Ë
    T^(hg ã                   ód  — d Z ddlZddlZddlmZmZmZ ddl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 dd
lmZmZ ddlmZmZmZ ddlmZ ddlmZmZm Z m!Z! ddl"m#Z#  e«       rddlm$Z$  e!jJ                  e&«      Z'dZ(dZ)dZ*g d¢Z+dZ,dZ-dZ.dZ/dZ0	 	 dKdee1e1f   de2de1dee	jf                     de1dejh                  fd„Z5 G d „ d!ejl                  «      Z7 G d"„ d#ejl                  «      Z8 G d$„ d%ejl                  «      Z9 G d&„ d'ejl                  «      Z: G d(„ d)ejl                  «      Z; G d*„ d+ejl                  «      Z< G d,„ d-ejl                  «      Z= G d.„ d/e=«      Z> G d0„ d1ejl                  «      Z? G d2„ d3e?«      Z@ G d4„ d5e?«      ZAe?eAe@d6œZB G d7„ d8ejl                  «      ZC G d9„ d:ejl                  «      ZD G d;„ d<ejl                  «      ZE G d=„ d>e«      ZFd?ZGd@ZH edAeG«       G dB„ dCeF«      «       ZI edDeG«       G dE„ dFeF«      «       ZJ edGeG«       G dH„ dIeF«      «       ZKg dJ¢ZLy)LzPyTorch SEW model.é    N)ÚOptionalÚTupleÚUnion)Únn)ÚCrossEntropyLossé   )ÚACT2FN)Úis_deepspeed_zero3_enabled)Úis_fsdp_managed_module)Ú!flash_attn_supports_top_left_maskÚis_flash_attn_available)ÚBaseModelOutputÚCausalLMOutputÚSequenceClassifierOutput)ÚPreTrainedModel)Úadd_code_sample_docstringsÚadd_start_docstringsÚ%add_start_docstrings_to_model_forwardÚloggingé   )Ú	SEWConfig)Ú_flash_attention_forwardr   zasapp/sew-tiny-100k-ft-ls100h)r   i$  i   z_'MISTER QUILTER IS THE APPOSTILE OF THE MIDDLE CLASSES AND WE ARE GLAD TO WELCOME HIS GOSPOLLE'gáz®GáÚ?z(anton-l/sew-mid-100k-ft-keyword-spottingz'_unknown_'g
×£p=
#@ÚshapeÚ	mask_probÚmask_lengthÚattention_maskÚ	min_masksÚreturnc                 óà  ‡‡‡‡‡— | \  }Š‰dk  rt        d«      ‚‰‰kD  rt        d‰› d‰› d�«      ‚t        j                  j                  d«      j	                  «       Šˆˆˆˆˆfd„}|�-|j                  «       j                  d«      j                  «       nt        |«      D �cg c]  }‰‘Œ c}}t        j                  |‰ft        ¬	«      }	g }
 |‰«      }|d
k(  r|	S |D ]¯  } ||«      }t        j                  j                  t        j                  |‰dz
  z
  «      |d¬«      }t        |«      d
k(  r‰dz
  }n|d
   }t        j                  |t        j                  ||z
  t        j                   ¬	«      |z  g«      }|
j#                  |«       Œ± t        j$                  |
«      }
t        j&                  |
dd…dd…df   ||‰f«      }
|
j)                  ||‰z  «      }
t        j                  ‰«      dddd…f   }t        j&                  |||‰f«      j)                  ||‰z  «      }|
|z   }
|
j+                  «       ‰dz
  kD  r‰dz
  |
|
‰dz
  kD  <   t        j,                  |	|
dd«       |	S c c}w )af  
    Computes random mask spans for a given shape. Used to implement [SpecAugment: A Simple Data Augmentation Method for
    ASR](https://arxiv.org/abs/1904.08779). Note that this method is not optimized to run on TPU and should be run on
    CPU as part of the preprocessing during training.

    Args:
        shape: The shape for which to compute masks. This should be of a tuple of size 2 where
               the first element is the batch size and the second element is the length of the axis to span.
        mask_prob:  The percentage of the whole axis (between 0 and 1) which will be masked. The number of
                    independently generated mask spans of length `mask_length` is computed by
                    `mask_prob*shape[1]/mask_length`. Note that due to overlaps, `mask_prob` is an upper bound and the
                    actual percentage will be smaller.
        mask_length: size of the mask
        min_masks: minimum number of masked spans
        attention_mask: A (right-padded) attention mask which independently shortens the feature axis of
                        each batch dimension.
    r   z&`mask_length` has to be bigger than 0.zO`mask_length` has to be smaller than `sequence_length`, but got `mask_length`: z and `sequence_length`: ú`c                 óœ   •— t        ‰| z  ‰z  ‰z   «      }t        |‰«      }|‰z  ‰kD  r‰‰z  }| ‰dz
  z
  |k  rt        | ‰dz
  z
  d«      }|S )z;Given input length, compute how many spans should be maskedr   r   )ÚintÚmax)Úinput_lengthÚnum_masked_spanÚepsilonr   r   r   Úsequence_lengths     €€€€€úb/var/www/skyplay_api_hub/venv/lib/python3.12/site-packages/transformers/models/sew/modeling_sew.pyÚcompute_num_masked_spanz6_compute_mask_indices.<locals>.compute_num_masked_spanm   so   ø€ ä˜i¨,Ñ6¸ÑDÀwÑNÓOˆÜ˜o¨yÓ9ˆð ˜[Ñ(¨?Ò:Ø-°Ñ<ˆOð ˜;¨™?Ñ+¨oÒ=Ü! ,°+À±/Ñ"BÀAÓFˆOàÐó    Néÿÿÿÿ©Údtyper   F)Úreplace)Ú
ValueErrorÚnpÚrandomÚrandÚitemÚdetachÚsumÚtolistÚrangeÚzerosÚboolÚchoiceÚarangeÚlenÚconcatenateÚonesÚint32ÚappendÚarrayÚbroadcast_toÚreshaper#   Úput_along_axis)r   r   r   r   r   Ú
batch_sizer)   Ú_Úinput_lengthsÚspec_aug_maskÚspec_aug_mask_idxsÚmax_num_masked_spanr$   r%   Úspec_aug_mask_idxÚdummy_mask_idxÚoffsetsr&   r'   s    `` `            @@r(   Ú_compute_mask_indicesrN   G   s­  ü€ ð0 #(Ñ€J�à�Q‚ÜÐAÓBÐBà�_Ò$ÜØ]Ð^iÐ]jØ& Ð&7°qð:ó
ð 	
ô �i‰i�n‰n˜QÓ×$Ñ$Ó&€G÷ð ð$ Ð%ð 	×ÑÓ×#Ñ# BÓ'×.Ñ.Ô0ä',¨ZÓ'8Ö9 !ŠoÒ9ð ô —H‘H˜j¨/Ð:Ä$ÔG€MØÐá1°/ÓBÐà˜aÒØÐà%ò 5ˆá1°,Ó?ˆô ŸI™I×,Ñ,Ü�I‰I�l k°A¡oÑ6Ó7¸ÐRWð -ó 
Ðô Ð Ó! QÒ&ð -¨qÑ0‰Nà.¨qÑ1ˆNäŸN™NØ¤§¡Ð(;¸oÑ(MÔUW×U]ÑU]Ô ^ÐaoÑ oÐpó
Ðð 	×!Ñ!Ð"3Õ4ð/5ô2 Ÿ™Ð"4Ó5Ðô Ÿ™Øš1ša ˜:Ñ&¨Ð5HÈ+Ð(VóÐð ,×3Ñ3°JÐ@SÐVaÑ@aÓbÐô �i‰i˜Ó$ T¨4² ]Ñ3€GÜ�o‰o˜g¨
Ð4GÈÐ'UÓV×^Ñ^ØÐ'¨+Ñ5ó€Gð ,¨gÑ5Ðð ×ÑÓ /°AÑ"5Ò5ØGVÐYZÑGZÐÐ-°À!Ñ0CÑCÑDô ×Ñ�mÐ%7¸¸BÔ?àÐùòw :s   Â$	I+c                   ó&   ‡ — e Zd Zdˆ fd„	Zd„ Zˆ xZS )ÚSEWNoLayerNormConvLayerc                 ód  •— t         ‰| �  «        |dkD  r|j                  |dz
     nd| _        |j                  |   | _        t        j                  | j                  | j                  |j                  |   |j                  |   |j                  ¬«      | _
        t        |j                     | _        y )Nr   r   ©Úkernel_sizeÚstrideÚbias)ÚsuperÚ__init__Úconv_dimÚin_conv_dimÚout_conv_dimr   ÚConv1dÚconv_kernelÚconv_strideÚ	conv_biasÚconvr	   Úfeat_extract_activationÚ
activation©ÚselfÚconfigÚlayer_idÚ	__class__s      €r(   rW   z SEWNoLayerNormConvLayer.__init__À   s—   ø€ Ü‰ÑÔØ<DÀqºL˜6Ÿ?™?¨8°a©<Ò8ÈaˆÔØ"ŸO™O¨HÑ5ˆÔä—I‘IØ×ÑØ×ÑØ×*Ñ*¨8Ñ4Ø×%Ñ% hÑ/Ø×!Ñ!ô
ˆŒ	ô ! ×!?Ñ!?Ñ@ˆ�r*   c                 óJ   — | j                  |«      }| j                  |«      }|S ©N)r_   ra   ©rc   Úhidden_statess     r(   ÚforwardzSEWNoLayerNormConvLayer.forwardÎ   s$   € ØŸ	™	 -Ó0ˆØŸ™¨Ó6ˆØÐr*   ©r   ©Ú__name__Ú
__module__Ú__qualname__rW   rk   Ú__classcell__©rf   s   @r(   rP   rP   ¿   s   ø„ õAör*   rP   c                   ó&   ‡ — e Zd Zdˆ fd„	Zd„ Zˆ xZS )ÚSEWLayerNormConvLayerc                 ó°  •— t         ‰| �  «        |dkD  r|j                  |dz
     nd| _        |j                  |   | _        t        j                  | j                  | j                  |j                  |   |j                  |   |j                  ¬«      | _
        t        j                  | j                  d¬«      | _        t        |j                     | _        y )Nr   r   rR   T)Úelementwise_affine)rV   rW   rX   rY   rZ   r   r[   r\   r]   r^   r_   Ú	LayerNormÚ
layer_normr	   r`   ra   rb   s      €r(   rW   zSEWLayerNormConvLayer.__init__Ö   s¯   ø€ Ü‰ÑÔØ<DÀqºL˜6Ÿ?™?¨8°a©<Ò8ÈaˆÔØ"ŸO™O¨HÑ5ˆÔä—I‘IØ×ÑØ×ÑØ×*Ñ*¨8Ñ4Ø×%Ñ% hÑ/Ø×!Ñ!ô
