Ë
    ßGìj€N  ã                    óô  — d Z ddlmZ ddlZddlmZ ddlZddlZddl	m
Z
 ddl	mZ ddlmZ ddlmZ dd	lmZmZ dd
lmZ ddlmZ 	 ddlmZ eZej6                  j8                  Zej6                  j:                  Zdededededef
d„Z dededededef
d„Z!dededededef
d„Z"dededededef
d„Z#dedededede$e   f
d„Z%dededededef
d„Z&dededededef
d„Z'd„ Z(dededededede)e*eef   e*eef   e*eef   e*eef   f   fd„Z+dejX                  defd„Z- eed¬ «      	 dAd!ed"ed#eded$edededed%e.dedejX                  de)eeef   fd&„«       Z/ eej`                  d'¬(«      d!ed"ed#ed)e*eef   d*e*eef   d+e*eef   d,e*eef   d$edededed%e.dede)eeef   fd-„«       Z1d$ede)e)eef   ef   fd.„Z2d/ed$efd0„Z3d1ed$edefd2„Z4d!ed"ed#ed3ed$edededed%e.dedejX                  fd4„Z5dededed%e.ded5efd6„Z6 e
jn                  d7«      Z8d8e8_9        e8ju                   eejv                  e8«      «       e8jy                  e6«       erG ejz                  e8ej|                  d9¬:«        e?ed;«      r ejz                  e8ej€                  d<¬:«       dededed%e.dedejX                  fd=„ZAdededed%e.ded5efd>„ZB e
jn                  d?«      ZCd8eC_9        eCju                   eejv                  eC«      «       eCjy                  eB«       erG ejz                  eCejˆ                  d9¬:«        e?ed@«      r ejz                  eCejŠ                  d<¬:«       e/j�                  e5eA«       y# e$ r dZY �Œ(w xY w)Ba¡
  `jax.experimental.rnn`: GPU accelerated RNN

----------------------------------------------

This module provides experimental support to CUDNN-backed LSTM.

Currently, the only supported RNN flavor is LSTM with double-bias. We use
notations and variable names similar to
https://pytorch.org/docs/stable/generated/torch.nn.LSTM.html#torch.nn.LSTM

and CUDNN_LSTM entry in
https://docs.nvidia.com/deeplearning/cudnn/api/index.html#cudnnRNNMode_t.

Note that a bidirectional LSTM is treated as having twice the number of layers,
where a forward layer i is followed by a reverse layer i. Each direction has
its own associated weights. We use pseudo-layer to denote such layers
following CUDNN documentation
https://docs.nvidia.com/deeplearning/cudnn/api/index.html#cudnnGetRNNWeightParams.

CUDNN takes an opaque 1D weight array that densely packs all the weight arrays
in a sparsely documented layout. Through trial-and-error and testing, we believe
the layout is the following. Assume 2-layer bi-LSTM with double-bias, so 4
pseudo-layers in total (forward-0, reverse-0, forward-1, reverse-1).

There are 4 kinds of weights: W_ih, W_hh, b_ih and b_hh, where

W_ih = (W_ii, W_if, W_ig, W_io) concatenated on leading axis,
W_hh = (W_hi, W_hf, W_hg, W_ho) concatenated on leading axis,
b_ih = (b_ii, b_if, b_ig, b_io) concatenated on leading axis,
b_hh = (b_hi, b_hf, b_hg, b_ho) concatenated on leading axis.

Say W_ih^0 denotates W_ih from pseudo-layer 0. The linear weights are packed
together from all pseudo-layers followed by bias weights from all pseudo-layers.
In particular, for each layer, W_ih is followed by W_hh and b_ih by b_hh.

(W_ih^0, W_hh^0, W_ih^1, W_hh^1, W_ih^2, W_hh^2, W_ih^3, W_hh^3,
 b_ih^0, b_hh^0, b_ih^1, b_hh^1, b_ih^2, b_hh^2, b_ih^3, b_hh^3)

See `get_params_shapes_in_lstm`.

