Ë
    T^(hÎ ã                   ó´  — d dl Z d dlmZmZ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 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mZ ddl m!Z!  e«       rddlm"Z"  ejF                  e$«      Z%dZ&dZ' G d„ de	jP                  «      Z) G d„ de	jP                  «      Z* G d„ de	jP                  «      Z+ G d„ de	jP                  «      Z, G d„ de	jP                  «      Z- G d„ de	jP                  «      Z. G d„ de	jP                  «      Z/ G d„ d e	jP                  «      Z0 G d!„ d"e0«      Z1 G d#„ d$e0«      Z2 G d%„ d&e	jP                  «      Z3e0e2e1d'œZ4 G d(„ d)e	jP                  «      Z5 G d*„ d+e	jP                  «      Z6 G d,„ d-e	jP                  «      Z7 G d.„ d/e	jP                  «      Z8 G d0„ d1e	jP                  «      Z9 G d2„ d3e«      Z:	 	 dLd4ee;e;f   d5e<d6e;d7eejz                     d8e;d9ej|                  fd:„Z?d;Z@d<ZA ed=e@«       G d>„ d?e:«      «       ZBdZCd@ZDdAZE edBe@«       G dC„ dDe:«      «       ZFdEZGdFZHdGZI edHe@«       G dI„ dJe:«      «       ZJg dK¢ZKy)Mé    N)ÚOptionalÚTupleÚUnion)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loggingÚreplace_return_docstringsé   )ÚHubertConfig)Ú_flash_attention_forwardzfacebook/hubert-large-ls960-ftr   c                   ó$   ‡ — e Zd Zˆ fd„Zd„ Zˆ xZS )ÚHubertPositionalConvEmbeddingc                 ó  •— t         ‰| �  «        t        j                  |j                  |j                  |j
                  |j
                  dz  |j                  ¬«      | _        d | _        |j                  r&t        j                  |j                  «      | _        �n¯t        j                  j                  }t        t        j                  j                  d«      r$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"                  j1                  | |«       |j"                  j1                  | |«       n || j                  dd¬«      | _        t3        |j
                  «      | _        t6        |j8                     | _        y # 1 sw Y   �Œ'xY w)	Né   )Úkernel_sizeÚpaddingÚgroupsÚweight_normr   ©Úmodifier_rankÚweight)ÚnameÚdimÚparametrizations)ÚsuperÚ__init__ÚnnÚConv1dÚhidden_sizeÚnum_conv_pos_embeddingsÚnum_conv_pos_embedding_groupsÚconvÚ
batch_normÚconv_pos_batch_normÚBatchNorm1dÚutilsr    Úhasattrr&   r	   Ú	deepspeedÚzeroÚGatheredParametersr#   Ú	original0Ú	original1Úweight_gÚweight_vÚregister_external_parameterÚHubertSamePadLayerr   r   Úfeat_extract_activationÚ
activation)ÚselfÚconfigr    r4   r9   r:   Ú	__class__s         €úh/var/www/skyplay_api_hub/venv/lib/python3.12/site-packages/transformers/models/hubert/modeling_hubert.pyr(   z&HubertPositionalConvEmbedding.__init__-   sË  ø€ Ü‰ÑÔÜ—I‘IØ×ÑØ×ÑØ×6Ñ6Ø×2Ñ2°aÑ7Ø×7Ñ7ô
ˆŒ	ð ˆŒØ×%Ò%Ü Ÿn™n¨V×-?Ñ-?Ó@ˆDŽOäŸ(™(×.Ñ.ˆKÜ”r—x‘x×0Ñ0°-Ô@Ü Ÿh™h×7Ñ7×CÑC�ä)Õ+Û à—^‘^×6Ñ6°t·y±y×7GÑ7GÐWXÐ6ÓYñ MÙ +¨D¯I©I¸HÈ!Ô L�D”I÷Mä˜4Ÿ9™9Ð&8Ô9Ø#Ÿy™y×9Ñ9×@Ñ@×JÑJ�HØ#Ÿy™y×9Ñ9×@Ñ@×JÑJ‘Hà#Ÿy™y×1Ñ1�HØ#Ÿy™y×1Ñ1�HØ—‘×:Ñ:¸4ÀÔJØ—‘×:Ñ:¸4ÀÕJá'¨¯	©	¸ÀaÔH�”	ä)¨&×*HÑ*HÓIˆŒÜ  ×!?Ñ!?Ñ@ˆ�÷Mñ Mús   Ä?I?É?J	c                 óî   — |j                  dd«      }| j                  �| j                  |«      }| j                  |«      }| j                  |«      }| j	                  |«      }|j                  dd«      }|S ©Nr   r   )Ú	transposer/   r.   r   r>   ©r?   Úhidden_statess     rB   Úforwardz%HubertPositionalConvEmbedding.forwardR   sn   € Ø%×/Ñ/°°1Ó5ˆØ�?‰?Ð&Ø ŸO™O¨MÓ:ˆMØŸ	™	 -Ó0ˆØŸ™ ]Ó3ˆØŸ™¨Ó6ˆà%×/Ñ/°°1Ó5ˆØÐó    ©Ú__name__Ú