ˆŒ	ô Ÿ,™, t×'8Ñ'8ÈTÔRˆŒÜ  ×!?Ñ!?Ñ@ˆ�r*   c                 ó´   — | j                  |«      }|j                  dd«      }| j                  |«      }|j                  dd«      }| j                  |«      }|S )Néþÿÿÿr+   )r_   Ú	transposerx   ra   ri   s     r(   rk   zSEWLayerNormConvLayer.forwardå   sV   € ØŸ	™	 -Ó0ˆà%×/Ñ/°°BÓ7ˆØŸ™¨Ó6ˆØ%×/Ñ/°°BÓ7ˆàŸ™¨Ó6ˆØÐr*   rl   rm   rr   s   @r(   rt   rt   Õ   s   ø„ õAör*   rt   c                   ó&   ‡ — e Zd Zdˆ fd„	Zd„ Zˆ xZS )ÚSEWGroupNormConvLayerc                 óÆ  •— t         ‰| �  «        |dkD  r|j                  |dz
     nd| _        |j                  |   | _        t        j                  | j                  | j                  |j                  |   |j                  |   |j                  ¬«      | _
        t        |j                     | _        t        j                  | j                  | j                  d¬«      | _        y )Nr   r   rR   T)Ú
num_groupsÚnum_channelsÚaffine)rV   rW   rX   rY   rZ   r   r[   r\   r]   r^   r_   r	   r`   ra   Ú	GroupNormrx   rb   s      €r(   rW   zSEWGroupNormConvLayer.__init__ò   s¹   ø€ Ü‰ÑÔØ<DÀqºL˜6Ÿ?™?¨8°a©<Ò8ÈaˆÔØ"ŸO™O¨HÑ5ˆÔä—I‘IØ×ÑØ×ÑØ×*Ñ*¨8Ñ4Ø×%Ñ% hÑ/Ø×!Ñ!ô
ˆŒ	ô ! ×!?Ñ!?Ñ@ˆŒäŸ,™,°$×2CÑ2CÐRV×RcÑRcÐlpÔqˆ�r*   c                 ól   — | j                  |«      }| j                  |«      }| j                  |«      }|S rh   )r_   rx   ra   ri   s     r(   rk   zSEWGroupNormConvLayer.forward  s2   € ØŸ	™	 -Ó0ˆØŸ™¨Ó6ˆØŸ™¨Ó6ˆØÐr*   rl   rm   rr   s   @r(   r}   r}   ñ   s   ø„ õrö r*   r}   c                   ó$   ‡ — e Zd Zˆ fd„Zd„ Zˆ xZS )ÚSEWPositionalConvEmbeddingc                 ó¼  •— t         ‰| �  «        t        j                  |j                  |j                  |j
                  |j
                  dz  |j                  |j                  ¬«      | _        t        j                  j                  }t        t        j                  j                  d«      r$t        j                  j                  j                  }t        «       �r(dd l}|j                  j!                  | j                  j"                  d¬«      5   || j                  dd¬«      | _        d d d «       t        | j                  d«      rU| j                  j                  j"                  j$                  }| j                  j                  j"                  j&                  }n,| j                  j(                  }| j                  j*                  }|j                  j-                  | |«       |j                  j-                  | |«       n || j                  dd¬«      | _        t/        |j
                  «      | _        t2        |j4                     | _        y # 1 sw Y   �Œ'xY w)	Né   )rS   ÚpaddingÚgroupsrT   Úweight_normr   ©Úmodifier_rankÚweight)ÚnameÚdimÚparametrizations)rV   rW   r   r[   Úhidden_sizeÚnum_conv_pos_embeddingsÚnum_conv_pos_embedding_groupsÚsqueeze_factorr_   ÚutilsrŠ   Úhasattrr�   r
   Ú	deepspeedÚzeroÚGatheredParametersr�   Ú	original0Ú	original1Úweight_gÚweight_vÚregister_external_parameterÚSEWSamePadLayerrˆ   r	   r`   ra   )rc   rd   rŠ   r—   rœ   r�   rf   s         €r(   rW   z#SEWPositionalConvEmbedding.__init__
  s®  ø€ Ü‰ÑÔÜ—I‘IØ×ÑØ×ÑØ×6Ñ6Ø×2Ñ2°aÑ7Ø×7Ñ7Ø×(Ñ(ô
ˆŒ	ô —h‘h×*Ñ*ˆÜ”2—8‘8×,Ñ,¨mÔ<ÜŸ(™(×3Ñ3×?Ñ?ˆKä%Õ'Ûà—‘×2Ñ2°4·9±9×3CÑ3CÐSTÐ2ÓUñ IÙ'¨¯	©	¸ÀaÔH�”	÷Iä�t—y‘yÐ"4Ô5ØŸ9™9×5Ñ5×<Ñ<×FÑF�ØŸ9™9×5Ñ5×<Ñ<×FÑF‘àŸ9™9×-Ñ-�ØŸ9™9×-Ñ-�Ø�N‰N×6Ñ6°t¸XÔFØ�N‰N×6Ñ6°t¸XÕFá# D§I¡I°HÀ!ÔDˆDŒIä& v×'EÑ'EÓFˆŒÜ  ×!?Ñ!?Ñ@ˆ�÷Iñ Iús   ÄIÉIc                 ól   — | j                  |«      }| j                  |«      }| j                  |«      }|S rh   )r_   rˆ   ra   ri   s     r(   rk   z"SEWPositionalConvEmbedding.forward,  s2   € ØŸ	™	 -Ó0ˆØŸ™ ]Ó3ˆØŸ™¨Ó6ˆàÐr*   rm   rr   s   @r(   r…   r…   	  s   ø„ ô AöDr*   r…   c                   ó$   ‡ — e Zd Zˆ fd„Zd„ Zˆ xZS )rŸ   c                 óP   •— t         ‰| �  «        |dz  dk(  rd| _        y d| _        y )Nr‡   r   r   )rV   rW   Únum_pad_remove)rc   r’   rf   s     €r(   rW   zSEWSamePadLayer.__init__6  s)   ø€ Ü‰ÑÔØ#:¸QÑ#>À!Ò#C˜aˆÕÈˆÕr*   c                 óV   — | j                   dkD  r|d d …d d …d | j                    …f   }|S ©Nr   )r£   ri   s     r(   rk   zSEWSamePadLayer.forward:  s6   € Ø×Ñ Ò"Ø)ª!ªQÐ0F°4×3FÑ3FÐ2FÐ0FÐ*FÑGˆMØÐr*   rm   rr   s   @r(   rŸ   rŸ   5  s   ø„ ôKör*   rŸ   c                   ó$   ‡ — e Zd Zˆ fd„Zd„ Zˆ xZS )ÚSEWUpsamplingc                 óì   •— t         ‰| �  «        t        j                  |j                  |j                  |j
                  z  «      | _        t        |j                     | _	        |j
                  | _        y rh   )
rV   rW   r   ÚLinearr‘   r”   Ú
projectionr	   r`   ra   ©rc   rd   rf   s     €r(   rW   zSEWUpsampling.__init__A  sW   ø€ Ü‰ÑÔÜŸ)™) F×$6Ñ$6¸×8JÑ8JÈV×MbÑMbÑ8bÓcˆŒÜ  ×!?Ñ!?Ñ@ˆŒØ$×3Ñ3ˆÕr*   c                 ó.  — | j                  |«      }| j                  |«      }| j                  dkD  rc|j                  «       \  }}}|| j                  z  }|| j                  z  }|j	                  ||| j                  |«      }|j	                  |||«      }|S )Nr   )rª   ra   r”   ÚsizerC   )rc   rj   ÚbszÚsrc_lenÚsrc_embed_dimÚtgt_lenÚtgt_embed_dims          r(   rk   zSEWUpsampling.forwardG  s˜   € ØŸ™¨Ó6ˆØŸ™¨Ó6ˆà×Ñ Ò"à*7×*<Ñ*<Ó*>Ñ'ˆC�˜-Ø × 3Ñ 3Ñ3ˆGØ)¨T×-@Ñ-@Ñ@ˆMØ)×1Ñ1°#°wÀ×@SÑ@SÐUbÓcˆMØ)×1Ñ1°#°wÀÓNˆMàÐr*   rm   rr   s   @r(   r§   r§   @  s   ø„ ô4ör*   r§   c                   ó.   ‡ — e Zd ZdZˆ fd„Zd„ Zd„ Zˆ xZS )ÚSEWFeatureEncoderz.Construct the features from raw audio waveformc           	      óØ  •— t         ‰| �  «        |j                  dk(  rDt        |d¬«      gt	        |j
                  dz
  «      D �cg c]  }t        ||dz   ¬«      ‘Œ c}z   }nV|j                  dk(  r.t	        |j
                  «      D �cg c]  }t        ||¬«      ‘Œ }}nt        d|j                  › d�«      ‚t        j                  |«      | _        d| _        d	| _        y c c}w c c}w )
NÚgroupr   )re   r   Úlayerz`config.feat_extract_norm` is z), but has to be one of ['group', 'layer']FT)rV   rW   Úfeat_extract_normr}   r7   Únum_feat_extract_layersrP   rt   r/   r   Ú
ModuleListÚconv_layersÚgradient_checkpointingÚ_requires_grad)rc   rd   Úir»   rf   s       €r(   rW   zSEWFeatureEncoder.__init__Z  sé   ø€ Ü‰ÑÔà×#Ñ# wÒ.Ü0°À!ÔDÐEÜINÈv×OmÑOmÐpqÑOqÓIröIØDEÔ'¨¸¸Q¹Ö?òIñ ‰Kð ×%Ñ%¨Ò0ÜNSÐTZ×TrÑTrÓNsÖtÈÔ0°À!ÖDÐtˆKÑtäØ0°×1IÑ1IÐ0JÐJsÐtóð ô Ÿ=™=¨Ó5ˆÔØ&+ˆÔ#Ø"ˆÕùòIùò us   ÁC"Â	C'c                 óJ   — | j                  «       D ]	  }d|_        Œ d| _        y ©NF)Ú
parametersÚrequires_gradr½   ©rc   Úparams     r(   Ú_freeze_parametersz$SEWFeatureEncoder._freeze_parametersk  s(   € Ø—_‘_Ó&ò 	(ˆEØ"'ˆEÕð	(à#ˆÕr*   c                 ó
  — |d d …d f   }| j                   r| j                  rd|_        | j                  D ]K  }| j                   r5| j                  r)| j                  r| j                  |j                  |«      }ŒD ||«      }ŒM |S )NT)r½   ÚtrainingrÂ   r»   r¼   Ú_gradient_checkpointing_funcÚ__call__)rc   Úinput_valuesrj   Ú
conv_layers       r(   rk   zSEWFeatureEncoder.forwardp  s…   € Ø$¢Q¨ WÑ-ˆð ×Ò 4§=¢=Ø*.ˆMÔ'à×*Ñ*ò 	:ˆJØ×"Ò" t×'BÒ'BÀtÇ}Â}Ø $× AÑ AØ×'Ñ'Ø!ó!‘ñ
 !+¨=Ó 9‘ð	:ð Ðr*   )rn   ro   rp   Ú__doc__rW   rÅ   rk   rq   rr   s   @r(   r´   r´   W  s   ø„ Ù8ô#ò"$ö