Example usage:
```
  x = jax.random.normal(
      k1, (batch_size, seq_len, input_size), dtype=jnp.float32)
  h_0 = jax.random.normal(
      k2, (num_directions * num_layers, batch_size, hidden_size),
      dtype=jnp.float32)
  c_0 = jax.random.normal(
      k3, (num_directions * num_layers, batch_size, hidden_size),
      dtype=jnp.float32)
  seq_lengths = jnp.ones((batch_size,), dtype=jnp.int32) * seq_len
  weights = rnn.init_lstm_weight(k4, input_size, hidden_size, num_layers,
                                 bidirectional)
  y, h_n, c_n = rnn.lstm(
      x,
      h_0,
      c_0,
      weights,
      seq_lengths=seq_lengths,
      input_size=input_size,
      hidden_size=hidden_size,
      num_layers=num_layers,
      dropout=False,
      bidirectional=bidirectional)
```

TODO:
  - Add support for input and weight dtypes other than float32.
  - Support ragged inputs.
  - Support RNNs other than LSTM.
é    )ÚpartialN)ÚAny)Úcore)Údispatch)Úmlir)Ú
custom_vjp)ÚArrayÚShape)Úlax)Úgpu_rnnÚlayer_iÚ
input_sizeÚhidden_sizeÚbidirectionalÚreturnc                 óH   — | dk(  s| dk(  r	|rd|z  |fS |rdnd}d|z  ||z  fS )zSShape of W_ii|W_if|W_ig|W_io.