__module__Ú__qualname__r(   rH   Ú__classcell__©rA   s   @rB   r   r   ,   s   ø„ ô#AöJ	rI   r   c                   ó$   ‡ — e Zd Zˆ fd„Zd„ Zˆ xZS )r<   c                 óP   •— t         ‰| �  «        |dz  dk(  rd| _        y d| _        y )Nr   r   r   )r'   r(   Únum_pad_remove)r?   r,   rA   s     €rB   r(   zHubertSamePadLayer.__init___   s)   ø€ Ü‰ÑÔØ#:¸QÑ#>À!Ò#C˜aˆÕÈˆÕrI   c                 óV   — | j                   dkD  r|d d …d d …d | j                    …f   }|S ©Nr   )rR   rF   s     rB   rH   zHubertSamePadLayer.forwardc   s6   € Ø×Ñ Ò"Ø)ª!ªQÐ0F°4×3FÑ3FÐ2FÐ0FÐ*FÑGˆMØÐrI   rJ   rO   s   @rB   r<   r<   ^   s   ø„ ôKörI   r<   c                   ó&   ‡ — e Zd Zdˆ fd„	Zd„ Zˆ xZS )ÚHubertNoLayerNormConvLayerc                 ód  •— t         ‰| �  «        |dkD  r|j                  |dz
     nd| _        |j                  |   | _        t        j                  | j                  | j                  |j                  |   |j                  |   |j                  ¬«      | _
        t        |j                     | _        y )Nr   r   ©r   ÚstrideÚbias)r'   r(   Úconv_dimÚin_conv_dimÚout_conv_dimr)   r*   Úconv_kernelÚconv_strideÚ	conv_biasr.   r   r=   r>   ©r?   r@   Úlayer_idrA   s      €rB   r(   z#HubertNoLayerNormConvLayer.__init__j   s—   ø€ Ü‰ÑÔØ<DÀqºL˜6Ÿ?™?¨8°a©<Ò8ÈaˆÔØ"ŸO™O¨HÑ5ˆÔä—I‘IØ×ÑØ×ÑØ×*Ñ*¨8Ñ4Ø×%Ñ% hÑ/Ø×!Ñ!ô
ˆŒ	ô ! ×!?Ñ!?Ñ@ˆ�rI   c                 óJ   — | j                  |«      }| j                  |«      }|S ©N)r.   r>   rF   s     rB   rH   z"HubertNoLayerNormConvLayer.forwardx   s$   € ØŸ	™	 -Ó0ˆØŸ™¨Ó6ˆØÐrI   ©r   rJ   rO   s   @rB   rV   rV   i   s   ø„ õAörI   rV   c                   ó&   ‡ — e Zd Zdˆ fd„	Zd„ Zˆ xZS )ÚHubertLayerNormConvLayerc                 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   rX   T)Úelementwise_affine)r'   r(   r[   r\   r]   r)   r*   r^   r_   r`   r.   Ú	LayerNormÚ
layer_normr   r=   r>   ra   s      €rB   r(   z!HubertLayerNormConvLayer.__init__   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�rI   c                 ó´   — | j                  |«      }|j                  dd«      }| j                  |«      }|j                  dd«      }| j                  |«      }|S )Néþÿÿÿéÿÿÿÿ)r.   rE   rk   r>   rF   s     rB   rH   z HubertLayerNormConvLayer.forwardŽ   sV   € ØŸ	™	 -Ó0ˆà%×/Ñ/°°BÓ7ˆØŸ™¨Ó6ˆØ%×/Ñ/°°BÓ7ˆàŸ™¨Ó6ˆØÐrI   re   rJ   rO   s   @rB   rg   rg   ~   s   ø„ õAörI   rg   c                   ó&   ‡ — e Zd Zdˆ fd„	Zd„ Zˆ xZS )ÚHubertGroupNormConvLayerc                 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   rX   T)Ú
num_groupsÚnum_channelsÚaffine)r'   r(   r[   r\   r]   r)   r*   r^   r_   r`   r.   r   r=   r>   Ú	GroupNormrk   ra   s      €rB   r(   z!HubertGroupNormConvLayer.__init__š   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�rI   c                 ól   — | j                  |«      }| j                  |«      }| j                  |«      }|S rd   )r.   rk   r>   rF   s     rB   rH   z HubertGroupNormConvLayer.forwardª   s2   € ØŸ	™	 -Ó0ˆØŸ™¨Ó6ˆØŸ™¨Ó6ˆØÐrI   re   rJ   rO   s   @rB   rp   rp   ™   s   ø„ õrö rI   rp   c                   ó.   ‡ — e Zd ZdZˆ fd„Zd„ Zd„ Zˆ xZS )ÚHubertFeatureEncoderz.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   )rb   r   Úlayerz`config.feat_extract_norm` is z), but has to be one of ['group', 'layer']FT)r'   r(   Úfeat_extract_normrp   ÚrangeÚnum_feat_extract_layersrV   rg   Ú
ValueErrorr)   Ú
ModuleListÚconv_layersÚgradient_checkpointingÚ_requires_grad)r?   r@   Úir�   rA   s       €rB   r(   zHubertFeatureEncoder.__init__´   sê   ø€ Ü‰ÑÔà×#Ñ# wÒ.Ü3°FÀQÔGÐHÜLQÐRX×RpÑRpÐstÑRtÓLuöLØGHÔ*¨6¸AÀ¹EÖBòLñ ‰Kð ×%Ñ%¨Ò0ÜQVÐW]×WuÑWuÓQvÖwÈAÔ3°FÀQÖGÐwˆKÑwäØ0°×1IÑ1IÐ0JÐJsÐtóð ô Ÿ=™=¨Ó5ˆÔØ&+ˆÔ#Ø"ˆÕùòLùò xs   ÁC"Â	C'c                 óJ   — | j                  «       D ]	  }d|_        Œ d| _        y ©NF)Ú