r*   r´   c                   ó   ‡ — e Zd Zˆ fd„Zˆ xZS )ÚSEWFeatureExtractorc                 óÐ   •— t         ‰| �  |«       t        j                  d| j                  j
                  › d| j                  j                  d   j
                  › d�t        «       y )NzThe class `zD` has been depreciated and will be removed in Transformers v5. Use `r   z
` instead.)rV   rW   ÚwarningsÚwarnrf   rn   Ú	__bases__ÚFutureWarningr«   s     €r(   rW   zSEWFeatureExtractor.__init__„  s[   ø€ Ü‰Ñ˜Ô Ü�‰Ø˜$Ÿ.™.×1Ñ1Ð2ð 3à—N‘N×,Ñ,¨QÑ/×8Ñ8Ð9¸ðEô õ		
r*   )rn   ro   rp   rW   rq   rr   s   @r(   rÎ   rÎ   ƒ  s   ø„ ÷
ð 
r*   rÎ   c                   ó†  ‡ — e Zd ZdZ	 	 	 	 	 ddededededededee   fˆ fd	„Z	d
e
j                  dedefd„Z	 	 	 	 	 dde
j                  dee
j                     deee
j                        dee
j                     dee
j                     dedee
j                  ee
j                     eee
j                        f   fd„Zˆ xZS )ÚSEWAttentionz=Multi-headed attention from 'Attention Is All You Need' paperÚ	embed_dimÚ	num_headsÚdropoutÚ
is_decoderrU   Ú	is_causalrd   c                 ó
  •— t         ‰| �  «        || _        || _        || _        ||z  | _        || _        | j
                  |z  | j                  k7  rt        d| j                  › d|› d�«      ‚| j
                  dz  | _        || _	        || _
        t        j                  |||¬«      | _        t        j                  |||¬«      | _        t        j                  |||¬«      | _        t        j                  |||¬«      | _        y )Nz;embed_dim must be divisible by num_heads (got `embed_dim`: z and `num_heads`: z).g      à¿)rU   )rV   rW   rÖ   r×   rØ   Úhead_dimrd   r/   ÚscalingrÙ   rÚ   r   r©   Úk_projÚv_projÚq_projÚout_proj)	rc   rÖ   r×   rØ   rÙ   rU   rÚ   rd   rf   s	           €r(   rW   zSEWAttention.__init__’  sä   ø€ ô 	‰ÑÔØ"ˆŒØ"ˆŒØˆŒØ! YÑ.ˆŒØˆŒà�M‰M˜IÑ%¨$¯.©.Ò8ÜØMÈdÏnÉnÐM]Ø$ Y K¨rð3óð ð —}‘} dÑ*ˆŒØ$ˆŒØ"ˆŒä—i‘i 	¨9¸4Ô@ˆŒÜ—i‘i 	¨9¸4Ô@ˆŒÜ—i‘i 	¨9¸4Ô@ˆŒÜŸ	™	 )¨Y¸TÔBˆ�r*   ÚtensorÚseq_lenr®   c                 óŽ   — |j                  ||| j                  | j                  «      j                  dd«      j	                  «       S )Nr   r‡   )Úviewr×   rÜ   r{   Ú
contiguous©rc   râ   rã   r®   s       r(   Ú_shapezSEWAttention._shape±  s7   € Ø�{‰{˜3 ¨¯©¸¿¹ÓG×QÑQÐRSÐUVÓW×bÑbÓdÐdr*   rj   Úkey_value_statesÚpast_key_valuer   Úlayer_head_maskÚoutput_attentionsr   c                 ó
  — |du}|j                  «       \  }}	}
| j                  |«      | j                  z  }|r0|�.|d   j                  d   |j                  d   k(  r|d   }|d   }�n
|rE| j	                  | j                  |«      d|«      }| j	                  | j                  |«      d|«      }nÃ|�}| j	                  | j                  |«      d|«      }| j	                  | j                  |«      d|«      }t        j                  |d   |gd¬«      }t        j                  |d   |gd¬«      }nD| j	                  | j                  |«      d|«      }| j	                  | j                  |«      d|«      }| j                  r||f}|| j                  z  d| j                  f} | j	                  ||	|«      j                  |Ž } |j                  |Ž } |j                  |Ž }|j                  d«      }t        j                  ||j                  dd«      «      }|j                  «       || j                  z  |	|fk7  r/t!        d|| j                  z  |	|f› d|j                  «       › �«      ‚|�{|j                  «       |d|	|fk7  r#t!        d	|d|	|f› d|j                  «       › �«      ‚|j                  || j                  |	|«      |z   }|j                  || j                  z  |	|«      }t"        j$                  j'                  |d¬«      }|�›|j                  «       | j                  fk7  r*t!        d
| j                  f› d|j                  «       › �«      ‚|j                  dddd«      |j                  || j                  |	|«      z  }|j                  || j                  z  |	|«      }|r?|j                  || j                  |	|«      }|j                  || j                  z  |	|«      }nd}t"        j$                  j)                  || j(                  | j*                  ¬«      }t        j                  ||«      }|j                  «       || j                  z  |	| j                  fk7  r9t!        d|| j                  z  |	| j                  f› d|j                  «       › �«      ‚|j                  || j                  |	| j                  «      }|j                  dd«      }|j                  ||	| j,                  «      }| j/                  |«      }|||fS )ú#Input shape: Batch x Time x ChannelNr   r‡   r   r+   ©r�   z$Attention weights should be of size ú	, but is z!Attention mask should be of size z/Head mask for a single layer should be of size )ÚprÇ   ú `attn_output` should be of size )r­   rà   rÝ   r   rè   rÞ   rß   ÚtorchÚcatrÙ   r×   rÜ   rå   rC   Úbmmr{   r/   r   Ú
functionalÚsoftmaxrØ   rÇ   rÖ   rá   )rc   rj   ré   rê   r   rë   rì   Úis_cross_attentionr®   r±   rF   Úquery_statesÚ
key_statesÚvalue_statesÚ
proj_shaper¯   Úattn_weightsÚattn_weights_reshapedÚ
attn_probsÚattn_outputs                       r(   rk   zSEWAttention.forward´  s  € ð .°TÐ9Ðà'×,Ñ,Ó.‰ˆˆW�að —{‘{ =Ó1°D·L±LÑ@ˆñ ØÐ*Ø˜qÑ!×'Ñ'¨Ñ*Ð.>×.DÑ.DÀQÑ.GÒGð (¨Ñ*ˆJØ)¨!Ñ,ŠLÙàŸ™ T§[¡[Ð1AÓ%BÀBÈÓLˆJØŸ;™; t§{¡{Ð3CÓ'DÀbÈ#ÓN‰LØÐ'àŸ™ T§[¡[°Ó%?ÀÀSÓIˆJØŸ;™; t§{¡{°=Ó'AÀ2ÀsÓKˆLÜŸ™ N°1Ñ$5°zÐ#BÈÔJˆJÜ Ÿ9™9 n°QÑ&7¸Ð%FÈAÔN‰Lð Ÿ™ T§[¡[°Ó%?ÀÀSÓIˆJØŸ;™; t§{¡{°=Ó'AÀ2ÀsÓKˆLà�?Š?ð )¨,Ð7ˆNà˜DŸN™NÑ*¨B°·±Ð>ˆ
ØC�t—{‘{ <°¸#Ó>×CÑCÀZÐPˆØ'�Z×'Ñ'¨Ð4ˆ
Ø+�|×+Ñ+¨ZÐ8ˆà—/‘/ !Ó$ˆÜ—y‘y ¨z×/CÑ/CÀAÀqÓ/IÓJˆà×ÑÓ 3¨¯©Ñ#7¸À'Ð"JÒJÜØ6¸¸d¿n¹nÑ8LÈgÐW^Ð7_Ð6`ð aØ ×%Ñ%Ó'Ð(ð*óð ð
 Ð%Ø×"Ñ"Ó$¨¨a°¸'Ð(BÒBÜ Ø7¸¸aÀÈ'Ð8RÐ7SÐS\Ð]k×]pÑ]pÓ]rÐ\sÐtóð ð (×,Ñ,¨S°$·.±.À'È7ÓSÐVdÑdˆLØ'×,Ñ,¨S°4·>±>Ñ-AÀ7ÈGÓTˆLä—}‘}×,Ñ,¨\¸rÐ,ÓBˆàÐ&Ø×#Ñ#Ó%¨$¯.©.Ð):Ò:Ü ØEÀtÇ~Á~ÐFWÐEXð YØ'×,Ñ,Ó.Ð/ð1óð ð +×/Ñ/°°2°q¸!Ó<¸|×?PÑ?PÐQTÐVZ×VdÑVdÐfmÐovÓ?wÑwˆLØ'×,Ñ,¨S°4·>±>Ñ-AÀ7ÈGÓTˆLáð
 %1×$5Ñ$5°c¸4¿>¹>È7ÐT[Ó$\Ð!Ø0×5Ñ5°c¸D¿N¹NÑ6JÈGÐU\Ó]‰Là$(Ð!ä—]‘]×*Ñ*¨<¸4¿<¹<ÐRV×R_ÑR_Ð*Ó`ˆ
ä—i‘i 
¨LÓ9ˆà×ÑÓ #¨¯©Ñ"6¸ÀÇÁÐ!OÒOÜØ2°C¸$¿.¹.Ñ4HÈ'ÐSW×S`ÑS`Ð3aÐ2bð cØ×$Ñ$Ó&Ð'ð)óð ð
 "×&Ñ& s¨D¯N©N¸GÀTÇ]Á]ÓSˆØ!×+Ñ+¨A¨qÓ1ˆð "×)Ñ)¨#¨w¸¿¹ÓGˆà—m‘m KÓ0ˆàÐ1°>ÐAÐAr*   )ç        FTFN©NNNNF)rn   ro   rp   rÌ   r"   Úfloatr9   r   r   rW   ró   ÚTensorrè   r   rk   rq   rr   s   @r(   rÕ   rÕ   �  sM  ø„ ÙGð Ø ØØØ&*ñCàðCð ðCð ð	Cð
 ðCð ðCð ðCð ˜Ñ#õCð>e˜UŸ\™\ð e°Cð e¸có eð 48Ø8<Ø15Ø26Ø"'ñvBà—|‘|ðvBð # 5§<¡<Ñ0ðvBð !  u§|¡|Ñ!4Ñ5ð	vBð
 ! §¡Ñ.ðvBð " %§,¡,Ñ/ðvBð  ðvBð 
ˆu�|‰|˜X e§l¡lÑ3°X¸eÀEÇLÁLÑ>QÑ5RÐRÑ	S÷vBr*   rÕ   c                   óV  ‡ — e Zd ZdZˆ fd„Zdej                  dedefd„Z	 	 	 	 	 ddej                  de	ej                     d	e	e
ej                        d
e	ej                     de	ej                     dede
ej                  e	ej                     e	e
ej                        f   fd„Zˆ xZS )ÚSEWFlashAttention2aB  
    SEW flash attention module. This module inherits from `SEWAttention` as the weights of the module stays
    untouched. The only required change would be on the forward pass where it needs to correctly call the public API of
    flash attention and deal with padding tokens in case the input contains any of them.