  Note that layer_i is an index of pseudo-layers.
  r   é   é   é   © )r   r   r   r   Únum_directionss        úXC:\Users\Stabi\Downloads\virtual-tryon(1)\venv\Lib\site-packages\jax/experimental/rnn.pyÚ_W_ih_lr   l   s?   € ð �‚\�g ’l¡}Ø�‰O˜ZÐ(Ð(á'‘Q¨Q€NØ�‰O˜^¨kÑ9Ð:Ð:ó    c                 ó   — d|z  |fS )zShape of W_hi|W_hf|W_hg|W_ho.r   r   ©r   r   r   r   s       r   Ú_W_hh_lr   y   s   € ð ˆk‰/˜;Ð	'Ð'r   c                 ó   — d|z  fS )zShape of b_ii|b_if|b_ig|b_io.r   r   r   s       r   Ú_b_ih_lr      ó   € ð ˆk‰/Ð	Ðr   c                 ó   — d|z  fS )zShape of b_hi|b_hf|b_hg|b_ho.r   r   r   s       r   Ú_b_hh_lr"   …   r    r   Ú
num_layersc                 ó  — g }|rdnd}||z  }t         t        g}t        |«      D ]%  }|D ]  }	 |	|| ||«      }
|j                  |
«       Œ  Œ' t        t
        g}t        |«      D ]%  }|D ]  }	 |	|| ||«      }
|j                  |
«       Œ  Œ' |S )z?Get flat param shapes in LSTM. See module docstring for layout.r   r   )r   r   ÚrangeÚappendr   r"   )r   r   r#   r   Úlayer_shapesr   Únum_pseudo_layersÚlinear_weightsÚiÚw_kindÚlayer_shapeÚbias_weightss               r   Ú_get_params_shapes_in_lstmr.   ‹   s«   € ð €,Ù%‘1¨1€.Ø  >Ñ1ÐÜœWÐ%€.ÜÐ"Ö#€aÛ ˆÙ˜1˜j¨+°}ÓE€kØ×Ñ˜+Õ&ñ !ð $ô
 œ7Ð#€,ÜÐ"Ö#€aÛˆÙ˜1˜j¨+°}ÓE€kØ×Ñ˜+Õ&ñ ð $ð 
Ðr   c                 óF   — t        | |||«      }t        d„ |D «       «      }|S )zGet param count in LSTM.c              3   óF   K  — | ]  }t        j                  |«      –— Œ y ­w©N)ÚmathÚprod)Ú.0Úshapes     r   Ú	<genexpr>z)get_num_params_in_lstm.<locals>.<genexpr>¥   s   è ø€ Ð?±,¨”D—I‘I˜e×$±,ùs   ‚!)r.   Úsum)r   r   r#   r   r'   Úparam_counts         r   Úget_num_params_in_lstmr9       s-   € ô ,¨J¸ÀZØ,9ó;€,äÑ?±,Ó?Ó?€+Ø	Ðr   Úrngc                 ó¶   — t        ||||«      }t        j                  d|z  «      }t        j                  j                  | |ft        j                  | |¬«      S )zDRandom initialize LSTM weights from U(-k, k), k=sqrt(1/hidden_size).g      ð?)r5   ÚdtypeÚminvalÚmaxval)r9   ÚnpÚsqrtÚjaxÚrandomÚuniformÚjnpÚfloat32)r:   r   r   r#   r   r8   Úks          r   Úinit_lstm_weightrG   ©   s[   € ô ' z°;À
Ø'4ó6€+ä‡g�gˆc�KÑÓ €!Ü	�‰×	Ñ	Ø	�+�¤s§{¡{¸A¸2Àað 
ó 
Ið Ir   c           
      ó
  — t        j                  | «      } t        ||||«      }|rdnd}d}t        |«      D ]É  }t        |«      D ]¹  }	dD ]²  }
|j	                  d«      }t        j                  |«      }| |||z    j                  |«      }t        j                  |dd¬«      }t        j                  |d   |d   |d   |d   gd¬«      }| j                  |||z    j                  |j                  «       «      } ||z  }Œ´ Œ» ŒË | S )zCSwaps the weights for the input and output gates for an LSTM model.r   r   r   ©ÚW_ihÚW_hhÚb_ihÚb_hhr   ©Úaxisé   )rD   Úasarrayr.   r%   Úpopr2   r3   ÚreshapeÚsplitÚconcatenateÚatÚsetÚflatten)Úweightsr   r   r#   r   Úflat_shapesr   Ú	w_offsetsÚlÚ	directionÚ	gate_namer5   Ú	num_elemsÚmatrixÚgatesÚswapped_matrixs                   r   Úswap_lstm_gatesrc   ²   s	  € ä�K‰K˜Ó €'Ü*¨:°{ÀJÐP]Ó^€+Ù%‘1¨1€.à€)Ü�Ö€aÜ˜>Ö*ˆ	ã7ˆ)Ø—‘ Ó"ˆÜ—I‘I˜eÓ$ˆ	Ø˜ 9¨yÑ#8Ð9×AÑAÀ%ÓHˆô —	‘	˜& !¨!Ô,ˆÜŸ™¨%°©(°E¸!±H¸eÀA¹hÈÈaÉÐ)QÐXYÔZˆð —*‘*˜Y y°9Ñ'<Ð=×AÑAÀ.×BXÑBXÓBZÓ[ˆØ�YÑ‰	ñ 8ñ +ð ð  
€.r   rY   c                 ó´  — t        ||||«      }d}d}|rdnd}||z  }	i }
i }t        |	«      D ]I  }|
|fD ]@  }||   }|dz  }t        j                  |«      }| |||z    j	                  |«      ||<   ||z  }ŒB ŒK i }i }t        |	«      D ]I  }||fD ]@  }||   }|dz  }t        j                  |«      }| |||z    j	                  |«      ||<   ||z  }ŒB ŒK |
|||fS )a«  Unpack cudnn LSTM weights into individual weights.