parametersÚrequires_gradrƒ   ©r?   Úparams     rB   Ú_freeze_parametersz'HubertFeatureEncoder._freeze_parametersÅ   s(   € Ø—_‘_Ó&ò 	(ˆEØ"'ˆEÕð	(à#ˆÕrI   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)rƒ   Útrainingrˆ   r�   r‚   Ú_gradient_checkpointing_funcÚ__call__)r?   Úinput_valuesrG   Ú
conv_layers       rB   rH   zHubertFeatureEncoder.forwardÊ   s…   € Ø$¢Q¨ WÑ-ˆð ×Ò 4§=¢=Ø*.ˆMÔ'à×*Ñ*ò 	:ˆJØ×"Ò" t×'BÒ'BÀtÇ}Â}Ø $× AÑ AØ×'Ñ'Ø!ó!‘ñ
 !+¨=Ó 9‘ð	:ð ÐrI   )rK   rL   rM   Ú__doc__r(   r‹   rH   rN   rO   s   @rB   rx   rx   ±   s   ø„ Ù8ô#ò"$ö
rI   rx   c                   ó$   ‡ — e Zd Zˆ fd„Zd„ Zˆ xZS )ÚHubertFeatureProjectionc                 ón  •— t         ‰| �  «        |j                  | _        | j                  r3t        j                  |j
                  d   |j                  ¬«      | _        t        j                  |j
                  d   |j                  «      | _
        t        j                  |j                  «      | _        y )Nrn   ©Úeps)r'   r(   Úfeat_proj_layer_normr)   rj   r[   Úlayer_norm_epsrk   ÚLinearr+   Ú
projectionÚDropoutÚfeat_proj_dropoutÚdropout©r?   r@   rA   s     €rB   r(   z HubertFeatureProjection.__init__Þ   s}   ø€ Ü‰ÑÔØ$*×$?Ñ$?ˆÔ!Ø×$Ò$Ü Ÿl™l¨6¯?©?¸2Ñ+>ÀF×DYÑDYÔZˆDŒOÜŸ)™) F§O¡O°BÑ$7¸×9KÑ9KÓLˆŒÜ—z‘z &×":Ñ":Ó;ˆ�rI   c                 ó„   — | j                   r| j                  |«      }| j                  |«      }| j                  |«      }|S rd   )r˜   rk   r›   rž   rF   s     rB   rH   zHubertFeatureProjection.forwardæ   s;   € à×$Ò$Ø ŸO™O¨MÓ:ˆMØŸ™¨Ó6ˆØŸ™ ]Ó3ˆØÐrI   rJ   rO   s   @rB   r”   r”   Ý   s   ø„ ô<örI   r”   c                   ó†  ‡ — e Zd ZdZ	 	 	 	 	 ddededededededee   fˆ fd	„Z	d
e
j                  dedefd„Z	 	 	 	 	 dde
j                  dee
j                     deee
j                        dee
j                     dee
j                     dedee
j                  ee
j                     eee
j                        f   fd„Zˆ xZS )ÚHubertAttentionz=Multi-headed attention from 'Attention Is All You Need' paperÚ	embed_dimÚ	num_headsrž   Ú
is_decoderrZ   Ú	is_causalr@   c                 ó
  •— t         ‰| �  «        || _        || _        || _        ||z  | _        || _        | j
                  |z  | j                  k7  rt        d| j                  › d|› d�«      ‚| j
                  dz  | _        || _	        || _
        t        j                  |||¬«      | _        t        j                  |||¬«      | _        t        j                  |||¬«      | _        t        j                  |||¬«      | _        y )Nz;embed_dim must be divisible by num_heads (got `embed_dim`: z and `num_heads`: z).g      à¿)rZ   )r'   r(   r£   r¤   rž   Úhead_dimr@   r   Úscalingr¥   r¦   r)   rš   Úk_projÚv_projÚq_projÚout_proj)	r?   r£   r¤   rž   r¥   rZ   r¦   r@   rA   s	           €rB   r(   zHubertAttention.__init__ò   sä   ø€ ô 	‰ÑÔØ"ˆŒØ"ˆŒØˆŒØ! YÑ.ˆŒØˆŒà�M‰M˜IÑ%¨$¯.©.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�rI   ÚtensorÚseq_lenÚbszc                 óŽ   — |j                  ||| j                  | j                  «      j                  dd«      j	                  «       S rD   )Úviewr¤   r¨   rE   Ú