    c                 óB   •— t        ‰| �  |i |¤Ž t        «       | _        y rh   )rV   rW   r   Ú_flash_attn_uses_top_left_mask)rc   ÚargsÚkwargsrf   s      €r(   rW   zSEWFlashAttention2.__init__5  s#   ø€ Ü‰Ñ˜$Ð) &Ò)ô
 /PÓ.QˆÕ+r*   râ   rã   r®   c                 óR   — |j                  ||| j                  | j                  «      S rh   )rå   r×   rÜ   rç   s       r(   Ú_reshapezSEWFlashAttention2._reshape=  s   € Ø�{‰{˜3 ¨¯©¸¿¹ÓGÐGr*   rj   ré   rê   r   rë   rì   r   c           
      óÎ  — |rt        d«      ‚|d u}|j                  «       \  }}	}
| j                  | j                  |«      d|«      }|rP|�N|d   j                  d   |j                  d   k(  r,|d   j                  dd«      }|d   j                  dd«      }�n*|rE| j                  | j                  |«      d|«      }| j                  | j                  |«      d|«      }nã|��| j                  | j                  |«      d|«      }| j                  | j                  |«      d|«      }t        j                  |d   j                  dd«      |gd¬«      }t        j                  |d   j                  dd«      |gd¬«      }nD| j                  | j                  |«      d|«      }| j                  | j                  |«      d|«      }| j                  r$|j                  dd«      |j                  dd«      f}|j                  d   }|�||d   j                  d   z  }|j                  }|t        j                  k(  rÂt        j                  «       rt        j                  «       }nMt        | j                   d«      r| j                   j"                  }n | j                  j$                  j                  }t&        j)                  d	|› d
�«       |j+                  |«      }|j+                  |«      }|j+                  |«      }t-        |||||	| j.                  r| j0                  nd| j2                  | j4                  ¬«      }|j7                  ||	d«      }| j9                  |«      }|sd }||fS )Nz?SEWFlashAttention2 attention does not support output_attentionsr+   r   r‡   r   rï   rz   Ú_pre_quantization_dtypez¾The input hidden states seems to be silently casted in float32, this might be related to the fact you have upcasted embedding or layer norm layers in float32. We will cast back the input in ú.r  )rØ   rÚ   Úuse_top_left_mask)r/   r­   r  rà   r   r{   rÞ   rß   ró   rô   rÙ   r-   Úfloat32Úis_autocast_enabledÚget_autocast_gpu_dtyper–   rd   r  r�   ÚloggerÚwarning_onceÚtor   rÇ   rØ   rÚ   r  rC   rá   )rc   rj   ré   rê   r   rë   rì   rø   r®   Úq_lenrF   rù   rú   rû   Ú
kv_seq_lenÚinput_dtypeÚtarget_dtyper   rý   s                      r(   rk   zSEWFlashAttention2.forward@  s0  € ñ ÜÐ^Ó_Ð_ð .°TÐ9Ðà%×*Ñ*Ó,‰ˆˆU�Að —}‘} T§[¡[°Ó%?ÀÀSÓIˆñ ØÐ*Ø˜qÑ!×'Ñ'¨Ñ*Ð.>×.DÑ.DÀQÑ.GÒGð (¨Ñ*×4Ñ4°Q¸Ó:ˆJØ)¨!Ñ,×6Ñ6°q¸!Ó<ŠLÙàŸ™ t§{¡{Ð3CÓ'DÀbÈ#ÓNˆJØŸ=™=¨¯©Ð5EÓ)FÈÈCÓP‰LØÐ'àŸ™ t§{¡{°=Ó'AÀ2ÀsÓKˆJØŸ=™=¨¯©°]Ó)CÀRÈÓMˆLÜŸ™ N°1Ñ$5×$?Ñ$?ÀÀ1Ó$EÀzÐ#RÐXYÔZˆJÜ Ÿ9™9 n°QÑ&7×&AÑ&AÀ!ÀQÓ&GÈÐ%VÐ\]Ô^‰Lð Ÿ™ t§{¡{°=Ó'AÀ2ÀsÓKˆJØŸ=™=¨¯©°]Ó)CÀRÈÓMˆLà�?Š?ð )×2Ñ2°1°aÓ8¸,×:PÑ:PÐQRÐTUÓ:VÐWˆNà×%Ñ% bÑ)ˆ
ØÐ%Ø˜.¨Ñ+×1Ñ1°"Ñ5Ñ5ˆJð #×(Ñ(ˆØœ%Ÿ-™-Ò'Ü×(Ñ(Ô*Ü$×;Ñ;Ó=‘ä˜Ÿ™Ð&?Ô@Ø#Ÿ{™{×BÑB‘à#Ÿ{™{×1Ñ1×7Ñ7�ä×Ñðà �> ð$ôð (Ÿ?™?¨<Ó8ˆLØ#Ÿ™ |Ó4ˆJØ'Ÿ?™?¨<Ó8ˆLä.ØØØØØØ$(§M¢M�D—L’L°sØ—n‘nØ"×AÑAô	
ˆð "×)Ñ)¨#¨u°bÓ9ˆØ—m‘m KÓ0ˆá ØˆLà˜L¨.Ð8Ð8r*   r  )rn   ro   rp   rÌ   rW   ró   r  r"   r  r   r   r9   rk   rq   rr   s   @r(   r  r  .  sæ   ø„ ñôRðH˜uŸ|™|ð H°cð HÀó Hð 48Ø8<Ø15Ø26Ø"'ñi9à—|‘|ði9ð # 5§<¡<Ñ0ði9ð !  u§|¡|Ñ!4Ñ5ð	i9ð
 ! §¡Ñ.ði9ð " %§,¡,Ñ/ði9ð  ði9ð 
ˆu�|‰|˜X e§l¡lÑ3°X¸eÀEÇLÁLÑ>QÑ5RÐRÑ	S÷i9r*   r  c                   ó$  ‡ — e Zd Z	 	 	 	 	 d	dej                  deej                     deeej                        deej                     deej                     dedeej                  eej                     eeej                        f   fˆ fd„Zˆ xZ	S )
ÚSEWSdpaAttentionrj   ré   rê   r   rë   rì   r   c                 óz  •— |s|�*t         j                  d«       t        ‰| �  ||||||¬«      S |du}|j	                  «       \  }}	}
| j                  |«      }|r0|�.|d   j                  d   |j                  d   k(  r|d   }|d   }�n
|rE| j                  | j                  |«      d|«      }| j                  | j                  |«      d|«      }nÃ|�}| j                  | j                  |«      d|«      }| j                  | j                  |«      d|«      }t        j                  |d   |gd¬«      }t        j                  |d   |gd¬«      }nD| j                  | j                  |«      d|«      }| j                  | j                  |«      d|«      }| j                  r||f}| j                  ||	|«      }| j                  r	|€|	dkD  rd	nd
}t        j                  j                  j!                  ||||| j"                  r| j$                  nd|¬«      }|j	                  «       || j&                  |	| j(                  fk7  r7t+        d|| j&                  |	| j(                  f› d|j	                  «       › �«      ‚|j-                  dd«      }|j/                  ||	| j0                  «      }| j3                  |«      }|d|fS )rî   Na¡  SEWModel is using SEWSdpaAttention, but `torch.nn.functional.scaled_dot_product_attention` does not support `output_attentions=True` or `layer_head_mask` not None. Falling back to the manual attention implementation, but specifying the manual implementation will be required from Transformers version v5.0.0 onwards. This warning can be removed using the argument `attn_implementation="eager"` when loading the model.)ré   rê   r   rë   rì   r   r‡   r   r+   rï   TFr  )Ú	attn_maskÚ	dropout_prÚ   rò   rð   )r  r  rV   rk   r­   rà   r   rè   rÞ   rß   ró   rô   rÙ   rÚ   r   rö   Úscaled_dot_product_attentionrÇ   rØ   r×   rÜ   r/   r{   rC   rÖ   rá   )rc   rj   ré   rê   r   rë   rì   rø   r®   r±   rF   rù   rú   rû   rÚ   r   rf   s                   €r(   rk   zSEWSdpaAttention.forward®  sÕ  ø€ ñ  Ð ;ä×Ñðlôô ‘7‘?ØØ!1Ø-Ø-Ø /Ø"3ð #ó ð ð .°TÐ9Ðà'×,Ñ,Ó.‰ˆˆW�að —{‘{ =Ó1ˆñ ØÐ*Ø˜qÑ!×'Ñ'¨Ñ*Ð.>×.DÑ.DÀQÑ.GÒGð (¨Ñ*ˆJØ)¨!Ñ,ŠLÙàŸ™ T§[¡[Ð1AÓ%BÀBÈÓLˆJØŸ;™; t§{¡{Ð3CÓ'DÀbÈ#ÓN‰LØÐ'àŸ™ T§[¡[°Ó%?ÀÀSÓIˆJØŸ;™; t§{¡{°=Ó'AÀ2ÀsÓKˆLÜŸ™ N°1Ñ$5°zÐ#BÈÔJˆJÜ Ÿ9™9 n°QÑ&7¸Ð%FÈAÔN‰Lð Ÿ™ T§[¡[°Ó%?ÀÀSÓIˆJØŸ;™; t§{¡{°=Ó'AÀ2ÀsÓKˆLà�?Š?ð )¨,Ð7ˆNà—{‘{ <°¸#Ó>ˆð
 !ŸNšN¨~Ð/EÈ'ÐTUÊ+‘DÐ[`ˆ	ô —h‘h×)Ñ)×FÑFØØØØ$Ø&*§m¢m�d—l’l¸Øð Gó 
ˆð ×ÑÓ # t§~¡~°wÀÇÁÐ!NÒNÜØ2°C¸¿¹ÈÐRV×R_ÑR_Ð3`Ð2að bØ×$Ñ$Ó&Ð'ð)óð ð
 "×+Ñ+¨A¨qÓ1ˆð "×)Ñ)¨#¨w¸¿¹ÓGˆà—m‘m KÓ0ˆà˜D .Ð0Ð0r*   r  )
rn   ro   rp   ró   r  r   r   r9   rk   rq   rr   s   @r(   r  r  ¬  s¿   ø„ ð
 48Ø8<Ø15Ø26Ø"'ñf1à—|‘|ðf1ð # 5§<¡<Ñ0ðf1ð !  u§|¡|Ñ!4Ñ5ð	f1ð
 ! §¡Ñ.ðf1ð " %§,¡,Ñ/ðf1ð  ðf1ð 
ˆu�|‰|˜X e§l¡lÑ3°X¸eÀEÇLÁLÑ>QÑ5RÐRÑ	S÷f1ñ f1r*   r  )ÚeagerÚsdpaÚflash_attention_2c                   ó$   ‡ — e Zd Zˆ fd„Zd„ Zˆ xZS )ÚSEWFeedForwardc                 óö  •— t         ‰| �  «        t        j                  |j                  «      | _        t        j                  |j                  |j                  «      | _	        t        |j                  t        «      rt        |j                     | _        n|j                  | _        t        j                  |j                  |j                  «      | _        t        j                  |j                   «      | _        y rh   )rV   rW   r   ÚDropoutÚactivation_dropoutÚintermediate_dropoutr©   r‘   Úintermediate_sizeÚintermediate_denseÚ