  CUDNN LSTM weight layout: (num_layers, num_directions, W_ih, W_hh, b_ih, b_hh)
  Returns W_ih, W_hh, b_ih, b_hh. e.g. W_ih[2][1] is the concat weights of
  4 weights (W_ii, W_if, W_ig, W_io), each of shape (hidden_size, input_size)
  at 2nd layer for the reverse direction. See notations from
  https://pytorch.org/docs/stable/generated/torch.nn.LSTM.html#torch.nn.LSTM.
  r   r   r   )r.   r%   r2   r3   rS   )rY   r   r   r#   r   rZ   Úflat_shapes_offsetr[   r   r(   rJ   rK   r\   r+   r5   r_   rL   rM   s                     r   Úunpack_lstm_weightsrf   Ì   s5  € ô +¨:°{ÀJØ+8ó:€+àÐØ€)Ù%‘1¨1€.Ø  >Ñ1Ðà€$Ø€$ÜÐ"Ö#€aØ˜“,ˆØÐ,Ñ-€eØ˜AÑÐÜ—)‘)˜EÓ"€iØ˜) I°	Ñ$9Ð:×BÑBÀ5ÓI€fˆQ�iØ�9Ñ�iñ ð $ð €$Ø€$ÜÐ"Ö#€aØ˜“,ˆØÐ,Ñ-€eØ˜AÑÐÜ—)‘)˜EÓ"€iØ˜) I°	Ñ$9Ð:×BÑBÀ5ÓI€fˆQ�iØ�9Ñ�iñ ð $ð 
ˆt�T˜4Ð	Ðr   Ú	precisionc                 óš  — t        j                  | «      } | �t        | t        «      rt	        | «      dk(  sy| x  r& dk(  r!\  t         j
                  j                  k(  r  y x  r& dk(  r!\  t         j
                  j                  k(  r  y   r/ dk(  r*\  t         j
                  j                  k(  r t        d«      ‚ 	 t        d| › �«      ‚)Nr   TFz)bfloat16 support not implemented for LSTMz%Unexpected precision specifier value )r   Úcanonicalize_precisionÚ
isinstanceÚtupleÚlenÚ	PrecisionÚHIGHESTÚHIGHÚDEFAULTÚNotImplementedErrorÚ
ValueError)rg   s    r   Ú_lstm_cudnn_allow_tf32rs   ö   s®   € ô ×(Ñ(¨Ó3€)ØÐœz¨)´UÔ;ÄÀIÃÐRSÒ@SØàß	#Ñ	#Œ#�-‰-×
Ô
 Ð	#Øð 
$ç	 Ñ	 Œ#�-‰-×
Ô
˜aÐ	 Øð 
!ç	#Ð	#Œ#�-‰-×
Ô
 ÜÐ KÓLÐLð 
$à	
ÜÐ>¸y¸kÐJÓKÐKr   )é   é   é   é   é	   é
   )Únondiff_argnumsÚxÚh_0Úc_0Úseq_lengthsÚdropoutc                 óF   — t        | |||||||||	|
¬«      \  \  }}}}|||fS )a`  LSTM via CuDNN or HIPDNN (not-yet-supported).

  Assume batch-first inputs.

  Arguments:
    x: (batch_size, max_seq_length, input_size)
    h_0: (num_directions * num_layers, batch_size, hidden_size)
    c_0: (num_directions * num_layers, batch_size, hidden_size)
    weights: (num_params,) where num_params = get_num_params_in_lstm(...)
    seq_lengths: (batch_size,)
  Returns: (y, h_n, c_n, reserve_space).
    y: (batch_size, max_seq_length, hidden_size * num_directions)
    h_n: (num_directions * num_layers, batch_size, hidden_size)
    c_n: (num_directions * num_layers, batch_size, hidden_size)
  )r   r   r#   r   r   rg   )Úlstm_fwd)r{   r|   r}   rY   r~   r   r   r#   r   r   rg   ÚyÚh_nÚc_nÚ_s                  r   Úlstmr†     sH   € ô& ØØ	Ø	ØØØØØØØ!ØôÑ�-€1€cˆ3�ð 
ˆC�ˆÐr   )rw   rx   ry   é   é   )Ústatic_argnumsrJ   rK   rL   rM   c           	      ó  — |j                   t        j                   d«      k7  rt        d«      ‚|dk7  rt        d«      ‚d„ }d„ }| j                  ddd	«      }|sÙg }g }t	        |
«      D ]‡  }t        |||   ||   ||   ||   ¬
«      }t        ||«      }t        j                  j                  |||   ||   f|«      }t        ||«      \  \  }}}|j                  |«       |j                  |«       Œ‰ t        j                  |«      }t        j                  |«      }|j                  ddd	«      ||fS g }g }t	        |
d	z  «      D ]ü  }t        |||   ||   ||   ||   ¬
«      }t        ||«      }|d	z  dk(  r<t        j                  j                  |||   ||   f|«      }t        ||«      \  \  }}}nlt        ||«      }t        j                  j                  |||   ||   f|«      }t        ||«      \  \  }}}t        ||«      }t        j                  |gd¬«      }|j                  |«       |j                  |«       Œþ t        j                  |«      }t        j                  |«      }|j                  ddd	«      ||fS )z¸Reference implementation of LSTM.