contiguous©r?   r®   r¯   r°   s       rB   Ú_shapezHubertAttention._shape  s7   € Ø�{‰{˜3 ¨¯©¸¿¹ÓG×QÑQÐRSÐUVÓW×bÑbÓdÐdrI   rG   Úkey_value_statesÚpast_key_valueÚattention_maskÚlayer_head_maskÚoutput_attentionsÚreturnc                 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$                  j)                  || j(                  | j*                  ¬«      }t        j                  ||«      }|j                  «       || j                  z  |	| j                  fk7  r9t!        d|| j                  z  |	| j                  f› d|j                  «       › �«      ‚|j                  || j                  |	| j                  «      }|j                  dd«      }|j                  ||	| j,                  «      }| j/                  |«      }|||fS )ú#Input shape: Batch x Time x ChannelNr   r   r   rn   ©r%   z$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pr�   ú `attn_output` should be of size )Úsizer¬   r©   Úshaperµ   rª   r«   ÚtorchÚcatr¥   r¤   r¨   r²   ÚreshapeÚbmmrE   r   r)   Ú
functionalÚsoftmaxrž   r�   r£   r­   )r?   rG   r¶   r·   r¸   r¹   rº   Úis_cross_attentionr°   Útgt_lenÚ_Úquery_statesÚ
key_statesÚvalue_statesÚ
proj_shapeÚsrc_lenÚattn_weightsÚattn_weights_reshapedÚ
attn_probsÚattn_outputs                       rB   rH   zHubertAttention.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ArI   )ç        FTFN©NNNNF)rK   rL   rM   r’   ÚintÚfloatÚboolr   r   r(   rÄ   ÚTensorrµ   r   rH   rN   rO   s   @rB   r¢   r¢   ï   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rI   r¢   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HubertFlashAttention2aH  
    Hubert flash attention module. This module inherits from `HubertAttention` 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 rd   )r'   r(   r   Ú_flash_attn_uses_top_left_mask)r?   ÚargsÚkwargsrA   s      €rB   r(   zHubertFlashAttention2.__init__”  s#   ø€ Ü‰Ñ˜$Ð) &Ò)ô
 /PÓ.QˆÕ+rI   r®   r¯   r°   c                 óR   — |j                  ||| j                  | j                  «      S rd   )r²   r¤   r¨   r´   s       rB   Ú_reshapezHubertFlashAttention2._reshapeœ  s   € Ø�{‰{˜3 ¨¯©¸¿¹ÓGÐGrI   rG   r¶   r·   r¸   r¹   rº   r»   c           
      óÎ  — |rt        d«      ‚|d u}|j                  «       \  }}	}
| j                  | j                  |«      d|«      }|rP|�N|d   j                  d   |j                  d   k(  r,|d   j                  dd«      }|d   j                  dd«      }�n*|rE| j                  | j                  |«      d|«      }| j                  | j                  |«      d|«      }nã|��| j                  | j                  |«      d|«      }| j                  | j                  |«      d|«      }t        j                  |d   j                  dd«      |gd¬«      }t        j                  |d   j                  dd«      |gd¬«      }nD| j                  | j                  |«      d|«      }| j                  | j                  |«      d|«      }| j                  r$|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 )NzBHubertFlashAttention2 attention does not support output_attentionsrn   r   r   r   r¾   rm   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.rÖ   )rž   r¦   Úuse_top_left_mask)r   rÂ   rã   r¬   rÃ   rE   rª   r«   rÄ   rÅ   r¥   ÚdtypeÚfloat32Úis_autocast_enabledÚget_autocast_gpu_dtyper3   r@   rå   r#   ÚloggerÚwarning_onceÚtor   r�   rž   r¦   rß   rÆ   r­   )r?   rG   r¶   r·   r¸   r¹   rº   rÊ   r°   Úq_lenrÌ   rÍ   rÎ   rÏ   Ú
kv_seq_lenÚinput_dtypeÚtarget_dtyperÕ   rÒ   s                      rB   rH   zHubertFlashAttention2.forwardŸ  s0  € ñ ÜÐaÓbÐbð .°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$(§M¢M�D—L’L°sØ—n‘nØ"×AÑAô	
ˆð "×)Ñ)¨#¨u°bÓ9ˆØ—m‘m KÓ0ˆá ØˆLà˜L¨.Ð8Ð8rI   r×   )rK   rL   rM   r’   r(   rÄ   rÛ   rØ   rã   r   r   rÚ   rH   rN   rO   s   @rB   rÝ   rÝ   �  sæ   ø„ ñôRðH˜uŸ|™|ð H°cð HÀó Hð 48Ø8<Ø15Ø26Ø"'ñi9à—|‘|ði9ð # 5§<¡<Ñ0ði9ð !  u§|¡|Ñ!4Ñ5ð	i9ð