isinstanceÚ
hidden_actÚstrr	   Úintermediate_act_fnÚoutput_denseÚhidden_dropoutÚoutput_dropoutr«   s     €r(   rW   zSEWFeedForward.__init__   s«   ø€ Ü‰ÑÔÜ$&§J¡J¨v×/HÑ/HÓ$IˆÔ!ä"$§)¡)¨F×,>Ñ,>À×@XÑ@XÓ"YˆÔÜ�f×'Ñ'¬Ô-Ü'-¨f×.?Ñ.?Ñ'@ˆDÕ$à'-×'8Ñ'8ˆDÔ$äŸI™I f×&>Ñ&>À×@RÑ@RÓSˆÔÜ Ÿj™j¨×)>Ñ)>Ó?ˆÕr*   c                 ó°   — | j                  |«      }| j                  |«      }| j                  |«      }| j                  |«      }| j	                  |«      }|S rh   )r+  r/  r)  r0  r2  ri   s     r(   rk   zSEWFeedForward.forward-  sX   € Ø×/Ñ/°Ó>ˆØ×0Ñ0°Ó?ˆØ×1Ñ1°-Ó@ˆà×)Ñ)¨-Ó8ˆØ×+Ñ+¨MÓ:ˆØÐr*   rm   rr   s   @r(   r%  r%    s   ø„ ô@ör*   r%  c                   ó&   ‡ — e Zd Zˆ fd„Zdd„Zˆ xZS )ÚSEWEncoderLayerc                 óÈ  •— t         ‰| �  «        t        |j                     |j                  |j
                  |j                  d¬«      | _        t        j                  |j                  «      | _        t        j                  |j                  |j                  ¬«      | _        t        |«      | _        t        j                  |j                  |j                  ¬«      | _        y )NF)rÖ   r×   rØ   rÙ   ©Úeps)rV   rW   ÚSEW_ATTENTION_CLASSESÚ_attn_implementationr‘   Únum_attention_headsÚattention_dropoutÚ	attentionr   r'  r1  rØ   rw   Úlayer_norm_epsrx   r%  Úfeed_forwardÚfinal_layer_normr«   s     €r(   rW   zSEWEncoderLayer.__init__9  s¥   ø€ Ü‰ÑÔÜ.¨v×/JÑ/JÑKØ×(Ñ(Ø×0Ñ0Ø×,Ñ,Øô	
ˆŒô —z‘z &×"7Ñ"7Ó8ˆŒÜŸ,™, v×'9Ñ'9¸v×?TÑ?TÔUˆŒÜ*¨6Ó2ˆÔÜ "§¡¨V×-?Ñ-?ÀV×EZÑEZÔ [ˆÕr*   c                 óè   — |}| j                  |||¬«      \  }}}| j                  |«      }||z   }| j                  |«      }|| j                  |«      z   }| j	                  |«      }|f}|r||fz  }|S )N©r   rì   )r=  rØ   rx   r?  r@  )rc   rj   r   rì   Úattn_residualrý   rF   Úoutputss           r(   rk   zSEWEncoderLayer.forwardG  s’   € Ø%ˆØ)-¯©Ø¨.ÐL]ð *8ó *
Ñ&ˆ�| Qð Ÿ™ ]Ó3ˆØ%¨Ñ5ˆàŸ™¨Ó6ˆØ%¨×(9Ñ(9¸-Ó(HÑHˆØ×-Ñ-¨mÓ<ˆà Ð"ˆáØ˜�Ñ&ˆGàˆr*   rÀ   rm   rr   s   @r(   r5  r5  8  s   ø„ ô\÷r*   r5  c                   ó.   ‡ — e Zd Zˆ fd„Z	 	 	 	 dd„Zˆ xZS )Ú
SEWEncoderc                 ó>  •— t         ‰| �  «        || _        t        |«      | _        t        j                  |j                  |j                  «      | _        t        j                  |j                  |j                  ¬«      | _        t        j                  |j                  «      | _        t        j                   t#        |j$                  «      D �cg c]  }t'        |«      ‘Œ c}«      | _        t+        |«      | _        d| _        |j0                  dk(  | _        y c c}w )Nr7  Fr#  )rV   rW   rd   r…   Úpos_conv_embedr   Ú	AvgPool1dr”   Úpoolrw   r‘   r>  rx   r'  r1  rØ   rº   r7   Únum_hidden_layersr5  Úlayersr§   Úupsampler¼   r:  Ú_use_flash_attention_2)rc   rd   rF   rf   s      €r(   rW   zSEWEncoder.__init__\  sÍ   ø€ Ü‰ÑÔØˆŒÜ8¸Ó@ˆÔÜ—L‘L ×!6Ñ!6¸×8MÑ8MÓNˆŒ	ÜŸ,™, v×'9Ñ'9¸v×?TÑ?TÔUˆŒÜ—z‘z &×"7Ñ"7Ó8ˆŒÜ—m‘mÄeÈF×LdÑLdÓFeÖ$fÀ¤_°VÕ%<Ò$fÓgˆŒÜ% fÓ-ˆŒØ&+ˆÔ#Ø&,×&AÑ&AÐEXÑ&XˆÕ#ùò %gs   ÃDc           	      óŠ  — |rdnd }|rdnd }|��´|j                  d«      j                  dd|j                  d   «      }| j                  rd|| <   |�d|v r|nd }�ngd|| <   |j	                  «       j                  d«      }	|	| j                  j                  z  }
|j                  d   | j                  j                  z  }t        j                  d||
j                  ¬«      j                  dd«      j                  |
j                  d   d«      }||
j                  dd«      k  j	                  «       }d|d d …d d d d …f   j                  |j                  ¬	«      z
  }|t        j                  |j                  «      j                   z  }|j                  |j                  d   d|j                  d   |j                  d   «      }|j                  d   }|j#                  dd«      }| j%                  |«      }| j'                  |«      }t!        |j)                  d«      |j)                  d«      «      }|d
d |…f   |d
d |…f   z   }|j#                  dd«      }| j+                  |«      }| j-                  |«      }t/        «       xs t1        | «      }| j2                  D ]£  }|r||fz   }t        j4                  g «      }| j6                  r|| j                  j8                  k  rdnd}|r|rG| j:                  r+| j6                  r| j=                  |j>                  |||«      }n ||||¬«      }|d   }|rd}|sŒ›|d   fz   }Œ¥ |r||fz   }| jA                  |«      }|j                  d   |k  r4tB        jD                  jG                  |ddd||j                  d   z
  f«      }|stI        d„ |||fD «       «      S tK        |||¬«      S )N© r+   r   r‡   r  r   ©Údeviceç      ð?r,   .TFrB  ©NNc              3   ó&   K  — | ]	  }|€Œ|–— Œ y ­wrh   rP  )Ú.0Úvs     r(   ú	<genexpr>z%SEWEncoder.forward.<locals>.<genexpr>Â  s   è ø€ Òm˜qÐ_`Ñ_lœÑmùs   ‚Š©Úlast_hidden_staterj   Ú
attentions)&Ú	unsqueezeÚrepeatr   rN  Úlongr5   rd   r”   ró   r;   rR  rå   Úexpandr  r-   ÚfinfoÚminr{   rH  rJ  r­   rx   rØ   r
   r   rL  r2   rÇ   Ú	layerdropr¼   rÈ   rÉ   rM  r   rö   ÚpadÚtupler   )rc   rj   r   rì   Úoutput_hidden_statesÚreturn_dictÚall_hidden_statesÚall_self_attentionsÚexpand_attention_maskrG   Úoutput_lengthsÚmax_encoder_lengthÚattention_idsÚn_input_timestepsÚposition_embeddingsÚpooled_hidden_statesÚ
min_lengthÚsynced_gpusr·   Údropout_probabilityÚskip_the_layerÚlayer_outputss                         r(   rk   zSEWEncoder.forwardh  sõ  € ñ #7™B¸DÐÙ$5™b¸4ÐàÑ%Ø$2×$<Ñ$<¸RÓ$@×$GÑ$GÈÈ1Èm×NaÑNaÐbcÑNdÓ$eÐ!Ø×*Ò*à8;�Ð4Ð4Ñ5à4BÐ4NÐSTÐXfÑSf¡Ðmq’ð 9<�Ð4Ð4Ñ5Ø!/×!4Ñ!4Ó!6× ;Ñ ;¸BÓ ?�à!.°$·+±+×2LÑ2LÑ!L�Ø%2×%8Ñ%8¸Ñ%;¸t¿{¹{×?YÑ?YÑ%YÐ"ä—L‘L Ð$6¸~×?TÑ?TÔUß‘T˜!˜R“[ß‘V˜N×0Ñ0°Ñ3°RÓ8ð ð
 #0°.×2EÑ2EÀbÈ!Ó2LÑ"L×!RÑ!RÓ!T�ð "% ~²a¸¸tÂQÐ6FÑ'G×'JÑ'JÐQ^×QdÑQdÐ'JÓ'eÑ!e�Ø!/´%·+±+¸m×>QÑ>QÓ2R×2VÑ2VÑ!V�Ø!/×!6Ñ!6Ø"×(Ñ(¨Ñ+¨Q°×0DÑ0DÀRÑ0HÈ.×J^ÑJ^Ð_aÑJbó"�ð *×/Ñ/°Ñ2Ðà%×/Ñ/°°1Ó5ˆØ"×1Ñ1°-Ó@ÐØ#Ÿy™y¨Ó7ÐÜÐ,×1Ñ1°"Ó5Ð7K×7PÑ7PÐQSÓ7TÓUˆ
Ø,¨S°+°:°+Ð-=Ñ>ÐATÐUXÐZeÐ[eÐZeÐUeÑAfÑfˆØ%×/Ñ/°°1Ó5ˆàŸ™¨Ó6ˆØŸ™ ]Ó3ˆä0Ó2ÒRÔ6LÈTÓ6Rˆà—[‘[ò 	PˆEÙ#Ø$5¸Ð8HÑ$HÐ!ô #(§*¡*¨R£.Ðà%)§]¢]Ð8KÈdÏkÉk×NcÑNcÒ8c™TÐjoˆNÙ!¡[à×.Ò.°4·=²=Ø$(×$EÑ$EØŸ™Ø%Ø&Ø)ó	%‘Mñ %*Ø%°nÐXiô%�Mð !.¨aÑ 0�áØ ,�â Ø&9¸]È1Ñ=MÐ<OÑ&OÑ#ð7	Pñ:  Ø 1°]Ð4DÑ DÐàŸ™ mÓ4ˆØ×Ñ˜qÑ!Ð$5Ò5ÜŸM™M×-Ñ-¨m¸aÀÀAÐGXÐ[h×[nÑ[nÐopÑ[qÑGqÐ=rÓsˆMáÜÑm ]Ð4EÐGZÐ$[ÔmÓmÐmÜØ+Ø+Ø*ô
ð 	
r*   )NFFTrm   rr   s   @r(   rF  rF  [  s   ø„ ô
Yð ØØ"Ø÷_
r*   rF  c                   ó|   — e Zd ZdZeZdZdZdZdZ	dZ
d„ Zdeej                  ef   fd„Zded	ej                  fd
„Zy)ÚSEWPreTrainedModelz†
    An abstract class to handle weights initialization and a simple interface for downloading and loading pretrained
    models.
