o p i@s\ddlZddlmZddlmZmZddlmZmZddlm Z d ddZ  d d d Z dS) N)_C_ops) check_typecheck_variable_and_dtype)Variablein_dynamic_or_pir_mode) LayerHelperc Cstrt||\}}||fStd it}|jdd}t|dttfd|D] }t |dgddq(||d}||d}|j d||d||fS) ay Check if input X contains all finite data, if yes, scale it by input Scale. $$Out = X / scale$$ If any tensor in X contains Inf or Nan, the Out will generate a indicator. FoundInfinite will be 1 (True), and Out will not be scaled. In this case, the data of Out should not be used, and its data may not be deterministic. Otherwise, FoundInfinite will be 0 (False). Args: x(list|tuple): The input tensors of check_finite_and_unscale operator. scale: The scale of check_finite_and_unscale operator. float_status(Tensor): (Only used on NPU) The float status to check overflow. check_finite_and_unscalebool)dtypexfloat16float32float64Zuint16)XZScale)Out FoundInfinite)typeinputsoutputsN)r) rrZcheck_finite_and_unscale_rlocalsZ"create_variable_for_type_inferencertuplelistr append_op) r scalenameZ float_status found_infhelpererrr_/home/app/PaddleOCR-VL/.venv_paddleocr/lib/python3.10/site-packages/paddle/static/amp/amp_nn.pyrs&   rFc  Cstrt|||||||||| |St|dddgdt|dttfd|D])} t| dgdd| jtj tj fvrF|jtj ksEJdq'|j| jksPJdq't dit } |||||d } ||||d }||||d }t| try| | d <n| |d <| jd| ||d|S)a Update loss scaling according to overall gradients. If all gradients is finite after incr_every_n_steps, loss scaling will increase by incr_ratio. Otherwise, loss scaling will decrease by decr_ratio after decr_every_n_nan_or_inf steps and each step some gradients are infinite. Args: x(list|tuple): The input tensors of update_loss_scaling operator. found_inf (Variable): A boolean variable indicates whether there is any infinite gradient. prev_loss_scaling (Variable): Previous loss scaling. num_good_steps (Variable): A variable accumulates good steps in which all gradients are finite. num_bad_steps (Variable): A variable accumulates bad steps in which some gradients are infinite. incr_every_n_steps (int): A variable represents increasing loss scaling every n consecutive steps with finite gradients. decr_every_n_nan_or_inf (int): A variable represents decreasing loss scaling every n accumulated steps with nan or inf gradients. incr_ratio(float): The multiplier to use when increasing the loss scaling. decr_ratio(float): The less-than-one-multiplier to use when decreasing loss scaling. prev_loss_scalingrrupdate_loss_scalingr r z\The dtype of prev_loss_scaling should be float32 when the dtype of x is float16 or bfloat16.zAThe dtype of prev_loss_scaling should be equal to the dtype of x.)rrZPrevLossScalingZ InGoodStepsZ InBadSteps)rZ LossScalingZ OutGoodStepsZ OutBadSteps)incr_every_n_stepsdecr_every_n_nan_or_inf incr_ratio decr_ratioZ StopUpdate stop_update)rrrattrsN)r")rrZupdate_loss_scaling_rrrrr paddler Zbfloat16rrr isinstancerr)r rr!Znum_good_stepsZ num_bad_stepsr#r$r%r&r'rrrrrr(rrr r"Bsv'    r")NN)FN) r)rZpaddle.base.data_feederrrZpaddle.base.frameworkrrZpaddle.base.layer_helperrrr"rrrr s   3