 ! §¡Ñ.ði9ð " %§,¡,Ñ/ði9ð  ði9ð 
ˆu�|‰|˜X e§l¡lÑ3°X¸eÀEÇLÁLÑ>QÑ5RÐRÑ	S÷i9rI   rÝ   c                   ó$  ‡ — 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HubertSdpaAttentionrG   r¶   r·   r¸   r¹   rº   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 )r½   Na§  HubertModel is using HubertSdpaAttention, 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.)r¶   r·   r¸   r¹   rº   r   r   r   rn   r¾   TFrÖ   )Ú	attn_maskÚ	dropout_pr¦   rÁ   r¿   )rì   rí   r'   rH   rÂ   r¬   rÃ   rµ   rª   r«   rÄ   rÅ   r¥   r¦   r)   rÈ   Úscaled_dot_product_attentionr�   rž   r¤   r¨   r   rE   rÆ   r£   r­   )r?   rG   r¶   r·   r¸   r¹   rº   rÊ   r°   rË   rÌ   rÍ   rÎ   rÏ   r¦   rÕ   rA   s                   €rB   rH   zHubertSdpaAttention.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0rI   r×   )
rK   rL   rM   rÄ   rÛ   r   r   rÚ   rH   rN   rO   s   @rB   rô   rô     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1rI   rô   c                   ó$   ‡ — e Zd Zˆ fd„Zd„ Zˆ xZS )ÚHubertFeedForwardc                 óö  •— t         ‰| �  «        t        j                  |j                  «      | _        t        j                  |j                  |j                  «      | _	        t        |j                  t        «      rt        |j                     | _        n|j                  | _        t        j                  |j                  |j                  «      | _        t        j                  |j                   «      | _        y rd   )r'   r(   r)   rœ   Úactivation_dropoutÚintermediate_dropoutrš   r+   Úintermediate_sizeÚintermediate_denseÚ
isinstanceÚ
hidden_actÚstrr   Úintermediate_act_fnÚoutput_denseÚhidden_dropoutÚoutput_dropoutrŸ   s     €rB   r(   zHubertFeedForward.__init__v  s«   ø€ Ü‰ÑÔÜ$&§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rI   c                 ó°   — | j                  |«      }| j                  |«      }| j                  |«      }| j                  |«      }| j	                  |«      }|S rd   )rÿ   r  rý   r  r  rF   s     rB   rH   zHubertFeedForward.forwardƒ  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rI   rJ   rO   s   @rB   rú   rú   u  s   ø„ ô@örI   rú   )ÚeagerÚsdpaÚflash_attention_2c                   ó&   ‡ — e Zd Zˆ fd„Zdd„Zˆ xZS )ÚHubertEncoderLayerc                 óÈ  •— t         ‰| �  «        t        |j                     |j                  |j
                  |j                  d¬«      | _        t        j                  |j                  «      | _        t        j                  |j                  |j                  ¬«      | _        t        |«      | _        t        j                  |j                  |j                  ¬«      | _        y )NF©r£   r¤   rž   r¥   r–   )r'   r(   ÚHUBERT_ATTENTION_CLASSESÚ_attn_implementationr+   Únum_attention_headsÚattention_dropoutÚ	attentionr)   rœ   r  rž   rj   r™   rk   rú   Úfeed_forwardÚfinal_layer_normrŸ   s     €rB   r(   zHubertEncoderLayer.__init__•  s¥   ø€ Ü‰ÑÔÜ1°&×2MÑ2MÑNØ×(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-¨fÓ5ˆÔÜ "§¡¨V×-?Ñ-?ÀV×EZÑEZÔ [ˆÕrI   c                 óè   — |}| j                  |||¬«      \  }}}| j                  |«      }||z   }| j                  |«      }|| j                  |«      z   }| j	                  |«      }|f}|r||fz  }|S ©N©r¸   rº   )r  rž   rk   r  r  ©r?   rG   r¸   rº   Úattn_residualrÒ   rÌ   Úoutputss           rB   rH   zHubertEncoderLayer.forward£  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rI   r†   rJ   rO   s   @rB   r  r  ”  s   ø„ ô\÷rI   r  c                   ór   ‡ — e Zd Zˆ fd„Z	 	 	 	 ddej