    ÚsewrÊ   Tc           
      óš  — t        |t        «      r³t        j                  j	                  |j
                  j                  ddt        j                  d|j
                  j                  d   |j
                  j                  z  z  «      z  ¬«       t        j                  j                  |j
                  j                  d«       �nt        |t        j                  «      r=|j                  j                  j	                  d| j                  j                   ¬«       �n½t        |t        j"                  t        j$                  f«      rK|j                  j                  j'                  «        |j                  j                  j)                  d«       �nHt        |t        j*                  «      �r-t-        «       rðddl}t1        |d«      r|t1        |d	«      rp|j2                  j5                  |j6                  |j8                  gd¬
«      5  t        j                  j;                  |j                  j                  «       ddd«       n—|j2                  j5                  |j                  d¬
«      5  t        j                  j;                  |j                  j                  «       ddd«       n3t        j                  j;                  |j                  j                  «       t        |t        j                  t        j*                  f«      r2|j                  �%|j                  j                  j'                  «        yyy# 1 sw Y   ŒfxY w# 1 sw Y   ŒrxY w)zInitialize the weightsr   r‡   r   )ÚmeanÚstdr  rS  Nr�   rœ   r‹   )r,  r…   r   ÚinitÚnormal_r_   r�   ÚmathÚsqrtrS   Úin_channelsÚ	constant_rU   r©   Údatard   Úinitializer_rangerw   r‚   Úzero_Úfill_r[   r
   r—   r–   r˜   r™   r�   rœ   Úkaiming_normal_)rc   Úmoduler—   s      r(   Ú_init_weightsz SEWPreTrainedModel._init_weights×  sB  € ä�fÔ8Ô9Ü�G‰G�O‰OØ—‘×"Ñ"ØØœŸ	™	 ! v§{¡{×'>Ñ'>¸qÑ'AÀFÇKÁK×D[ÑD[Ñ'[Ñ"\Ó]Ñ]ð ô ô
 �G‰G×Ñ˜fŸk™k×.Ñ.°Ö2Ü˜¤§	¡	Ô*ð �M‰M×Ñ×&Ñ&¨C°T·[±[×5RÑ5RÐ&ÖSÜ˜¤§¡¬r¯|©|Ð <Ô=Ø�K‰K×Ñ×"Ñ"Ô$Ø�M‰M×Ñ×$Ñ$ SÖ)Ü˜¤§	¡	Õ*Ü)Ô+Û ä˜6 :Ô.´7¸6À:Ô3NØ"Ÿ™×:Ñ:¸F¿O¹OÈVÏ_É_Ð;]ÐmnÐ:Óoñ DÜŸ™×/Ñ/°·±×0BÑ0BÔC÷Dð Dð #Ÿ™×:Ñ:¸6¿=¹=ÐXYÐ:ÓZñ DÜŸ™×/Ñ/°·±×0BÑ0BÔC÷Dð Dô —‘×'Ñ'¨¯©×(:Ñ(:Ô;ä�fœrŸy™y¬"¯)©)Ð4Ô5¸&¿+¹+Ð:QØ�K‰K×Ñ×"Ñ"Õ$ð ;RÐ5÷Dð Dú÷Dð Dús   È4L5É(4MÌ5L>ÍM
rG   c                 ó˜   — d„ }t        | j                  j                  | j                  j                  «      D ]  \  }} ||||«      }Œ |S )zH
        Computes the output length of the convolutional layers
        c                 ó>   — t        j                  | |z
  |d¬«      dz   S )NÚfloor)Úrounding_moder   )ró   Údiv)r$   rS   rT   s      r(   Ú_conv_out_lengthzMSEWPreTrainedModel._get_feat_extract_output_lengths.<locals>._conv_out_lengthü  s"   € ô —9‘9˜\¨KÑ7¸ÈwÔWÐZ[Ñ[Ð[r*   )Úziprd   r\   r]   )rc   rG   r�  rS   rT   s        r(   Ú _get_feat_extract_output_lengthsz3SEWPreTrainedModel._get_feat_extract_output_lengths÷  sQ   € ò
	\ô
 $' t§{¡{×'>Ñ'>ÀÇÁ×@WÑ@WÓ#Xò 	QÑˆK˜Ù,¨]¸KÈÓP‰Mð	Qð Ðr*   Úfeature_vector_lengthr   c                 óä  — | j                  |j                  d«      «      j                  t        j                  «      }|j
                  d   }t        j                  ||f|j                  |j                  ¬«      }d|t        j                  |j
                  d   |j                  ¬«      |dz
  f<   |j                  dg«      j                  d«      j                  dg«      j                  «       }|S )Nr+   r   )r-   rR  r   rQ  )r�  r5   r  ró   r^  r   r8   r-   rR  r;   ÚflipÚcumsumr9   )rc   r�  r   rj  rE   s        r(   Ú"_get_feature_vector_attention_maskz5SEWPreTrainedModel._get_feature_vector_attention_mask  s×   € Ø×>Ñ>¸~×?QÑ?QÐRTÓ?UÓV×YÑYÔZ_×ZdÑZdÓeˆØ#×)Ñ)¨!Ñ,ˆ
äŸ™ØÐ.Ð/°~×7KÑ7KÐTb×TiÑTiô
ˆð uvˆœŸ™ ^×%9Ñ%9¸!Ñ%<À^×EZÑEZÔ[Ð]kÐnoÑ]oÐpÑqØ'×,Ñ,¨b¨TÓ2×9Ñ9¸"Ó=×BÑBÀBÀ4ÓH×MÑMÓOˆØÐr*   N)rn   ro   rp   rÌ   r   Úconfig_classÚbase_model_prefixÚmain_input_nameÚsupports_gradient_checkpointingÚ_supports_flash_attn_2Ú_supports_sdpar‡  r   ró   Ú
LongTensorr"   r�  r”  rP  r*   r(   rv  rv  Ê  sg   „ ñð
 €LØÐØ$€OØ&*Ð#Ø!ÐØ€Nò%ð@¸eÀE×DTÑDTÐVYÐDYÑ>Zó ð
Èð 
Ð]b×]mÑ]mô 
r*   rv  aó  
    SEW was proposed in [Performance-Efficiency Trade-offs in Unsupervised Pre-training for Speech
    Recognition](https://arxiv.org/abs/2109.06870) by Felix Wu, Kwangyoun Kim, Jing Pan, Kyu Han, Kilian Q. Weinberger,
    Yoav Artzi.

    This model inherits from [`PreTrainedModel`]. Check the superclass documentation for the generic methods the
    library implements for all its model (such as downloading or saving etc.).

    This model is a PyTorch [torch.nn.Module](https://pytorch.org/docs/stable/nn.html#torch.nn.Module) sub-class. Use
    it as a regular PyTorch Module and refer to the PyTorch documentation for all matter related to general usage and
    behavior.

    Parameters:
        config ([`SEWConfig`]): Model configuration class with all the parameters of the model.
            Initializing with a config file does not load the weights associated with the model, only the
            configuration. Check out the [`~PreTrainedModel.from_pretrained`] method to load the model weights.
aØ  
    Args:
        input_values (`torch.FloatTensor` of shape `(batch_size, sequence_length)`):
            Float values of input raw speech waveform. Values can be obtained by loading a `.flac` or `.wav` audio file
            into an array of type `List[float]` or a `numpy.ndarray`, *e.g.* via the soundfile library (`pip install
            soundfile`). To prepare the array into `input_values`, the [`AutoProcessor`] should be used for padding and
            conversion into a tensor of type `torch.FloatTensor`. See [`Wav2Vec2Processor.__call__`] for details.
        attention_mask (`torch.LongTensor` of shape `(batch_size, sequence_length)`, *optional*):
            Mask to avoid performing convolution and attention on padding token indices. Mask values selected in `[0,
            1]`:

            - 1 for tokens that are **not masked**,
            - 0 for tokens that are **masked**.

            [What are attention masks?](../glossary#attention-mask)

        output_attentions (`bool`, *optional*):
            Whether or not to return the attentions tensors of all attention layers. See `attentions` under returned
            tensors for more detail.
        output_hidden_states (`bool`, *optional*):
            Whether or not to return the hidden states of all layers. See `hidden_states` under returned tensors for
            more detail.
        return_dict (`bool`, *optional*):
            Whether or not to return a [`~utils.ModelOutput`] instead of a plain tuple.
z]The bare SEW Model transformer outputting raw hidden-states without any specific head on top.c                   óV  ‡ — e Zd Zdefˆ fd„Z	 	 ddej                  deej                     deej                     fd„Z	 e
e«       eeeede¬«      	 	 	 	 	 dd	eej"                     deej"                     deej                     d
ee   dee   dee   deeef   fd„«       «       Zˆ xZS )ÚSEWModelrd   c                 óÄ  •— t         ‰| �  |«       || _        t        |«      | _        t        j                  |j                  d   |j                  ¬«      | _	        |j                  d   |j                  k7  | _        | j                  r2t        j                  |j                  d   |j                  «      | _        t        j                  |j                  «      | _        |j"                  dkD  s|j$                  dkD  rEt        j&                  t)        j*                  |j                  «      j-                  «       «      | _        t1        |«      | _        | j5                  «        y )Nr+   r7  r  )rV   rW   rd   r´   Úfeature_extractorr   rw   rX   r>  rx   r‘   Úproject_featuresr©   Úfeature_projectionr'  Úfeat_proj_dropoutÚfeature_dropoutÚmask_time_probÚmask_feature_probÚ	Parameterró   r  Úuniform_Úmasked_spec_embedrF  ÚencoderÚ	post_initr«   s     €r(   rW   zSEWModel.__init__F  sÿ   ø€ Ü‰Ñ˜Ô ØˆŒÜ!2°6Ó!:ˆÔÜŸ,™, v§¡°rÑ':À×@UÑ@UÔVˆŒà &§¡°Ñ 3°v×7IÑ7IÑ IˆÔØ× Ò Ü&(§i¡i°·±ÀÑ0CÀV×EWÑEWÓ&XˆDÔ#Ü!Ÿz™z¨&×*BÑ*BÓCˆÔà× Ñ  3Ò&¨&×*BÑ*BÀSÒ*HÜ%'§\¡\´%·,±,¸v×?QÑ?QÓ2R×2[Ñ2[Ó2]Ó%^ˆDÔ"ä! &Ó)ˆŒð 	�‰Õr*   rj   Úmask_time_indicesr   c                 óÎ  — t        | j                  dd«      s|S |j                  «       \  }}}|�)| j                  j	                  |j
                  «      ||<   nË| j                  j                  dkD  r²| j                  r¦t        ||f| j                  j                  | j                  j                  || j                  j                  ¬«      }t        j                  ||j                  t        j                  ¬«      }| j                  j	                  |j
                  «      ||<   | j                  j                  dkD  r¨| j                  rœt        ||f| j                  j                  | j                  j                   | j                  j"                  ¬«      }t        j                  ||j                  t        j                  ¬«      }|dd…df   j%                  d|d«      }d||<   |S )	zš
        Masks extracted features along time axis and/or along feature axis according to
        [SpecAugment](https://arxiv.org/abs/1904.08779).
