o 闦i*@sddlmZmZmZmZmZddlZddlmZddlm Z m Z m Z m Z m Z mZmZmZmZmZmZmZmZmZmZmZmZmZmZddgZGdddeZd d ed ed ed e d e d ed e_deedeedeedeedeedeedeedeededeee fdeee fdeee fde de dedededef$d d!Z!deedeedeedeedeedeedeedeededeee fdeee fdeee fde de dedededef$d"d#Z"deedeedeedeedeedeedeedeedede de deee fde de dedededed$df&d%d&Z#e e!d'  ( (    (d,deedeedeedeedeedeed)eededed*eedeedeededede de dee efde de def(d+dZ$dS)-)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 _params_doc_stack_if_compiling_use_grad_for_differentiable _view_as_real DeviceDictDeviceDtypeDict OptimizerParamsTAdamWadamwcseZdZ     dddddddded eeefd eeeefeeeffd ed ed edede ededede effddZ fddZ ddZ e dddZZS)rMbP?g?g+?:0yE>{Gz?FN)maximizeforeach capturabledifferentiablefusedparamslrbetaseps weight_decayamsgradr"r#r$r%r&c st|tr|r | s td|dkrtdd|ks"td|d|ks-td|d|dkr9dksCntd |dd|dkrOdksYntd |dd|ksdtd |t|dtrrt|dtst|dtrt|dtstd t|dtr| s|rtd |ddkrtdt|dtr| s|rtd|ddkrtdt|||||||| | | d } t|| | r| rtdd|_ |rtddSdS)NElr as a Tensor is not supported for capturable=False and foreach=TruerzTensor lr must be 1-elementzInvalid learning rate: zInvalid epsilon value: r?z#Invalid beta parameter at index 0: z#Invalid beta parameter at index 1: zInvalid weight_decay value: z0betas must be either both floats or both TensorszKbetas[0] as a Tensor is not supported for capturable=False and foreach=Truez!Tensor betas[0] must be 1-elementzKbetas[1] as a Tensor is not supported for capturable=False and foreach=Truez!Tensor betas[1] must be 1-element) r(r)r*r+r,r#r"r$r%r&z)`fused` does not support `differentiable`Tz0`fused` and `foreach` cannot be `True` together.) isinstancer ValueErrornumelfloatdictsuper__init__ RuntimeError_step_supports_amp_scaling) selfr'r(r)r*r+r,r"r#r$r%r&defaults __class__O/home/ubuntu/transcripts/venv/lib/python3.10/site-packages/torch/optim/adamw.pyr6#sz       zAdamW.__init__cst||jD]e}|dd|dd|dd|dd|dd|dd}|dD]:}|j|g}t|d krmt|d smt |d }|dsW|drctj |t |d |j d ntj |t d |d <q3q dS)Nr,Fr"r#r$r%r&r'rstepis_fuseddtypedevicerC) r5 __setstate__ param_groups setdefaultstategetlentorch is_tensorr3tensorrrD)r9rIgroupr&pp_statestep_valr;r=r>rFms2          zAdamW.__setstate__c Cs|d} |dD]} | jdurq| t| O} || | jjr"td|| j|j| } t| dkr{|dr;t| |dsC|drQtj dt |dd| j d ntj d t d | d <tj | tjd | d<tj | tjd | d<|r{tj | tjd | d<|| d|| d|dr|| d|dr| d jrtd|drt|dtr|dstd|| d q| S)NFr'z'AdamW does not support sparse gradientsrr&r$r=r@rBr.rEr?) memory_formatexp_avg exp_avg_sqmax_exp_avg_sqr,r%zB`requires_grad` is not supported for `step` in differentiable moder#r(r-)gradrL is_complexappend is_sparser7rIrKr zerosrrDrN zeros_likepreserve_format requires_gradr0r) r9rOparams_with_gradgradsr,exp_avgs exp_avg_sqsmax_exp_avg_sqs state_steps has_complexrPrIr=r=r> _init_groupsh             zAdamW._init_groupcCs|d}|dur!t |}Wdn1swY|jD]]}g}g}g}g}g}g} |d} ttttf|d\} } ||||| |||| } t|||||| f| | | |d|d|d|d|d|d |d |d t |d dt |d d| dq$|S)zPerform a single optimization step. Args: closure (Callable, optional): A closure that reevaluates the model and returns the loss. Nr,r)r(r+r*r"r#r$r%r& grad_scale found_inf)r,beta1beta2r(r+r*r"r#r$r%r&rgrhre) _cuda_graph_capture_health_checkrL enable_gradrGrrr3rfrgetattr)r9closurelossrOr_r`rarbrcrdr,rirjrer=r=r>r?sb      z AdamW.step)rrr r!FN)__name__ __module__ __qualname__rrr3rrboolrr6rFrfrr? __classcell__r=r=r;r>r"sN        JKa 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-1}}^{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`_. z Args: a 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`rarbrcrdrgrhr,rirjr(r+r*r"r$r%rec &Cs<|dur|dus Jtjr$t| tsJt| tsJt| ts$Jt| tr2| j| jf| i}nd}t|D]b\}}|sC||n|| }||}||}||}tj st|rtt }|jj |jj krl|jj |vstJd|dt |rt|}t|}t|}|rt||||<t|}|d7}|d| | |j}|j}|j}|dur|j}||f}||vr| j||dd||<||}n| }||d||| j||d| d|s|rB|}d| |}d| |} | |}!