                  deej                     dededef
d„Z	ˆ xZ
S )	ÚHubertEncoderc                 óÀ  •— t         ‰| �  «        || _        t        |«      | _        t        j                  |j                  |j                  ¬«      | _	        t        j                  |j                  «      | _        t        j                  t        |j                  «      D �cg c]  }t!        |«      ‘Œ c}«      | _        d| _        |j&                  dk(  | _        y c c}w ©Nr–   Fr
  )r'   r(   r@   r   Úpos_conv_embedr)   rj   r+   r™   rk   rœ   r  rž   r€   r}   Únum_hidden_layersr  Úlayersr‚   r  Ú_use_flash_attention_2©r?   r@   rÌ   rA   s      €rB   r(   zHubertEncoder.__init__¸  s¤   ø€ Ü‰ÑÔØˆŒÜ;¸FÓCˆÔÜŸ,™, v×'9Ñ'9¸v×?TÑ?TÔUˆŒÜ—z‘z &×"7Ñ"7Ó8ˆŒÜ—m‘mÌÈv×OgÑOgÓIhÖ$iÀAÔ%7¸Õ%?Ò$iÓjˆŒØ&+ˆÔ#Ø&,×&AÑ&AÐEXÑ&XˆÕ#ùò %jó   Â!CrG   r¸   rº   Úoutput_hidden_statesÚreturn_dictc                 ó4  — |rdnd }|rdnd }|�Ý|j                  d«      j                  dd|j                  d   «      }d|| <   | j                  r|�d|v r|nd }n‘d|d d …d d d d …f   j	                  |j
                  ¬«      z
  }|t        j                  |j
                  «      j                  z  }|j                  |j                  d   d|j                  d   |j                  d   «      }| j                  |«      }	||	z   }| j                  |«      }| j                  |«      }t        «       xs t        | «      }
| j                  D ]£  }|r||fz   }t        j                   g «      }| j"                  r|| j$                  j&                  k  rdnd	}|r|
rG| j(                  r+| j"                  r| j+                  |j,                  |||«      }n ||||¬
«      }|d   }|rd}|sŒ›|d   fz   }Œ¥ |r||fz   }|st/        d„ |||fD «       «      S t1        |||¬«      S )N© rn   r   r   r   ç      ð?©rè   TFr  ©NNc              3   ó&   K  — | ]	  }|€Œ|–— Œ y ­wrd   r)  ©Ú.0Úvs     rB   ú	<genexpr>z(HubertEncoder.forward.<locals>.<genexpr>  ó   è ø€ Òm˜qÐ_`Ñ_lœÑmùó   ‚Š©Úlast_hidden_staterG   Ú
attentions)Ú	unsqueezeÚrepeatrÃ   r#  rî   rè   rÄ   ÚfinfoÚminÚexpandr   rk   rž   r	   r
   r"  Úrandr�   r@   Ú	layerdropr‚   rŽ   r�   Útupler   ©r?   rG   r¸   rº   r&  r'  Úall_hidden_statesÚall_self_attentionsÚexpand_attention_maskÚposition_embeddingsÚsynced_gpusr{   Údropout_probabilityÚskip_the_layerÚlayer_outputss                  rB   rH   zHubertEncoder.forwardÂ  s[  € ñ #7™B¸DÐÙ$5™b¸4ÐàÐ%à$2×$<Ñ$<¸RÓ$@×$GÑ$GÈÈ1Èm×NaÑNaÐbcÑNdÓ$eÐ!Ø45ˆMÐ0Ð0Ñ1Ø×*Ò*à4BÐ4NÐSTÐXfÑSf¡Ðmq‘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 #×1Ñ1°-Ó@ÐØ%Ð(;Ñ;ˆØ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$EØŸ™Ø%Ø&Ø)ó	%‘Mñ %*Ø%°nÐXiô%�Mð !.¨aÑ 0�áØ ,�â Ø&9¸]È1Ñ=MÐ<OÑ&OÑ#ð7	Pñ:  Ø 1°]Ð4DÑ DÐáÜÑm ]Ð4EÐGZÐ$[ÔmÓmÐmÜØ+Ø+Ø*ô
ð 	
rI   ©NFFT)rK   rL   rM   r(   rÄ   r®   r   rÛ   rÚ   rH   rN   rO   s   @rB   r  r  ·  s_   ø„ ôYð 26Ø"'Ø%*Ø ñG
à—|‘|ðG
ð ! §¡Ñ.ðG
ð  ð	G
ð
 #ðG
ð ÷G
rI   r  c                   ó>   ‡ — e Zd Zˆ fd„Zdej
                  fd„Zˆ xZS )ÚHubertAttnAdapterLayerc                 óœ  •— t         ‰| �  «        |j                  | _        |j                  | _        t        j                  | j
                  «      | _        t        j                  | j
                  | j                  «      | _
        t        j                  «       | _        t        j                  | j                  | j
                  «      | _        y)zŸ
        Implements adapter modules directly with 3D tensor weight as parameters and without using ModuleList to speed
        up training throughput.