  See https://pytorch.org/docs/stable/generated/torch.nn.LSTM.html#lstm
  https://docs.nvidia.com/deeplearning/cudnn/api/index.html#cudnnRNNMode_t
  Úint32ú `seq_lengths` can only be int32.g        zWDropout not supported in LSTM reference because we cannot determine CUDNN dropout mask.c                óÒ  — | \  }}t        j                  |dd¬«      \  }}	}
}t        j                  |dd¬«      \  }}}}t        j                  |dd¬«      \  }}}}t        j                  |dd¬«      \  }}}}t        ||j                  z  |d    z   ||j                  z  z   |d    z   «      }t        ||	j                  z  |d    z   ||j                  z  z   |d    z   «      }t	        ||
j                  z  |d    z   ||j                  z  z   |d    z   «      }t        ||j                  z  |d    z   ||j                  z  z   |d    z   «      }||z  ||z  z   }|t	        |«      z  }||f|fS )Nr   r   rN   )rD   rT   ÚsigmoidÚTÚtanh)Úcarryr{   rJ   rK   rL   rM   ÚhÚcÚW_iiÚW_ifÚW_igÚW_ioÚW_hiÚW_hfÚW_hgÚW_hoÚb_iiÚb_ifÚb_igÚb_ioÚb_hiÚb_hfÚb_hgÚb_hor*   ÚfÚgÚos                               r   Ú	lstm_cellzlstm_ref.<locals>.lstm_cellI  sg  € Ø�D€A€qÜ ŸY™Y t¨Q°QÔ7Ñ€Dˆ$��dÜ ŸY™Y t¨Q°QÔ7Ñ€Dˆ$��dÜ ŸY™Y t¨Q°QÔ7Ñ€Dˆ$��dÜ ŸY™Y t¨Q°QÔ7Ñ€Dˆ$��dÜ��D—F‘F‘
˜T $™ZÑ'¨!¨d¯f©f©*Ñ4°t¸D±zÑAÓB€AÜ��D—F‘F‘
˜T $™ZÑ'¨!¨d¯f©f©*Ñ4°t¸D±zÑAÓB€AÜˆQ�—‘‰Z˜$˜t™*Ñ$ q¨4¯6©6¡zÑ1°D¸±JÑ>Ó?€AÜ��D—F‘F‘
˜T $™ZÑ'¨!¨d¯f©f©*Ñ4°t¸D±zÑAÓB€AØ	ˆA‰��A‘‰€AØ	ŒD�‹G‰€AØˆqˆ6�1ˆ9Ðr   c                 ó&   —  | ||«      \  }}|||ffS r1   r   )Úcellr‘   r{   r‚   s       r   Úscan_fnzlstm_ref.<locals>.scan_fnZ  s    € Ù�E˜1‹~�H€Eˆ1Ø�5˜!�*ÐÐr   r   r   r   rI   éÿÿÿÿrN   )r<   rD   rq   Ú	transposer%   r   rA   r   ÚscanÚ_extract_outputr&   ÚstackÚ_flip_sequencerU   )r{   r|   r}   rJ   rK   rL   rM   r~   r   r   r#   r   r   r§   rª   Úseq_first_yÚfinal_hÚfinal_cr\   r©   Úcell_fnÚoutÚh_tÚc_trƒ   r„   Úseq_first_y_fwdÚseq_first_y_reversedÚseq_first_y_bwds                                r   Úlstm_refr»   0  sŒ  € ð ×Ñœ#Ÿ)™) GÓ,Ò,Ü
