o ^i,@sddlmZmZmZddlZddlmZddlmZmZm Z m Z m Z m Z m Z mZmZmZmZmZmZmZmZmZmZmZmZmZddgZGdddeZd d ed e d ed ed e d e d e_deedeedeedeedeedeedeedeedededee efdee efdee efde de dededed ef&d!d"Z!deedeedeedeedeedeedeedeedededee efdee efdee efde de dededed ef&d#d$Z"deedeedeedeedeedeedeedeededede de dee efde de dededed ed%df(d&d'Z#e e!d(  ) )    ) )d-deedeedeedeedeedeed*eededed+eedeedeeded edede de dee efde de def*d,dZ$dS).)castOptionalUnionN)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 _to_scalar_use_grad_for_differentiable _view_as_real DeviceDictDeviceDtypeDict OptimizerParamsTAdamadamcseZdZ     ddddddddded eeefd eeeefeeeffd ed ed ede edededede edeffddZ fddZ ddZ e dddZZS)rMbP?g?g+?:0yE>rFN)foreachmaximize capturabledifferentiablefuseddecoupled_weight_decayparamslrbetaseps weight_decayamsgradr r!r"r#r$r%c st|tr|r | s td|dkrtdd|ks"td|d|ks-td|d|dkr9dksCntd |dd|dkrOdksYntd |dd|ksdtd |t|dtrrt|dtst|dtrt|dtstd t|dtr| s|rtd |ddkrtdt|dtr| s|rtd|ddkrtd|||||||| | | | d } t|| | r| rtdd|_|rtddSdS)NElr as a Tensor is not supported for capturable=False and foreach=TruerTensor lr must be 1-elementzInvalid learning rate: zInvalid epsilon value: r?z#Invalid beta parameter at index 0: z#Invalid beta parameter at index 1: zInvalid weight_decay value: z0betas must be either both floats or both TensorszKbetas[0] as a Tensor is not supported for capturable=False and foreach=Truez!Tensor betas[0] must be 1-elementzKbetas[1] as a Tensor is not supported for capturable=False and foreach=Truez!Tensor betas[1] must be 1-element) r'r(r)r*r+r!r r"r#r$r%z)`fused` does not support `differentiable`Tz0`fused` and `foreach` cannot be `True` together.) isinstancer ValueErrornumelfloatsuper__init__ RuntimeError_step_supports_amp_scaling)selfr&r'r(r)r*r+r r!r"r#r$r%defaults __class__K/var/www/html/RAG/RAG_venv/lib/python3.10/site-packages/torch/optim/adam.pyr5#sz       z Adam.__init__cst||jD]k}|dd|dd|dd|dd|dd|dd|dd}|d D]:}|j|g}t|d krst|d sst |d }|ds]|dritj |t |d |j d ntj |t d|d <q9q dS)Nr+Fr!r r"r#r%r$r&rstepis_fuseddtypedevicerB) r4 __setstate__ param_groups setdefaultstategetlentorch is_tensorr3tensorrrC)r8rHgroupr$pp_statestep_valr:r<r=rErs4           zAdam.__setstate__c Cs~d}|dD]} | jdur|t| O}|| | jjr!td|| j|j| } t| dkr||dr:t| |dsB|drPtj dt |dd| j d ntj d t d | d <tj | tjd | d<tj | tjd | d<|dr|tj | tjd | d<|| d|| d|dr|| d|dr| d jrtd|drt|dr|dstd|| d q|S)NFr&zJAdam does not support sparse gradients, please consider SparseAdam insteadrr$r"r<r?rAr.rDr>) memory_formatexp_avg exp_avg_sqr+max_exp_avg_sqr#zB`requires_grad` is not supported for `step` in differentiable moder r'r,)gradrK is_complexappend is_sparser6rHrJr zerosrrCrM zeros_likepreserve_format requires_gradrL) r8rNparams_with_gradgradsexp_avgs exp_avg_sqsmax_exp_avg_sqs state_steps has_complexrOrHr<r<r= _init_groupsl             zAdam._init_groupc Cs|d}|dur!t |}Wdn1swY|jD]V}g}g}g}g}g}g} |d\} } |||||||| } t|||||| f|d| | | |d|d|d|d|d|d |d |d t|d dt|d d|ddq$|S)zPerform a single optimization step. Args: