m_t}/ \big(\sqrt{\widehat{v_t}} + \epsilon \big) \\ &\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 `Decoupled Weight Decay Regularization`_. a Args: params (iterable): iterable of parameters to optimize or dicts defining parameter groups lr (float, Tensor, optional): learning rate (default: 1e-3). A tensor LR is not yet supported for all our implementations. Please use a float LR if you are not also specifying fused=True or capturable=True. 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 coefficient (default: 1e-2) amsgrad (bool, optional): whether to use the AMSGrad variant of this algorithm from the paper `On the Convergence of Adam and Beyond`_ (default: False) z a8 .. Note:: A prototype implementation of Adam and AdamW for MPS supports `torch.float32` and `torch.float16`. .. _Decoupled Weight Decay Regularization: https://arxiv.org/abs/1711.05101 .. _On the Convergence of Adam and Beyond: https://openreview.net/forum?id=ryQu7f-RZ r#r_r`rarbrcrgrhr(rirjr$r'r&rr r!rdc ||Jtjjrt| tsJt |D];\}}|s||n|| }||}||}||}tj js\|rZt}|jj|jjk(r|jj|vs Jd|dtj|rqtj|}tj|}tj|}|rtj||||<tj|}|dz }|jd| | zz |j|d| z |j| j||d| z |s|r|}d| |zz }d| |zz }| |z }|j!}|j#}|ro|r||j%}n||}||j'tj(||||j#||zz j+| |z } n(|j#||zz j+| |z } |j-|| nt/|}d| |zz }d| |zz }| |z }|dz}|rDtj(|||||||j#|z j+| } n"|j#|z j+| } |j-|| | |stj||s!tj0||||<>y)NIIf capturable=True, params and state_steps must be on supported devices: .r )value?)out)rHjit is_scriptingr-rJ enumerate_utils is_compilingrr@typerW view_as_realmul_lerp_addcmul_negsqrtclonecopy_maximumadd_addcdiv_rview_as_complex)!r#r_r`rarbrcrgrhr(rirjr$r'r&rr r!rdiparamrVrSrTstep_tcapturable_supported_devicesr;bias_correction1bias_correction2 step_size step_size_negbias_correction2_sqrtrUdenoms! r8_single_tensor_adamwr=s*  )"33 3 yy"e$$$f%WK5'uQxeAhY1+ ^ Q||((*z+L+N ( !!V]]%7%77LL%%)EE {[[wZxxyz {F   E "%%d+D((1G++J7J%*%7%78J%K"&&u-E !   1rL(()  dAI&''d!e)'D D 5$;  5$; --I%MMOM$4$9$9$; !!%4Q%7%=%=%?N%4Q%7N"((~z)RS $A&++-1F1VW$s]*+ OO%)>)NO$s]*+ NN7E *f%D 5$;  5$; --I$4c$9 ! oa0*/RSBTU)+0025JJPPQTU#*-BBHHM NN7E)N < u''q 2!&!6!6q7I!JOA oWKr9c  't|dk(ryt| tr |s tdtj j s7|r5td't'fdt||Ds Jd'd|rJd||Jtj||||||g}|jD]"\\}}}}}}}ttt|}ttt|}ttt|}ttt|}ttt|}|r7|r'ttt|}t|||||nt|||||rt j |}tj j s=|dj"r.t j$|t j&d d d nt j$|d | dk7rt j(|d | | zz t j*||d | z t j(|| t j,|||d | z ~|rft j.| |} t j.| |}!t j0| d t j0|!d t j2|!t j4| | t j6| t j8|!| }"|!}#|rCttt|}t j:||t j<|}$nt j<|}$t j4|$|#t j$|$| t j4|$|"t j>|||$|D%cgc]}%d | tA|%zz } }%|D%cgc]}%d | tA|%zz }!}%tC| D&cgc] }&| |&z dz c}&}"|!D&cgc]}&|&dz }#}&|rCttt|}t j:||t j<|}$nt j<|}$t j4|$|#t j$|$| t j>|||$|"%ycc}%wcc}%wcc}&wcc}&w)Nrr*F) supports_xlac3K|]N\}}|jj|jjk(xr|jjvPywrp)r@r).0rMr;rs r8 z&_multi_tensor_adamw..sQ 4 HHMMT[[-- - > !