Ð@Ó
AÐAØ�‚^Ü
Øaóð òò"ð —‘˜A˜q !Ó$€+Ù	Ø€GØ€GÜ�:ÖˆÜØ
˜$˜q™'¨¨Q©°d¸1±gÀDÈÁGôM€dä˜ Ó&€gÜ�G‰G�L‰L˜ 3 q¡6¨3¨q©6Ð"2Ø-8ó:€cä /°¸SÓ AÑ�j€sˆC�+Ø‡n�n�SÔØ‡n�n�SÕð ô �)‰)�GÓ
€CÜ
�)‰)�GÓ
€CØ× Ñ   A qÓ)¨3°Ð3Ð3ð €'Ø€'Ü�˜a‘Ö €aÜØ˜˜Q™ d¨1¡g°D¸±GÀ$ÀqÁ'ôK€Dä�g˜tÓ$€GØˆ1�u�‚zÜ�G‰G�L‰L˜ 3 q¡6¨3¨q©6Ð"2Ø1<ó>€cä$3°KÀÓ$EÑ!�j€sˆC‘/ô ,¨K¸ÓEÐÜ�G‰G�L‰LØ
�C˜‘F˜C ™FÐ#Ð%9ó;€cä$3°KÀÓ$EÑ!�j€sˆC�/ä& ¸ÓD€oä—O‘O _°oÐ$FÈRÔP€kØ‡N�N�3ÔØ‡N�N�3Õð' !ô( 	�	‰	�'Ó€#Ü�	‰	�'Ó€#Ø	×	Ñ	˜q ! QÓ	'¨¨cÐ	1Ð1r   c                 ó  — |\  }\  \  }}}t        || «      }t        || «      }| d    t        j                  |j                  d   t        j                  ¬«      d d …d f   kD  }t        j
                  |d   |d«      }||f|fS )Nr   )r<   ).N)Ú_select_last_carryrD   Úaranger5   r‹   Úwhere)	r~   rµ   r…   ÚhsÚcsr±   r¶   r·   Úmasks	            r   r®   r®   Š  sŽ   € Ø"Ñ€!Ñ�h€rˆ2�Ü˜2˜{Ó+€#Ü˜2˜{Ó+€#ð 
�TÑ	œSŸZ™Z¨×(9Ñ(9¸!Ñ(<ÄCÇIÁIÔNÊqÐRVÈwÑWÑ	W€$ä—	‘	Ø
ˆ9�oØØó	€+ð ˆsˆ�[Ð	 Ð r   Ú	carry_seqc                 óV   — | |dz
  t        j                  | j                  d   «      f   S )Nr   )rD   r¾   r5   )rÃ   r~   s     r   r½   r½   ˜  s(   € Ø	�; ‘?¤C§J¡J¨y¯©¸qÑ/AÓ$BÐBÑ	CÐCr   Ú	sequencesc                 ó¦   — | j                   d   }||z
  } t        j                  t        t        j
                  d¬«      dd¬«      | |«      d d d…   S )Nr   rN   )r   r   r   )Úin_axesÚout_axesr«   )r5   rA   Úvmapr   rD   Úroll)rÅ   r~   Ú	max_stepsÚroll_amountss       r   r°   r°   ›  sW   € Ø�o‰o˜aÑ €)Ø˜[Ñ(€,ð
Œ�‰”'œ#Ÿ(™(¨Ô+°VØô
Ø˜\ó
+Ù+/¨R¨4ñ
1ð 1r   Úwc                 óØ   — |j                   t        j                   d«      k7  rt        d«      ‚t        |