closure (Callable, optional): A closure that reevaluates the model and returns the loss. Nr(r+r'r*r)r!r r"r#r$ grad_scale found_infr%)r+rdbeta1beta2r'r*r)r!r r"r#r$rfrgr%) _cuda_graph_capture_health_checkrK enable_gradrFrergetattr) r8closurelossrNr^r_r`rarbrcrhrirdr<r<r=r>s`       z Adam.step)rrrrFN)__name__ __module__ __qualname__rrr3rtupleboolrr5rErerr> __classcell__r<r<r:r=r"sT        OKaf Implements Adam 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)} \\ &\hspace{13mm} \lambda \text{ (weight decay)}, \: \textit{amsgrad}, \:\textit{maximize}, \: \epsilon \text{ (epsilon)} \\ &\textbf{initialize} : m_0 \leftarrow 0 \text{ ( first moment)}, v_0\leftarrow 0 \text{ (second moment)},\: 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}\textbf{if} \: \lambda \neq 0 \\ &\hspace{10mm} g_t \leftarrow g_t + \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}\textbf{if} \: amsgrad \\ &\hspace{10mm} v_t^{max} \leftarrow \mathrm{max}(v_{t-1}^{max},v_t) \\ &\hspace{10mm}\widehat{v_t} \leftarrow v_t^{max}/\big(1-\beta_2^t \big) \\ &\hspace{5mm}\textbf{else} \\ &\hspace{10mm}\widehat{v_t} \leftarrow v_t/\big(1-\beta_2^t \big) \\ &\hspace{5mm}\theta_t \leftarrow \theta_{t-1} - \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 `Adam: A Method for Stochastic Optimization`_. 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 (L2 penalty) (default: 0) decoupled_weight_decay (bool, optional): if True, this optimizer is equivalent to AdamW and the algorithm will not accumulate weight decay in the momentum nor variance. (default: False) 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 a= .. Note:: A prototype implementation of Adam and AdamW for MPS supports `torch.float32` and `torch.float16`. .. _Adam\: A Method for Stochastic Optimization: https://arxiv.org/abs/1412.6980 .. _On the Convergence of Adam and Beyond: https://openreview.net/forum?id=ryQu7f-RZ r&r_r`rarbrcrfrgr+rdrhrir'r*r)r!r"r#r%c 'Csp|dur|dus Jtjr%t| tsJt| tsJt| ts$Jnt| } t| tr7| j| jf| i}nd}t |D]\}}|sH||n|| }||}||}||}tj sy|ryt }|jj |jj krq|jj |vsyJd|d|d7}| dkr|r|d| | n"|rt| tr| jr||| }n|j|| d}n|j|| d}t|rt|}t|}t|}|rt||||<t|}|j}|dur|j}||f}||vr| j||dd||<||}n| }||d||r)t| tr)| jr|jt|d| dn|| j||ttd| d n || j||d| d |s<|r|}|r[t| tr[| jrTd| |} n d| |} nd| |} |r~t| tr~| jrwd| |}!n d| |}!nd| |}!| | }"|"}#|!}$|r|r||}%n||}%||t|%||||$|#||#}&n ||$|#||#}&|r|||&nL|||&nEt|}d| |} d| |}!| | }"|!d }$|rtj|||||d |||$|}&n ||$|}&|j||&|" d |r5t||r5t ||||<q=dS) NIIf capturable=True, params and state_steps must be on supported devices: .rralphaT)rCrB non_blocking)weight)value?)out)!rKjit is_scriptingr0r3rrrCrB enumeratecompiler is_compilingrtypemul_r]addcmul_cloneaddrW view_as_realtolerp_squarernegsqrtcopy_maximumadd_addcdiv_rview_as_complex)'r&r_r`rarbrcrfrgr+rdrhrir'r*r)r!r"r#r% beta1_dictiparamrVrSrTstep_tcapturable_supported_devicesrCrBkey device_beta1r>bias_correction1bias_correction2 step_size step_size_negbias_correction2_sqrtrUdenomr<r<r=_single_tensor_adamYs                 rc + st|dkrdSttr|stddkrtdttr2|s(tddkr2tdttrG|s=tddkrGtdtjsi|rit d d t fd d t ||DsiJd d|durq|dussJ|ryJdt t ||||||g}ttrtjdkrjind}|D]\\}}}}}}}ttt|}ttt|}ttt|}ttt|}ttt|} |dj}!