== > sAArwrxz#_foreach ops don't support autogradr,cpu)r@)alphar rz)"rGr-rr3rHrrrallzipr"_group_tensors_by_device_and_dtypevaluesrrr _foreach_negis_cpu _foreach_add_rK _foreach_mul__foreach_lerp__foreach_addcmul_ _foreach_pow _foreach_sub_ _foreach_neg_ _foreach_div__foreach_reciprocal__foreach_sqrt__foreach_maximum_ _foreach_sqrt_foreach_addcdiv_rr)(r#r_r`rarbrcrgrhr(rirjr$r'r&rr r!rdgrouped_tensorsdevice_params_ device_grads_device_exp_avgs_device_exp_avg_sqs_device_max_exp_avg_sqs_device_state_steps__ device_params device_gradsdevice_exp_avgsdevice_exp_avg_sqsdevice_state_stepsdevice_max_exp_avg_sqsrrrrexp_avg_sq_sqrtr;bcrs( @r8_multi_tensor_adamwrs* 6{a"fj S  << $ $ &:'H( $ v{3   wWWsVttu v  w DDD   )"33 3BB + LO  " " $D  T&\>: DL-8 tF|-=>!$v,0CD!$v,0CD )-d6l@RS#("5"56L"M"'"5"56H"I   1F G    5    ;  # #M?O T;M 26EZ---   ;M 26EZ---   ,FV,Wb2g^,WXI#3%C% !%)-d6l@RS#("5"56L"M"'"5"56H"I   1F G    5  # #  DJ  -X%sS$7S)S. 4 S3returnc "|sy|r td||j|ini}||j|ini}t| tr&t | jdk7r| j| ind}t j ||||||g}|jD]t\\}}\\}}}}}}}ttt|}ttt|}ttt|} ttt|}!ttt|}"|jdk(r||Jd\}#}$|#|j||j|d}#|#|j||j|d}$|'||vr#|j|| j|d} tj|"dtj||| |!||"|| | | | | ||#|$ |$Rtj |"|$gt#|"zwy) Nz9Adam with fused=True does not support differentiable=Truermps)NNT) non_blocking)r@rr ) r(r$rirjr'r&rrgrh)r3r@r-rstrrritemsrrrrDtorHr _fused_adamw_rrG)%r#r_r`rarbrcrgrhr(rirjr$r'r&rr r!rdgrad_scale_dictfound_inf_dictlr_dictrr@rrrrrrrrrrrrdevice_grad_scaledevice_found_infs% r8 _fused_adamwrjs^* VWW,6+A  J'r*3)>  9%B&b&1c"))n6MBSW  BB + LO    6        "  T&\>: DL-8 tF|-=>!$v,0CD!$v,0CD ;;% $); ;;.8++  ! / : : f4 @!   -88 V$ ?    6#8##V$?B .2      " %(& "  '   "%5$6=O9P$P i6r9)single_tensor_fnrr"c htjjstd|Ds t d| )|'t ||d\}}|rt |tr|sd}| d} |d}|r)tjjr t d| r)tjjr t d| r%tjjst}n-|r%tjjst}nt}|||||||| ||||||||| | | y) zpFunctional API that performs AdamW algorithm computation. See :class:`~torch.optim.AdamW` for details. c3PK|]}t|tj ywrp)r-rHr)rts r8rzadamw..s 3() 1ell#3s$&zPAPI has changed, `state_steps` argument must contain a list of singleton tensorsNF) use_fusedz6torch.jit.script not supported with foreach optimizersz4torch.jit.script not supported with fused optimizers) r(rirjr$r'r&rr r!rgrhrd) rHrrrr3r r-rr|r}rrr)r#r_r`rarbrcrr r!r"rgrhrdr(rirjr$r'r&rrfuncs r8rrs0< << $ $ &s3-830 ^   }1 Ne 7 z"f-jG }599))+STT '')QRR UYY++- //1"#  ! %%r9)NFFNNNF)#typingrrrrrrHr optimizerr r r r rrrrrrrrrrrrr__all__r__doc__rtrJrrrrrQr9r8rs65 * G SISn&N       'OB NtK LtK <tK6ltKf tK &\ tK f tK tKtKtK tK tK femtKtK tK !tK"#tK$%tK&'tKns Ls <s6lsf s &\ s f s sss s s femss s !s"#s$%s&'sl` L` <`6l`f ` &\ ` f ` ``` ` ` fem`` ` !`"#`$%`&'`( )`F 1EF#  #'"&S LS <S6lSf S &\ S f Sd^SSS D>S SSS"#S$ %S& 'S( eVm)S*+S, -S./SGSr9