«      }t        j                  | |||||||||	|¬«      \  }}}}|||f| ||||||ffS )Nr‹   rŒ   ©r   r   r#   r   r   Úcudnn_allow_tf32)r<   rD   rq   rs   Ú	rnn_fwd_pÚbind)r{   r|   r}   rÍ   r~   r   r   r#   r   r   rg   rÐ   r‚   rƒ   r„   Úreserve_spaces                   r   r�   r�   £  s‘   € ð ×Ñœ#Ÿ)™) GÓ,Ò,Ü
Ð@Ó
AÐAÜ+¨IÓ6ÐÜ(Ÿ~™~ØØ	Ø	ØØØØØØØ!Ø'ð  .ó  )Ñ€!€Sˆ#ˆ}ð ˆS�#ˆ˜˜C  a¨°a¸ÐGÐ	GÐGr   rÐ   c           
      ó.  — | j                   d   | j                   d   }}|	rdnd}||||z  f}t        j                  || j                  «      }t	        j
                  |||||||	|
«      \  }}t        j                  |ft        j                  «      }||||fS )Nr   r   r   )r5   r   ÚShapedArrayr<   r   Ú)compute_rnn_workspace_reserve_space_sizesrD   rE   )Úx_avalÚh_0_avalÚc_0_avalÚw_avalÚseq_lengths_avalr   r   r#   r   r   rÐ   Ú
batch_sizeÚmax_seq_lengthr   Úoutput_shapeÚoutput_avalr…   Úreserve_space_sizeÚreserve_space_avals                      r   Úrnn_abstract_evalrâ   ¸  s¥   € ð  &Ÿ|™|¨A™°·±¸Q±ˆn€*Ù%‘1¨1€.Ø˜n¨n¸{Ñ.JÐK€,Ü× Ñ  ¨v¯|©|Ó<€+ä×7Ñ7Ø
�k :¨z¸>Ø
�=Ð"2ó4ñ €!Ðô ×'Ñ'Ð);Ð(=¼s¿{¹{ÓKÐØ	�h Ð*<Ð	<Ð<r   Úrnn_fwdTÚcuda)ÚplatformÚmiopen_rnn_loweringÚrocmc                 óÀ   — t        |«      }|\  }	}
}}}}}|\  }}}t        j                  ||||	|
|||||| |||||¬«      \  }}}}||||t        j                  |«      fS )NrÏ   )rs   Ú	rnn_bwd_prÒ   rD   Ú
zeros_like)r   r   r#   r   r   rg   Ú	residualsÚ	gradientsrÐ   r{   r|   r}   rÍ   r~   r‚   rÓ   ÚdyÚdh_nÚdc_nÚdxÚdh_0Údc_0Údws                          r   Úlstm_bwdrô   Ò  s–   € ô ,¨IÓ6ÐØ2;Ñ/€!€Sˆ#ˆq�+˜q -Ø�.€"€dˆDÜ —~‘~ØØ
Ø
ØØ	Ø	ØØØØØØØØØ!Ø'ð! &ó )Ñ€"€dˆD�"ð" ˆd�D˜"œcŸn™n¨[Ó9Ð	:Ð:r   c                 ó   — ||||fS r1   r   )Údy_avalÚdhn_avalÚdcn_avalr×   Úh0_avalÚc0_avalrÚ   Úy_avalrá   rÛ   r   r   r#   r   r   rÐ   s                   r   Úrnn_bwd_abstract_evalrü   ì  s   € ð
 