        N)r'   r(   Úadapter_attn_dimÚ	input_dimr+   Ú
hidden_dimr)   rj   Únormrš   Úlinear_1ÚReLUÚact_fnÚlinear_2rŸ   s     €rB   r(   zHubertAttnAdapterLayer.__init__  s   ø€ ô
 	‰ÑÔØ×0Ñ0ˆŒØ ×,Ñ,ˆŒä—L‘L §¡Ó1ˆŒ	ÜŸ	™	 $§/¡/°4·>±>ÓBˆŒÜ—g‘g“iˆŒÜŸ	™	 $§.¡.°$·/±/ÓBˆ�rI   rG   c                 óŽ   — | j                  |«      }| j                  |«      }| j                  |«      }| j                  |«      }|S rd   )rO  rP  rR  rS  rF   s     rB   rH   zHubertAttnAdapterLayer.forward  s@   € ØŸ	™	 -Ó0ˆàŸ™ mÓ4ˆØŸ™ MÓ2ˆØŸ™ mÓ4ˆàÐrI   )rK   rL   rM   r(   rÄ   ÚFloatTensorrH   rN   rO   s   @rB   rJ  rJ    s   ø„ ôCð U×%6Ñ%6÷ rI   rJ  c                   óf   ‡ — e Zd Zˆ fd„Z	 	 ddej
                  deej
                     defd„Zˆ xZ	S )Ú!HubertEncoderLayerStableLayerNormc                 ó  •— t         ‰| �  «        t        |j                     |j                  |j
                  |j                  d¬«      | _        t        j                  |j                  «      | _        t        j                  |j                  |j                  ¬«      | _        t        |«      | _        t        j                  |j                  |j                  ¬«      | _        t%        |dd «      �t'        |«      | _        y d | _        y )NFr  r–   rL  )r'   r(   r  r  r+   r  r  r  r)   rœ   r  rž   rj   r™   rk   rú   r  r  ÚgetattrrJ  Úadapter_layerrŸ   s     €rB   r(   z*HubertEncoderLayerStableLayerNorm.__init__&  sÊ   ø€ Ü‰ÑÔÜ1°&×2MÑ2MÑNØ×(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-¨fÓ5ˆÔÜ "§¡¨V×-?Ñ-?ÀV×EZÑEZÔ [ˆÔä�6Ð-¨tÓ4Ð@Ü!7¸Ó!?ˆDÕà!%ˆDÕrI   rG   r¸   rº   c                 ó$  — |}| j                  |«      }| j                  |||¬«      \  }}}| j                  |«      }||z   }|| j                  | j	                  |«      «      z   }| j
                  �|| j                  |«      z   }|f}|r||fz  }|S r  )rk   r  rž   r  r  rZ  r  s           rB   rH   z)HubertEncoderLayerStableLayerNorm.forward8  s±   € ð &ˆØŸ™¨Ó6ˆØ)-¯©Ø¨.ÐL]ð *8ó *
Ñ&ˆ�| Qð Ÿ™ ]Ó3ˆØ%¨Ñ5ˆØ%¨×(9Ñ(9¸$×:OÑ:OÐP]Ó:^Ó(_Ñ_ˆà×ÑÐ)Ø)¨D×,>Ñ,>¸}Ó,MÑMˆMà Ð"ˆáØ˜�Ñ&ˆGàˆrI   r†   )
rK   rL   rM   r(   rÄ   rÛ   r   rÚ   rH   rN   rO   s   @rB   rW  rW  %  s>   ø„ ô&ð* 26Ø"'ñ	à—|‘|ðð ! §¡Ñ.ðð  ÷	rI   rW  c                   ó.   ‡ — e Zd Zˆ fd„Z	 	 	 	 dd„Zˆ xZS )ÚHubertEncoderStableLayerNormc                 óÀ  •— t         ‰| �  «        || _        t        |«      | _        t        j                  |j                  |j                  ¬«      | _	        t        j                  |j                  «      | _        t        j                  t        |j                  «      D �cg c]  }t!        |«      ‘Œ c}«      | _        d| _        |j&                  dk(  | _        y c c}w r  )r'   r(   r@   r   r   r)   rj   r+   r™   rk   rœ   r  rž   r€   r}   r!  rW  r"  r‚   r  r#  r$  s      €rB   r(   z%HubertEncoderStableLayerNorm.__init__S  s©   ø€ Ü‰ÑÔØˆŒÜ;¸FÓCˆÔÜŸ,™, v×'9Ñ'9¸v×?TÑ?TÔUˆŒÜ—z‘z &×"7Ñ"7Ó8ˆŒÜ—m‘mÜ@EÀf×F^ÑF^Ó@_Ö`¸1Ô.¨vÕ6Ò`ó
ˆŒð ',ˆÔ#Ø&,×&AÑ&AÐEXÑ&XˆÕ#ùò ar%  c                 óf  — |rdnd }|rdnd }|�ö|j                  d«      j                  dd|j                  d   «      }||j                  |j                  ¬«      z  }| j
                  r|�d|v r|nd }n‘d|d d …d d d d …f   j                  |j                  ¬«      z
  }|t        j                  |j                  «      j                  z  }|j                  |j                  d   d|j                  d   |j                  d   «      }| j                  |«      }	||	z   }| j                  |«      }t        «       xs t        | «      }