|dur|!|vrj|!dd||!<|r||!n}"| r|rttt|}#t|||||#nt|||||rt|}tjs*| djr*tj| tjdddddnt| d| dkrW|rCt|d| n|rOtj||| dntj||| d}t||ttd|"t|ttjr{t|d}$d}%n|}$d}%t||$||%~~$|rt | }&t | }'t!|&dt!|'dt"|'t#|&t$|&t%|'|&}(|'})|rttt|}#t&|#|t'|#}*nt'|}*t#|*|)t|*|t#|*|(t(|||*qfdd| D}&fdd| D}'t)fdd|&D}(dd|'D})|r.ttt|}#t&|#|t'|#}*nt'|}*t#|*|)t|*|t(|||*|(qdS)Nrr,rr-zHbeta1 as a Tensor is not supported for capturable=False and foreach=TruezTensor beta1 must be 1-elementzHbeta2 as a Tensor is not supported for capturable=False and foreach=TruezTensor beta2 must be 1-elementF) supports_xlac3s0|]\}}|jj|jjko|jjvVqdSro)rCr).0rOr>)rr<r= Vs   z%_multi_tensor_adam..rvrwz#_foreach ops don't support autogradcpuTrCrzr/)rCrxcg|] }dt|qSrrrr>)rhr<r= z&_multi_tensor_adam..crrrr)rir<r=rrcsg|]}|dqS)r<rbc)r'r<r=rscSsg|]}|dqS)r}r<rr<r<r=rs)*rJr0rr6r2r1rKrrrallziprr"_group_tensors_by_device_and_dtypestrrCvaluesrlistrr _foreach_negis_cpu _foreach_add_rM _foreach_mul_ _foreach_add_foreach_lerp_r3 _foreach_mul_foreach_addcmul_ _foreach_pow _foreach_sub_ _foreach_neg_ _foreach_div__foreach_reciprocal__foreach_sqrt__foreach_maximum_ _foreach_sqrt_foreach_addcdiv_r)+r&r_r`rarbrcrfrgr+rdrhrir'r*r)r!r"r#r%grouped_tensorsrdevice_params_ device_grads_device_exp_avgs_device_exp_avg_sqs_device_max_exp_avg_sqs_device_state_steps__ device_params device_gradsdevice_exp_avgsdevice_exp_avg_sqsdevice_state_stepsrCrdevice_max_exp_avg_sqsscaled_device_gradsr|rrrrexp_avg_sq_sqrtr<)rhrirr'r=_multi_tensor_adam s&                                      rreturnc 'Cs|sdS|r td|dur|j|ini}|dur|j|ini}t| tr1t| jdkr1| j| ind}t||||||g}|D]\\}}\\}}}}}}}tt t|}tt t|} tt t|}!tt t|}"tt t|}#d\}$}%|dur| ||j |dd}$|dur| ||j |dd}%|dur||vr| j |dd||<||} t |#d|st jnt j}&|&|| |!|"||#|| | | | |||$|%d|%durt |#|%gt|#qBdS) Nz9Adam with fused=True does not support differentiable=Truer)NNT)rzrr) r+r'rhrir*r)r!rfrg)r6rCr0rrrritemsrrrGrrKr _fused_adam_ _fused_adamw_rrJ)'r&r_r`rarbrcrfrgr+rdrhrir'r*r)r!r"r#r%grad_scale_dictfound_inf_dictlr_dictrrCrrrrrrrrrrrrdevice_grad_scaledevice_found_inffuncr<r<r= _fused_adams$  r)single_tensor_fnFr r$cCs| dur|durt||dd\}}|rt|tr|sd}| dur"d} |dur(d}tjs:tdd|Ds:td|rEtj rEtd| rPtj rPtd| rZtj sZt }n |rdtj sdt }nt }|||||||f|| ||||||||| | | d dS) znFunctional API that performs Adam algorithm computation. See :class:`~torch.optim.Adam` for details. NF) use_fusedcss|] }t|tjVqdSro)r0rKr)rtr<r<r=rs  zadam..zPAPI has changed, `state_steps` argument must contain a list of singleton tensorsz6torch.jit.script not supported with foreach optimizersz4torch.jit.script not supported with fused optimizers) r+rdrhrir'r*r)r!r"r#rfrgr%) rr0rrKrrrr6rrrrr)r&r_r`rarbrcr r"r#r$rfrgrdr%r+rhrir'r*r)r!rrr<r<r=rus^#  )NFFNNNFF)%typingrrrrKr optimizerrrr r r r r rrrrrrrrrrrrr__all__r__doc__rrtr3rrrrr<r<r<r=s Xr%G      H      v      `