�˜' 6Ð	)Ð)r   Úrnn_bwdÚmiopen_rnn_bwd_loweringr1   )GÚ__doc__Ú	functoolsr   r2   Útypingr   rA   Únumpyr?   Újax._srcr   r   Újax.interpretersr   Újax._src.custom_derivativesr   Újax._src.typingr	   r
   Újax._src.laxr   Ú	jax.numpyrD   Újax._src.libr   ÚImportErrorÚPRNGKeyArrayÚnnrŽ   r�   ÚintÚboolr   r   r   r"   Úlistr.   r9   rG   rc   rk   Údictrf   ÚPrecisionLikers   Úfloatr†   Újitr»   r®   r½   r°   r�   râ   Ú	PrimitiverÑ   Úmultiple_resultsÚdef_implÚapply_primitiveÚdef_abstract_evalÚregister_loweringÚcudnn_rnn_loweringÚhasattrræ   rô   rü   ré   Úcudnn_rnn_bwd_loweringrþ   Údefvjpr   r   r   Ú<module>r     s8  ðñFõN Û Ý ã 
Û Ý Ý Ý !Ý 2ß (Ý Ý ðÝ"ð €Ø
�&‰&�.‰.€Ø
‡v�v‡{�{€ð
;�Sð 
; cð 
;¸ð 
;Øð
;Ø$)ó
;ð(�Sð ( cð (¸ð (Øð(Ø$)ó(ð�Sð  cð ¸ð ØðØ$)óð�Sð  cð ¸ð ØðØ$)óð¨3ð ¸Sð Ø+.ðà.2ðà7;¸E±{óð* sð ¸ð È#ð Ø*.ðØ36óðI˜,ð I°Cð IÀcð IØ!$ðIØ59óIòð4' Øð' Ø #ð' Ø25ð' ØCFð' àð' ð ˆ4��U�
Ñ˜T # u *Ñ-¨t°C¸°JÑ/?ÀÀcØFKðGLñ BMð Mñ Nó' ðTL c×&7Ñ&7ð L¸Dó Lñ0 	ˆÐ%8Ô9ð >BñˆEð ˜ð  Eð °Eð Èð ØðØ'*ðØ8;ðØFKðàðà),×):Ñ):ðàFKÈEÐSXÐZ_ÐL_ÑF`òó :ðñB 	ˆ�‰Ð!3Ô4ðW2�ð W2˜Eð W2¨ð W2°T¸#¸u¸*Ñ5Eð W2Ø˜˜U˜
Ñ#ðW2Ø+/°°U°
Ñ+;ðW2à˜˜U˜
Ñ#ðW2à27ðW2àEHðW2ð ðW2ð ,/ðW2ð :?ðW2ð !ð	W2ð &+¨5°%¸Ð+>Ñ%?ò	W2ó 5ðW2ðr! ð !°°e¸EÀ5¸LÑ6IÈ5Ð6PÑ0Qó !ðD %ð D°eó Dð1˜eð 1°%ð 1¸Eó 1ðH�ð H˜Eð H¨ð H°%ð HÀeð HØðHØ+.ðHØ<?ðHØJOðHà ðHà-0×->Ñ->óHð*=Ø"%ð=Ø47ð=ØEHð=à$ð=à59ð=ð )-ó=ð  ˆD�N‰N˜9Ó%€	Ø!€	Ô Ø 	× Ñ ‘7˜8×3Ñ3°YÓ?Ô @Ø 	× Ñ Ð-Ô .Ù
Ø€$×Ñ˜ G×$>Ñ$>ÈÕPÙˆWÐ+Ô,Ø€D×Ñ˜9 g×&AÑ&AÈFÕSð;˜ð ;¨3ð ;¸Cð ;È%ð ;Ø ð;Ø-0×->Ñ->ó;ð4*à8;ð*àJMð*ð '*ð*ð 5:ð*ð KOð*ð -1ó	*ð ˆD�N‰N˜9Ó%€	Ø!€	Ô Ø 	× Ñ ‘7˜8×3Ñ3°YÓ?Ô @Ø 	× Ñ Ð1Ô 2Ù
Ø€$×ÑØ�×/Ñ/¸&õBáˆWÐ/Ô0Ø€D×ÑØ�7×2Ñ2¸VõEð ‡�ˆH�hÕ øðw ò Øƒ'ðús   ÁM, Í,M7Í6M7