o I5i7}*@sddlmZmZmZmZmZddlZddlmZddlm Z m Z m Z m Z m Z mZmZmZmZmZmZmZmZmZmZmZmZddgZGdddeZd d ed ed e d e d ed e_d eedeedeedeedeedeedeedeedededededeeefdededededef$dd Zd eedeedeedeedeedeedeedeedededededeeefdededededef$d!d"Z d eedeedeedeedeedeedeedeedededededeeefdededededed#df&d$d%Z!e ed&  ' '    'd+d eedeedeedeedeedeed(eededed)eedeedeedededededeeefdededef(d*dZ"dS),)castListOptionalTupleUnionN)Tensor)_capturable_doc_default_to_fused_or_foreach_device_dtype_check_for_fused_differentiable_doc_disable_dynamo_if_unsupported _foreach_doc _fused_doc!_get_capturable_supported_devices_get_scalar_dtype _get_value _maximize_doc_stack_if_compiling_use_grad_for_differentiable _view_as_real DeviceDict OptimizerParamsTAdamadamcseZdZ     dddddddded eeefd eeefd ed ed ede edededede effddZ fddZ ddZ e dddZZS)rMbP?g?g+?:0yE>rFN)foreachmaximize capturabledifferentiablefusedparamslrbetaseps weight_decayamsgradrr r!r"r#c st|tr|r | s td|dkrtdd|ks"td|d|ks-td|d|dkr9dksCntd |dd|dkrOdksYntd |dd|ksdtd |t|||||||| | | d } t|| | r| rtd d|_|rtddSdS)NElr as a Tensor is not supported for capturable=False and foreach=TruerzTensor lr must be 1-elementzInvalid learning rate: zInvalid epsilon value: r?z#Invalid beta parameter at index 0: z#Invalid beta parameter at index 1: zInvalid weight_decay value: ) r%r&r'r(r)r rr!r"r#z)`fused` does not support `differentiable`Tz0`fused` and `foreach` cannot be `True` together.) isinstancer ValueErrornumeldictsuper__init__ RuntimeError_step_supports_amp_scaling) selfr$r%r&r'r(r)rr r!r"r#defaults __class__^/lsinfo/ai/hellotax_ai/llm_service/venv_embed/lib/python3.10/site-packages/torch/optim/adam.pyr2!sL   z Adam.__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 rr!r"r#r$rstepis_fuseddtypedevicer?) r1 __setstate__ param_groups setdefaultstategetlentorch is_tensorfloattensorrr@)r5rEgroupr#pp_statestep_valr7r9r:rB[s2          zAdam.__setstate__c Cs~d}|dD]} | jdur|t| O}|| | jjr!td|| j|j| } t| dkr||dr:t| |dsB|drPtj dt |dd| j d ntj d t d | d <tj | tjd | d<tj | tjd | d<|dr|tj | tjd | d<|| d|| d|dr|| d|dr| d jrtd|drt|dr|dstd|| d q|S)NFr$zJAdam does not support sparse gradients, please consider SparseAdam insteadrr#r!r9r<r>r+rAr;) memory_formatexp_avg exp_avg_sqr)max_exp_avg_sqr"zB`requires_grad` is not supported for `step` in differentiable moderr%r*)gradrH is_complexappend is_sparser3rErGr zerosrr@rK zeros_likepreserve_format requires_gradrI) r5rLparams_with_gradgradsexp_avgs exp_avg_sqsmax_exp_avg_sqs state_steps has_complexrMrEr9r9r: _init_grouprsl             zAdam._init_groupc Cs|d}|dur!t |}Wdn1swY|jD]S}g}g}g}g}g}g} |d\} } |||||||| } t|||||| f|d| | | |d|d|d|d|d|d |d |d t|d dt|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 rr!r"r# grad_scale found_inf)r)rbbeta1beta2r%r(r'r rr!r"r#rdre) _cuda_graph_capture_health_checkrH enable_gradrCrcrgetattr) r5closurelossrLr\r]r^r_r`rarfrgrbr9r9r:r;s^       z Adam.step)rrrrFN)__name__ __module__ __qualname__rrrJrrboolrr2rBrcrr; __classcell__r9r9r7r:r sN         :Ka 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} \\ &\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}\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}\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-1} - \gamma \widehat{m_t}/ \big(\sqrt{\widehat{v_t}^{max}} + \epsilon \big) \\ &\hspace{5mm}\textbf{else} \\ &\hspace{10mm}\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`_. 