|!}"| }#|r.|r||}$n||}$||t|$||||#|"| |"}%n ||#|"| |"}%|||%nEt|}d| |}d| |} | |}!| d}#|rutj|||||d|||#| }%n ||#| }%|j||%|! d|rt ||rt||||<q8dS) NIIf capturable=True, params and state_steps must be on supported devices: .rT)rDrC non_blocking)value?)out)rLjit is_scriptingr0r3rrDrC enumeratecompiler is_compilingrtyperX view_as_realmul_tolerp_addcmul_negsqrtclonecopy_maximumadd_addcdiv_rview_as_complex)&r'r`rarbrcrdrgrhr,rirjr(r+r*r"r$r%re beta1_dictiparamrWrTrUstep_tcapturable_supported_devicesrDrCkey device_beta1r?bias_correction1bias_correction2 step_size step_size_negbias_correction2_sqrtrVdenomr=r=r>_single_tensor_adamwSs                rc *s<t|dkrdSttr|stdttr(|stddkr(tdttr=|s3tddkr=tdtjs_|r_t dd t fd d t ||Ds_Jd d |reJd|durm|dusoJt ||||||g}ttrtjdkrjind}|D]\\}}}}}}}ttt|}ttt|}ttt|}ttt|}ttt|}|dj} |dur| |vrՈj| dd|| <|r|| n}!|r|rttt|}"t|||||"nt|||||rt|}tjs|djrtj|tjdddddnt|d| dkr.t|d| t||d|!t|ttjrOt|d}#d}$n|}#d}$t||#||$~~#|rt|}%t|}&t|%dt|&dt|&t |%t!|%t"|&|%}'|&}(|rttt|}"t#|"|t$|"})nt$|})t |)|(t|)| t |)|'t%|||)qfdd|D}%fdd|D}&t&fdd|%D}'dd|&D}(|rttt|}"t#|"|t$|"})nt$|})t |)|(t|)| t%|||)|'qdS)Nrr-zHbeta1 as a Tensor is not supported for capturable=False and foreach=TruerzTensor beta1 must be 1-elementzHbeta2 as a Tensor is not supported for capturable=False and foreach=TruezTensor beta2 must be 1-elementF) supports_xlac3s0|]\}}|jj|jjko|jjvVqdSrp)rDr).0rPr?)rr=r> s   z&_multi_tensor_adamw..rvrwz#_foreach ops don't support autogradcpuTrDrxr/)rD)alphacg|] }dt|qSrrrr?)rir=r> z'_multi_tensor_adamw..crrrr)rjr=r>rrcsg|]}|dqS)r=rbc)r(r=r>rscSsg|]}|dqS)rzr=rr=r=r>rs)'rKr0rr7r1r2rLrrrallzipr"_group_tensors_by_device_and_dtypestrrDvaluesrrrr _foreach_negis_cpu _foreach_add_rN _foreach_mul__foreach_lerp_ _foreach_mul_foreach_addcmul_ _foreach_pow _foreach_sub_ _foreach_neg_ _foreach_div__foreach_reciprocal__foreach_sqrt__foreach_maximum_ _foreach_sqrt_foreach_addcdiv_r)*r'r`rarbrcrdrgrhr,rirjr(r+r*r"r$r%regrouped_tensorsrdevice_params_ device_grads_device_exp_avgs_device_exp_avg_sqs_device_max_exp_avg_sqs_device_state_steps__ device_params device_gradsdevice_exp_avgsdevice_exp_avg_sqsdevice_state_stepsrDrdevice_max_exp_avg_sqsscaled_device_gradsryrrrrexp_avg_sq_sqrtr=)rirjrr(r>_multi_tensor_adamws                                    rreturnc %Cs|sdS|r td|dur|j|ini}|dur|j|ini}t| tr1t| jdkr1| j| ind}t||||||g}|D]\\}}\\}}}}}}}tt t|}tt t|}tt t|} tt t|}!tt t|}"|j dkr|dur|dusJd\}#}$|dur| ||j |dd}#|dur| ||j |dd}$|dur||vr| || j |dd} t |"dt j||| |!||"|| | | | | ||#|$d |$durt |"|$gt|"qBdS) Nz9Adam with fused=True does not support differentiable=Truermps)NNT)rxrr) r,r(rirjr+r*r"rgrh)r7rDr0rrrritemsrrrrHrrLr _fused_adamw_rrK)%r'r`rarbrcrdrgrhr,rirjr(r+r*r"r$r%regrad_scale_dictfound_inf_dictlr_dictrrDrrrrrrrrrrrrdevice_grad_scaledevice_found_infr=r=r> _fused_adamws$   r)single_tensor_fnFr#r&c Cstjstdd|Dstd| dur.|dur.t||dd\}}|r.t|tr.|s.d}| dur4d} |dur:d}|rEtj rEtd| rPtj rPtd| rZtj sZt }n |rdtj sdt }nt }|||||||| ||||||||| | | d dS) zpFunctional API that performs AdamW algorithm computation. See :class:`~torch.optim.AdamW` for details. css|] }t|tjVqdSrp)r0rLr)rtr=r=r>rHs  zadamw..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*r"r$r%rgrhre) rLrrrr7r r0rr|r}rrr)r'r`rarbrcrdr#r$r%r&rgrhrer,rirjr(r+r*r"rfuncr=r=r>r*sZ  )NFFNNNF)%typingrrrrrrLr optimizerr r r r r rrrrrrrrrrrrrr__all__r__doc__rtr3rrrrr=r=r=r>sv Tm'F          b     c