o I5iy*@sddlmZmZmZmZmZddlZddlmZddlm Z m Z m Z m Z m Z mZmZmZmZmZmZmZmZmZmZmZmZddgZGdddeZd d ed ed e d e d ed e_d eedeedeedeedeedeedeedeededededeeefdedededededef$dd Zd eedeedeedeedeedeedeedeededededeeefdedededededef$d!d"Z d eedeedeedeedeedeedeedeededededeeefdedededededed#df&d$d%Z!e ed&  ' '    'd+d eedeedeedeedeedeed(eededed)eedeedeedededededeeefdededef(d*dZ"dS),)castListOptionalTupleUnionN)Tensor)_capturable_doc_default_to_fused_or_foreach_device_dtype_check_for_fused_differentiable_doc_disable_dynamo_if_unsupported _foreach_doc _fused_doc!_get_capturable_supported_devices_get_scalar_dtype _get_value _maximize_doc_stack_if_compiling_use_grad_for_differentiable _view_as_real DeviceDict OptimizerParamsTAdamWadamwcseZdZ     dddddddded eeefd eeefd ed ed edede ededede effddZ fddZ ddZ e dddZZS)rMbP?g?g+?:0yE>{Gz?FN)maximizeforeach capturabledifferentiablefusedparamslrbetaseps weight_decayamsgradr r!r"r#r$c st|tr|r | s td|dkrtdd|ks"td|d|ks-td|d|dkr9dksCntd |dd|dkrOdksYntd |dd|ksdtd |t|||||||| | | 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!r r"r#r$z)`fused` does not support `differentiable`Tz0`fused` and `foreach` cannot be `True` together.) isinstancer ValueErrornumeldictsuper__init__ RuntimeError_step_supports_amp_scaling) selfr%r&r'r(r)r*r r!r"r#r$defaults __class___/lsinfo/ai/hellotax_ai/llm_service/venv_embed/lib/python3.10/site-packages/torch/optim/adamw.pyr3!sL   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$r%rstepis_fuseddtypedevicer@) r2 __setstate__ param_groups setdefaultstategetlentorch is_tensorfloattensorrrA)r6rFgroupr$pp_statestep_valr8r:r;rCVs2          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$r"r:r=r?r,rBr<) memory_formatexp_avg exp_avg_sqmax_exp_avg_sqr*r#zB`requires_grad` is not supported for `step` in differentiable moder!r&r+)gradrI is_complexappend is_sparser4rFrHr zerosrrArL zeros_likepreserve_format requires_gradr.r) r6rMparams_with_gradgradsr*exp_avgs exp_avg_sqsmax_exp_avg_sqs state_steps has_complexrNrFr:r:r; _init_groupmsh             zAdamW._init_groupcCs|d}|dur!t |}Wdn1swY|jD]]}g}g}g}g}g}g} |d} ttttf|d\} } ||||| |||| } t|||||| f| | | |d|d|d|d|d|d |d |d t |d dt |d d| dq$|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$rerfrc) _cuda_graph_capture_health_checkrI enable_gradrDrrrKrdrgetattr)r6closurelossrMr]r^r_r`rarbr*rgrhrcr:r:r;r<sb      z AdamW.step)rrrrFN)__name__ __module__ __qualname__rrrKrrboolrr3rCrdrr< __classcell__r:r:r8r;r sN         5Ka 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_r`rarbrerfr*rgrhr&r)r(r r"r#rcc !Cs|dur|dus Jtjrt| tsJt|D]6\}}|s%||n|| }||}||}||}tjsV|rVt}|j j |j j krN|j j |vsVJd|dt |rzt |}t |}t |}|rut ||||<t |}|d7}| d| | ||d| | | j||d| d|s|r|}d| |}d| |}| |}|}|}|r|r||}n||}||t||||||| |} n |||| |} ||| nEt|}d| |}d| |}| |}|d}|r+tj|||||d|||| } n ||| } |j|| | d|rQt ||rQt||||<qdS)NIIf capturable=True, params and state_steps must be on supported devices: .r)value?)out)rIjit is_scriptingr.rK enumerate_utils is_compilingrrAtyperV view_as_realmul_lerp_addcmul_negsqrtclonecopy_maximumadd_addcdiv_rview_as_complex)!r%r^r_r`rarbrerfr*rgrhr&r)r(r r"r#rciparamrUrRrSstep_tcapturable_supported_devicesr<bias_correction1bias_correction2 step_size step_size_negbias_correction2_sqrtrTdenomr:r:r;_single_tensor_adamw=st             rc %sRt|dkrdSttr|stdtjs5|r5tddtfddt ||Ds5Jdd|r;Jd |durC|dusEJt ||||||g}| D]Q\\}}}}}}}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| dkrt|d| t||dt|t|||d~|rSt|} t|}!t| dt|!dt|!t| t| t|!| }"|!}#|r4t 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|jjvVqdSrn)rAr~).0rNr<)rr:r; s   z&_multi_tensor_adamw..rtruz#_foreach ops don't support autogradr-cpu)rA)alpharcg|] }dt|qSrrrr<)rgr:r; Hz'_multi_tensor_adamw..crrrr)rhr:r;rKrcsg|]}|dqS)r:rbc)r&r:r;rOscSsg|]}|dqS)rwr:rr:r:r;rQs)!rHr.rr4rIr|r}rallzipr"_group_tensors_by_device_and_dtypevaluesrrr _foreach_negis_cpu _foreach_add_rL _foreach_mul__foreach_lerp__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`rarbrerfr*rgrhr&r)r(r r"r#rcgrouped_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:)rgrhrr&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 |"|$gt|"qBdS) Nz9Adam with fused=True does not support differentiable=Truermps)NNT) non_blocking)rArr) r*r&rgrhr)r(r rerf)r4rAr.rstrrritemsrrr~rEtorIr _fused_adamw_rrH)%r%r^r_r`rarbrerfr*rgrhr&r)r(r r"r#rcgrad_scale_dictfound_inf_dictlr_dictrrArrrrrrrrrrrrdevice_grad_scaledevice_found_infr:r:r; _fused_adamwjs$   r)single_tensor_fnFr!r$c Cstjstdd|Dstd| dur.|dur.t||dd\}}|r.t|tr.|s.d}| dur4d} |dur:d}|rEtj rEtd| rPtj rPtd| rZtj sZt }n |rdtj sdt }nt }|||||||| ||||||||| | | d dS) zpFunctional API that performs AdamW algorithm computation. See :class:`~torch.optim.AdamW` for details. css|] }t|tjVqdSrn)r.rIr)rtr:r:r;rs  zadamw..zPAPI has changed, `state_steps` argument must contain a list of singleton tensorsNF) use_fusedz6torch.jit.script not supported with foreach optimizersz4torch.jit.script not supported with fused optimizers) r*rgrhr&r)r(r r"r#rerfrc) rIr|r}rr4r r.rryrzrrr)r%r^r_r`rarbr!r"r#r$rerfrcr*rgrhr&r)r(r rfuncr:r:r;rsZ  )NFFNNNF)#typingrrrrrrIr optimizerr r r r r rrrrrrrrrrrr__all__r__doc__rrrKrrrrr:r:r:r;sp LX'G     w     7     c