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 (L2 penalty) (default: 0) 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^r_r`rardrer)rbrfrgr%r(r'r r!r"c !Cs|dur|dus Jtjrt| tsJt|D]:\}}|s%||n|| }||}||}||}tjsV|rVt}|j j |j j krN|j j |vsVJd|d|d7}| dkre|j || d}t |rt |}t |}t |}|rt ||||<t |}||d| || j||d| d|s|r|}d| |}d| |}| |}|}|}|r|r||}n||}||t||||||||} n |||||} ||| nEt|}d| |}d| |}| |}|d}|r/tj|||||d||||} n |||} |j|| | d|rUt ||rUt||||<qdS) NIIf capturable=True, params and state_steps must be on supported devices: .rralpha)value?)out)rHjit is_scriptingr-rJ enumerate_utils is_compilingrr@typeaddrU view_as_reallerp_mul_addcmul_conjnegsqrtclonecopy_maximumadd_addcdiv_rview_as_complex)!r$r]r^r_r`rardrer)rbrfrgr%r(r'r r!r"iparamrTrQrRstep_tcapturable_supported_devicesr;bias_correction1bias_correction2 step_size step_size_negbias_correction2_sqrtrSdenomr9r9r:_single_tensor_adam@sv             rc %sdt|dkrdSttr|stdtjs5|r5tddtfddt ||Ds5Jdd|dur=|dus?J|rEJd t ||||||g}| D]Z\\}}}}}}}t tt|}t tt|}t tt|}t tt|}t tt|}| r|rt tt|}t|||||nt|||||rt|}tjs|djrtj|tjd d d d d nt|d| dkr|rtj||| d ntj||| d }t||dt|t|||d~|r\t|} t|}!t| dt|!dt|!t| t| t|!| }"|!}#|r=t tt|}t||t|}$nt|}$t|$|#t|$|t|$|"t |||$qTfdd|D} fdd|D}!t!fdd| D}"dd|!D}#|rt tt|}t||t|}$nt|}$t|$|#t|$|t |||$|"qTdS)Nrr*F) supports_xlac3s0|]\}}|jj|jjko|jjvVqdSrm)r@r).0rMr;)rr9r: s   z%_multi_tensor_adam..rsrtz#_foreach ops don't support autogradr,cpu)r@rurcg|] }dt|qSrrrr;)rfr9r: Pz&_multi_tensor_adam..crrrr)rgr9r:rSrcsg|]}|dqS)r9rbc)r%r9r:rWscSsg|]}|dqS)rxr9rr9r9r:rYs)"rGr-rr3rHr}r~rallzipr"_group_tensors_by_device_and_dtypevaluesrrr _foreach_negis_cpu _foreach_add_rK _foreach_add_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$r]r^r_r`rardrer)rbrfrgr%r(r'r r!r"grouped_tensorsdevice_params_ device_grads_device_exp_avgs_device_exp_avg_sqs_device_max_exp_avg_sqs_device_state_steps__ device_params device_gradsdevice_exp_avgsdevice_exp_avg_sqsdevice_state_stepsdevice_max_exp_avg_sqsrrrrexp_avg_sq_sqrtr9)rfrgrr%r:_multi_tensor_adams                              rreturnc %Cs|sdS|r td|dur|j|ini}|dur|j|ini}t| tr1t| jdkr1| j| ind}t||||||g}|D]\\}}\\}}}}}}}tt t|}tt t|}tt t|} tt t|}!tt t|}"|j dkr|dur|dusJd\}#}$|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) non_blocking)r@rr) r)r%rfrgr(r'r rdre)r3r@r-rstrrritemsrrrrDtorHr _fused_adam_rrG)%r$r]r^r_r`rardrer)rbrfrgr%r(r'r r!r"grad_scale_dictfound_inf_dictlr_dictrr@rrrrrrrrrrrrdevice_grad_scaledevice_found_infr9r9r: _fused_adamls$   r)single_tensor_fnFrr#c Cs| dur|durt||dd\}}|rt|tr|sd}| dur"d} |dur(d}tjs:tdd|Ds:td|rEtj rEtd| rPtj rPtd| rZtj sZt }n |rdtj sdt }nt }|||||||| | ||||||||| | d dS) znFunctional API that performs Adam algorithm computation. See :class:`~torch.optim.Adam` for details. NF) use_fusedcss|] }t|tjVqdSrm)r-rHr)rtr9r9r: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)rbrfrgr%r(r'r r!r"rdre) r r-rrHr}r~rr3rzr{rrr)r$r]r^r_r`rarr!r"r#rdrerbr)rfrgr%r(r'r rfuncr9r9r:rsZ"  )NFFNNNF)#typingrrrrrrHr optimizerr r r r r rrrrrrrrrrrr__all__r__doc__rqrJrrrrr9r9r9r:sp L['G      w      6      a