        Úapply_spec_augmentTNr   )r   r   r   r   )rR  r-   )r   r   r   r+   )Úgetattrrd   r­   r¨  r  r-   r¤  rÇ   rN   Úmask_time_lengthÚmask_time_min_masksró   râ   rR  r9   r¥  Úmask_feature_lengthÚmask_feature_min_masksr_  )rc   rj   r«  r   rE   r'   r‘   Úmask_feature_indicess           r(   Ú_mask_hidden_stateszSEWModel._mask_hidden_statesZ  sš  € ô �t—{‘{Ð$8¸$Ô?Ø Ð ð 4A×3EÑ3EÓ3GÑ0ˆ
�O [àÐ(à/3×/EÑ/E×/HÑ/HÈ×I\ÑI\Ó/]ˆMÐ+Ò,Ø�[‰[×'Ñ'¨!Ò+°·²Ü 5Ø˜_Ð-ØŸ+™+×4Ñ4Ø ŸK™K×8Ñ8Ø-ØŸ+™+×9Ñ9ô!Ðô !&§¡Ð->À}×G[ÑG[Ôch×cmÑcmÔ nÐØ/3×/EÑ/E×/HÑ/HÈ×I\ÑI\Ó/]ˆMÐ+Ñ,à�;‰;×(Ñ(¨1Ò,°·²ä#8Ø˜[Ð)ØŸ+™+×7Ñ7Ø ŸK™K×;Ñ;ØŸ+™+×<Ñ<ô	$Ð ô $)§<¡<Ð0DÈ]×MaÑMaÔin×isÑisÔ#tÐ Ø#7º¸4¸Ñ#@×#GÑ#GÈÈOÐ]_Ó#`Ð Ø23ˆMÐ.Ñ/àÐr*   Úaudio)Ú
checkpointÚoutput_typer•  ÚmodalityÚexpected_outputrÊ   rì   re  rf  r   c                 óZ  — |�|n| j                   j                  }|�|n| j                   j                  }|�|n| j                   j                  }| j	                  |«      }|j                  dd«      }| j                  |«      }| j                  r| j                  |«      }| j                  |«      }|�| j                  |j                  d   |«      }| j                  ||¬«      }| j                  |||||¬«      }	|	d   }|s	|f|	dd  z   S t        ||	j                  |	j                   ¬«      S )Nr   r‡   )r«  ©r   rì   re  rf  r   rY  )rd   rì   re  Úuse_return_dictrŸ  r{   rx   r   r¡  r£  r”  r   r´  r©  r   rj   r[  )
rc   rÊ   r   r«  rì   re  rf  Úextract_featuresrj   Úencoder_outputss
             r(   rk   zSEWModel.forwardˆ  sU  € ð" 2CÐ1NÑ-ÐTX×T_ÑT_×TqÑTqÐà$8Ð$DÑ È$Ï+É+×JjÑJjð 	ð &1Ð%<‘kÀ$Ç+Á+×B]ÑB]ˆà×1Ñ1°,Ó?ÐØ+×5Ñ5°a¸Ó;ÐØŸ?™?Ð+;Ó<Ðà× Ò Ø#×6Ñ6Ð7GÓHÐØ×,Ñ,Ð-=Ó>ˆàÐ%à!×DÑDÀ]×EXÑEXÐYZÑE[Ð]kÓlˆNà×0Ñ0°ÐRcÐ0ÓdˆàŸ,™,ØØ)Ø/Ø!5Ø#ð 'ó 
ˆð (¨Ñ*ˆáØ!Ð# o°a°bÐ&9Ñ9Ð9äØ+Ø)×7Ñ7Ø&×1Ñ1ô
ð 	
r*   rT  ©NNNNN)rn   ro   rp   r   rW   ró   ÚFloatTensorr   r›  r´  r   ÚSEW_INPUTS_DOCSTRINGr   Ú_CHECKPOINT_FOR_DOCr   Ú_CONFIG_FOR_DOCÚ_EXPECTED_OUTPUT_SHAPEr  r9   r   r   rk   rq   rr   s   @r(   r�  r�  A  s  ø„ ð
˜yõ ð. :>Ø59ñ	,à×(Ñ(ð,ð $ E×$5Ñ$5Ñ6ð,ð ! ×!1Ñ!1Ñ2ó	,ñ\ +Ð+?Ó@ÙØ&Ø#Ø$ØØ.ôð 26Ø9=Ø,0Ø/3Ø&*ñ.
à˜uŸ|™|Ñ,ð.
ð ! §¡Ñ.ð.
ð $ E×$5Ñ$5Ñ6ð	.
ð
 $ D™>ð.
ð ' t™nð.
ð ˜d‘^ð.
ð 
ˆu�oÐ%Ñ	&ò.
óó Aô.
r*   r�  zaSEW Model with a `language modeling` head on top for Connectionist Temporal Classification (CTC).c                   ó  ‡ — e Zd Zddee   fˆ fd„Zd„ Zd„ Zd„ Zd„ Z	 e
e«       eeeeee¬«      	 	 	 	 	 ddeej&                     d	eej&                     d
ee   dee   dee   deej&                     deeef   fd„«       «       Zˆ xZS )Ú	SEWForCTCÚtarget_langc                 ó®  •— t         ‰| �  |«       t        |«      | _        t	        j
                  |j                  «      | _        || _        |j                  €t        d| j                  › d�«      ‚t        |d«      r|j                  r|j                  n|j                  }t	        j                   ||j                  «      | _        | j%                  «        y )NzYou are trying to instantiate zò with a configuration that does not define the vocabulary size of the language model head. Please instantiate the model as follows: `SEWForCTC.from_pretrained(..., vocab_size=vocab_size)`. or define `vocab_size` of your model's configuration.Úadd_adapter)rV   rW   r�  rw  r   r'  Úfinal_dropoutrØ   rÇ  Ú
vocab_sizer/   rf   r–   rÉ  Úoutput_hidden_sizer‘   r©   Úlm_headrª  )rc   rd   rÇ  rÌ  rf   s       €r(   rW   zSEWForCTC.__init__Ç  s½   ø€ Ü‰Ñ˜Ô ä˜FÓ#ˆŒÜ—z‘z &×"6Ñ"6Ó7ˆŒà&ˆÔà×ÑÐ$ÜØ0°·±Ð0@ð AHð Hóð ô *1°¸Ô)GÈF×L^ÒL^ˆF×%Ò%Ðdj×dvÑdvð 	ô —y‘yÐ!3°V×5FÑ5FÓGˆŒð 	�‰Õr*   c                 óö   — | j                   }|�&t        | j                  dd«      €t        d|› d�«      ‚|€-t        | j                  dd«      �t        j                  d«       y|�| j                  |d¬«       yy)a'  
        This method overwrites [`~PreTrainedModel.tie_weights`] so that adapter weights can be correctly loaded when
        passing `target_lang=...` to `from_pretrained(...)`.

        This method is **not** supposed to be called by the user and is prone to be changed in the future.
        NÚadapter_attn_dimzCannot pass `target_lang`: z- if `config.adapter_attn_dim` is not defined.z)By default `target_lang` is set to 'eng'.T)Ú
force_load)rÇ  r®  rd   r/   r  ÚinfoÚload_adapter)rc   rÇ  s     r(   Útie_weightszSEWForCTC.tie_weightsÞ  sƒ   € ð ×&Ñ&ˆàÐ"¤w¨t¯{©{Ð<NÐPTÓ'UÐ']ÜÐ:¸;¸-ÐGtÐuÓvÐvØÐ ¤W¨T¯[©[Ð:LÈdÓ%SÐ%_Ü�K‰KÐCÕDØÐ$Ø×Ñ˜k°dÐÕ;ð %r*   c                 óX   — t        j                  dt        «       | j                  «        y)ú¨
        Calling this function will disable the gradient computation for the feature encoder so that its parameter will
        not be updated during training.
        úžThe method `freeze_feature_extractor` is deprecated and will be removed in Transformers v5. Please use the equivalent `freeze_feature_encoder` method instead.N©rÐ   rÑ   rÓ   Úfreeze_feature_encoder©rc   s    r(   Úfreeze_feature_extractorz"SEWForCTC.freeze_feature_extractoró  ó'   € ô
 	�‰ðQäô	
ð
 	×#Ñ#Õ%r*   c                 óL   — | j                   j                  j                  «        y©rÕ  N©rw  rŸ  rÅ   rÙ  s    r(   rØ  z SEWForCTC.freeze_feature_encoderÿ  ó   € ð
 	�‰×"Ñ"×5Ñ5Õ7r*   c                 óP   — | j                   j                  «       D ]	  }d|_        Œ y©zÒ
        Calling this function will disable the gradient computation for the base model so that its parameters will not
        be updated during training. Only the classification head will be updated.
        FN©rw  rÁ   rÂ   rÃ   s     r(   Úfreeze_base_modelzSEWForCTC.freeze_base_model  ó(   € ð
 —X‘X×(Ñ(Ó*ò 	(ˆEØ"'ˆEÕñ	(r*   )r¶  r·  r•  r¹  Úexpected_lossrÊ   r   rì   re  rf  Úlabelsr   c           
      ó¤  — |�|n| j                   j                  }|�I|j                  «       | j                   j                  k\  r"t	        d| j                   j                  › �«      ‚| j                  |||||¬«      }|d   }| j                  |«      }| j                  |«      }	d}
|��b|�|n$t        j                  |t        j                  ¬«      }| j                  |j                  d«      «      j                  t        j                  «      }|dk\  }|j                  d«      }|j                  |«      }t        j                   j#                  |	dt        j$                  ¬«      j'                  dd«      }t        j(                  j*                  j-                  d	¬
«      5  t        j                   j/                  ||||| j                   j0                  | j                   j2                  | j                   j4                  ¬«      }
ddd«       |s|	f|t6        d z   }|
�|
f|z   S |S t9        |
|	|j:                  |j<                  ¬«      S # 1 sw Y   ŒExY w)aà  
        labels (`torch.LongTensor` of shape `(batch_size, target_length)`, *optional*):
            Labels for connectionist temporal classification. Note that `target_length` has to be smaller or equal to
            the sequence length of the output logits. Indices are selected in `[-100, 0, ..., config.vocab_size - 1]`.
            All labels set to `-100` are ignored (masked), the loss is only computed for labels in `[0, ...,
            config.vocab_size - 1]`.
        Nz$Label values must be <= vocab_size: r»  r   r,   r+   )r�   r-   r   F)Úenabled)ÚblankÚ	reductionÚzero_infinity©ÚlossÚlogitsrj   r[  )rd   r¼  r#   rË  r/   rw  rØ   rÍ  ró   Ú	ones_liker^  r�  r5   r  Úmasked_selectr   rö   Úlog_softmaxr  r{   ÚbackendsÚcudnnÚflagsÚctc_lossÚpad_token_idÚctc_loss_reductionÚctc_zero_infinityÚ_HIDDEN_STATES_START_POSITIONr   rj   r[  )rc   rÊ   r   rì   re  rf  ræ  rD  rj   rî  rí  rG   Úlabels_maskÚtarget_lengthsÚflattened_targetsÚ	log_probsÚoutputs                    r(   rk   zSEWForCTC.forward  s'  € ð0 &1Ð%<‘kÀ$Ç+Á+×B]ÑB]ˆàÐ &§*¡*£,°$·+±+×2HÑ2HÒ"HÜÐCÀDÇKÁK×DZÑDZÐC[Ð\Ó]Ð]à—(‘(ØØ)Ø/Ø!5Ø#ð ó 
ˆð   ™
ˆØŸ™ ]Ó3ˆà—‘˜mÓ,ˆàˆØÑð #1Ð"<‘Ä%Ç/Á/ÐR^Ôfk×fpÑfpÔBqð ð !×AÑAÀ.×BTÑBTÐUWÓBXÓY×\Ñ\Ô]b×]gÑ]gÓhˆMð ! A™+ˆKØ(Ÿ_™_¨RÓ0ˆNØ &× 4Ñ 4°[Ó AÐô Ÿ™×1Ñ1°&¸bÌÏÉÐ1ÓV×`Ñ`ÐabÐdeÓfˆIä—‘×%Ñ%×+Ñ+°EÐ+Ó:ñ 	Ü—}‘}×-Ñ-ØØ%Ø!Ø"ØŸ+™+×2Ñ2Ø"Ÿk™k×<Ñ<Ø"&§+¡+×"?Ñ"?ð .ó �÷	ñ Ø�Y Ô)FÐ)GÐ!HÑHˆFØ)-Ð)9�T�G˜fÑ$ÐE¸vÐEäØ˜f°G×4IÑ4IÐV]×VhÑVhô
ð 	
÷	ð 	ús   ÆA#IÉIrh   r¿  )rn   ro   rp   r   r.  rW   rÓ  rÚ  rØ  rã  r   rÁ  r   rÂ  r   rÃ  Ú_CTC_EXPECTED_OUTPUTÚ_CTC_EXPECTED_LOSSró   r  r9   r   r   rk   rq   rr   s   @r(   rÆ  rÆ  Á  sí   ø„ ñ¨H°S©Mõ ò.<ò*
&ò8ò(ñ +Ð+?Ó@ÙØ&Ø"Ø$Ø,Ø(ôð 26Ø,0Ø/3Ø&*Ø)-ñD
à˜uŸ|™|Ñ,ðD
ð ! §¡Ñ.ðD
ð $ D™>ð	D
ð
 ' t™nðD
ð ˜d‘^ðD
ð ˜Ÿ™Ñ&ðD
ð 
ˆu�nÐ$Ñ	%òD
óó AôD
r*   rÆ  z’
    SEW Model with a sequence classification head on top (a linear layer over the pooled output) for tasks like SUPERB
    Keyword Spotting.
    c                   ó  ‡ — e Zd Zˆ fd„Zd„ Zd„ Zd„ Z ee«       e	e
eedee¬«      	 	 	 	 	 ddeej"                     deej"                     d	ee   d
ee   dee   deej"                     deeef   fd„«       «       Zˆ xZS )ÚSEWForSequenceClassificationc                 óü  •— t         ‰| �  |«       t        |d«      r|j                  rt	        d«      ‚t        |«      | _        |j                  dz   }|j                  r0t        j                  t        j                  |«      |z  «      | _        t        j                  |j                  |j                   «      | _        t        j                  |j                   |j$                  «      | _        | j)                  «        y )NrÉ  zZSequence classification does not support the use of SEW adapters (config.add_adapter=True)r   )rV   rW   r–   rÉ  r/   r�  rw  rK  Úuse_weighted_layer_sumr   r¦  ró   r>   Úlayer_weightsr©   r‘   Úclassifier_proj_sizeÚ	projectorÚ
num_labelsÚ
classifierrª  )rc   rd   Ú
num_layersrf   s      €r(   rW   z%SEWForSequenceClassification.__init__f  sÀ   ø€ Ü‰Ñ˜Ô ä�6˜=Ô)¨f×.@Ò.@ÜØlóð ô ˜FÓ#ˆŒØ×-Ñ-°Ñ1ˆ
Ø×(Ò(Ü!#§¡¬e¯j©j¸Ó.DÀzÑ.QÓ!RˆDÔÜŸ™ 6×#5Ñ#5°v×7RÑ7RÓSˆŒÜŸ)™) F×$?Ñ$?À×ARÑARÓSˆŒð 	�‰Õr*   c                 óX   — t        j                  dt        «       | j                  «        y)z©
        Calling this function will disable the gradient computation for the feature encoder so that its parameters will
        not be updated during training.
        rÖ  Nr×  rÙ  s    r(   rÚ  z5SEWForSequenceClassification.freeze_feature_extractorw  rÛ  r*   c                 óL   — | j                   j                  j                  «        yrÝ  rÞ  rÙ  s    r(   rØ  z3SEWForSequenceClassification.freeze_feature_encoderƒ  rß  r*   c                 óP   — | j                   j                  «       D ]	  }d|_        Œ yrá  râ  rÃ   s     r(   rã  z.SEWForSequenceClassification.freeze_base_modelŠ  rä  r*   rµ  )r¶  r·  r•  r¸  r¹  rå  rÊ   r   rì   re  rf  ræ  r   c                 ó<  — |�|n| j                   j                  }| j                   j                  rdn|}| j                  |||||¬«      }| j                   j                  rr|t           }t        j                  |d¬«      }t        j                  j                  | j                  d¬«      }	||	j                  ddd«      z  j                  d¬«      }n|d   }| j                  |«      }|€|j                  d¬«      }
n‰| j                  |j                   d   |«      }|j#                  d«      j%                  dd|j                   d   «      }d	|| <   |j                  d¬«      |j                  d¬«      j                  dd«      z  }
| j'                  |
«      }d}|�Ft)        «       } ||j                  d| j                   j*                  «      |j                  d«      «      }|s|f|t        d z   }|�|f|z   S |S t-        |||j.                  |j0                  ¬
«      S )a�  
        labels (`torch.LongTensor` of shape `(batch_size,)`, *optional*):
            Labels for computing the sequence classification/regression loss. Indices should be in `[0, ...,
            config.num_labels - 1]`. If `config.num_labels == 1` a regression loss is computed (Mean-Square loss), If
            `config.num_labels > 1` a classification loss is computed (Cross-Entropy).
        NTr»  r   rï   r+   r   r‡   r  rì  )rd   r¼  r  rw  rù  ró   Ústackr   rö   r÷   r  rå   r5   r  ry  r”  r   r\  r]  r	  r   r  r   rj   r[  )rc   rÊ   r   rì   re  rf  ræ  rD  rj   Únorm_weightsÚpooled_outputÚpadding_maskÚexpand_padding_maskrî  rí  Úloss_fctrþ  s                    r(   rk   z$SEWForSequenceClassification.forward’  s  € ð2 &1Ð%<‘kÀ$Ç+Á+×B]ÑB]ˆØ'+§{¡{×'IÒ'I™tÐOcÐà—(‘(ØØ)Ø/Ø!5Ø#ð ó 
ˆð �;‰;×-Ò-Ø#Ô$AÑBˆMÜ!ŸK™K¨¸1Ô=ˆMÜŸ=™=×0Ñ0°×1CÑ1CÈÐ0ÓLˆLØ*¨\×->Ñ->¸rÀ1ÀaÓ-HÑH×MÑMÐRSÐMÓT‰Mà# A™JˆMàŸ™ }Ó5ˆØÐ!Ø)×.Ñ.°1Ð.Ó5‰Mà×BÑBÀ=×CVÑCVÐWXÑCYÐ[iÓjˆLØ".×"8Ñ"8¸Ó"<×"CÑ"CÀAÀqÈ-×J]ÑJ]Ð^_ÑJ`Ó"aÐØ25ˆMÐ.Ð.Ñ/Ø)×-Ñ-°!Ð-Ó4°|×7GÑ7GÈAÐ7GÓ7N×7SÑ7SÐTVÐXYÓ7ZÑZˆMà—‘ Ó/ˆàˆØÐÜ'Ó)ˆHÙ˜FŸK™K¨¨D¯K©K×,BÑ,BÓCÀVÇ[Á[ÐQSÃ_ÓUˆDáØ�Y Ô)FÐ)GÐ!HÑHˆFØ)-Ð)9�T�G˜fÑ$ÐE¸vÐEä'ØØØ!×/Ñ/Ø×)Ñ)ô	
ð 	
r*   r¿  )rn   ro   rp   rW   rÚ  rØ  rã  r   rÁ  r   Ú_SEQ_CLASS_CHECKPOINTr   rÃ  Ú_SEQ_CLASS_EXPECTED_OUTPUTÚ_SEQ_CLASS_EXPECTED_LOSSr   ró   r  r9   r   r   rk   rq   rr   s   @r(   r  r  ]  sØ   ø„ ôò"
&ò8ò(ñ +Ð+?Ó@ÙØ(Ø,Ø$ØØ2Ø.ôð 26Ø,0Ø/3Ø&*Ø)-ñ<
à˜uŸ|™|Ñ,ð<
ð ! §¡Ñ.ð<
ð $ D™>ð	<
ð
 ' t™nð<
ð ˜d‘^ð<
ð ˜Ÿ™Ñ&ð<
ð 
ˆuÐ.Ð.Ñ	/ò<
óó Aô<
r*   r  )rÆ  r  r�  rv  r¥   )MrÌ   r}  rÐ   Útypingr   r   r   Únumpyr0   ró   Útorch.utils.checkpointr   Útorch.nnr   Úactivationsr	   Úintegrations.deepspeedr
   Úintegrations.fsdpr   Úmodeling_flash_attention_utilsr   r   Úmodeling_outputsr   r   r   Úmodeling_utilsr   r•   r   r   r   r   Úconfiguration_sewr   r   Ú
get_loggerrn   r  rù  rÃ  rÂ  rÄ  rÿ  r   r  r  r  r"   r  r›  ÚndarrayrN   ÚModulerP   rt   r}   r…   rŸ   r§   r´   rÎ   rÕ   r  r  r9  r%  r5  rF  rv  ÚSEW_START_DOCSTRINGrÁ  r�  rÆ  r  Ú__all__rP  r*   r(   ú<module>r(     s¶  ðñ ã Û ß )Ñ )ã Û Û Ý Ý %å !Ý @Ý 7ß hß YÑ YÝ -÷ó õ )ñ ÔÝJð 
ˆ×	Ñ	˜HÓ	%€ð !"Ð ð €ð 6Ð Ú&Ð ð fð ð Ð ð CÐ Ø*Ð ØÐ ð 26ØñtØ��c�‰?ðtàðtð ðtð ˜U×-Ñ-Ñ.ð	tð
 ðtð ‡Z�Zótôp˜bŸi™iô ô,˜BŸI™Iô ô8˜BŸI™Iô ô0( §¡ô (ôX�b—i‘iô ô�B—I‘Iô ô.)˜Ÿ	™	ô )ôX
Ð+ô 
ô[B�2—9‘9ô [Bô~{9˜ô {9ô|h1�|ô h1ðX ØØ+ñÐ ô�R—Y‘Yô ô2 �b—i‘iô  ôFl
�—‘ô l
ô^F˜ô FðRÐ ð&Ð ñ6 ØcØóôy
Ð!ó y
ó	ðy
ñx ØkØóô
T
Ð"ó T
óð
T
ñn ðð óôr
Ð#5ó r
óðr
òj Z�r*   