| j                  D ]£  }|r||fz   }t        j                  g «      }| j                   r|| j"                  j$                  k  rdnd	}|r|
rG| j&                  r+| j                   r| j)                  |j*                  |||«      }n ||||¬
«      }|d   }|rd}|sŒ›|d   fz   }Œ¥ | j-                  |«      }|r||fz   }|st/        d„ |||fD «       «      S t1        |||¬«      S )Nr)  rn   r   r   r+  r   r*  TFr  r,  c              3   ó&   K  — | ]	  }|€Œ|–— Œ y ­wrd   r)  r.  s     rB   r1  z7HubertEncoderStableLayerNorm.forward.<locals>.<genexpr>£  r2  r3  r4  )r7  r8  rÃ   rî   rè   r#  rÄ   r9  r:  r;  r   rž   r	   r
   r"  r<  r�   r@   r=  r‚   rŽ   r�   rk   r>  r   r?  s                  rB   rH   z$HubertEncoderStableLayerNorm.forward_  sn  € ñ #7™B¸DÐÙ$5™b¸4ÐàÐ%à$2×$<Ñ$<¸RÓ$@×$GÑ$GÈÈ1Èm×NaÑNaÐbcÑNdÓ$eÐ!Ø)Ð,A×,DÑ,DÈ=×K^ÑK^Ð,DÓ,_Ñ_ˆMØ×*Ò*à4BÐ4NÐSTÐXfÑSf¡Ðmq‘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 #×1Ñ1°-Ó@ÐØ%Ð(;Ñ;ˆØŸ™ ]Ó3ˆä0Ó2ÒRÔ6LÈTÓ6Rˆà—[‘[ò 	PˆEÙ#Ø$5¸Ð8HÑ$HÐ!ô #(§*¡*¨R£.Ðà%)§]¢]Ð8KÈdÏkÉk×NcÑNcÒ8c™TÐjoˆNÙ!¡[ð ×.Ò.°4·=²=Ø$(×$EÑ$EØŸ™Ø%Ø&Ø)ó	%‘Mñ %*Ø%°nÐXiô%�Mð !.¨aÑ 0�áØ ,�â Ø&9¸]È1Ñ=MÐ<OÑ&OÑ#ð9	Pð< Ÿ™¨Ó6ˆáØ 1°]Ð4DÑ DÐáÜÑm ]Ð4EÐGZÐ$[ÔmÓmÐmÜØ+Ø+Ø*ô
ð 	
rI   rH  rJ   rO   s   @rB   r]  r]  R  s   ø„ ô
Yð ØØ"Ø÷I
rI   r]  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HubertPreTrainedModelz†
    An abstract class to handle weights initialization and a simple interface for downloading and loading pretrained
    models.
    Úhubertr�   Tc                 óz  — t        |t        j                  «      rm|j                  j                  j                  d| j                  j                  ¬«       |j                  �%|j                  j                  j                  «        yyt        |t        j                  t        j                  t        j                  f«      rJ|j                  j                  j                  «        |j                  j                  j                  d«       yt        |t        j                  «      �r_t        «       rðddl}t#        |d«      r|t#        |d«      rp|j$                  j'                  |j(                  |j*                  gd¬«      5  t        j,                  j/                  |j                  j                  «       ddd«       n—|j$                  j'                  |j                  d¬«      5  t        j,                  j/                  |j                  j                  «       ddd«       n3t        j,                  j/                  |j                  j                  «       |j                  �%|j                  j                  j                  «        yyt        |t0        «      r2t#        |d	«      r%|j2                  j                  j5                  «        yyt        |t6        «      rMt#        |d
«      r@|j8                  j                  j                  d| j                  j:                  dz   z  «       yyy# 1 sw Y   ŒÛxY w# 1 sw Y   ŒçxY w)zInitialize the weightsrÖ   )ÚmeanÚstdNr*  r   r:   r9   r!   Úmasked_spec_embedÚlayer_weightsr   )r   r)   rš   r#   ÚdataÚnormal_r@   Úinitializer_rangerZ   Úzero_rj   ru   r1   Úfill_r*   r	   r4   r3   r5   r6   r:   r9   ÚinitÚkaiming_normal_ÚHubertModelrg  Úuniform_ÚHubertForSequenceClassificationrh  r!  )r?   Úmoduler4   s      rB   Ú_init_weightsz#HubertPreTrainedModel._init_weights¸  sP  € ä�fœbŸi™iÔ(ð �M‰M×Ñ×&Ñ&¨C°T·[±[×5RÑ5RÐ&ÔSØ�{‰{Ð&Ø—‘× Ñ ×&Ñ&Õ(ð 'ä˜¤§¡¬r¯|©|¼R¿^¹^Ð LÔMØ�K‰K×Ñ×"Ñ"Ô$Ø�M‰M×Ñ×$Ñ$ SÕ)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;à�{‰{Ð&Ø—‘× Ñ ×&Ñ&Õ(ð 'ä˜¤Ô,Ü�vÐ2Ô3Ø×(Ñ(×-Ñ-×6Ñ6Õ8ð 4ä˜Ô ?Ô@Ü�v˜Ô/Ø×$Ñ$×)Ñ)×/Ñ/°°t·{±{×7TÑ7TÐWXÑ7XÑ0YÕZð 0ð A÷Dð Dú÷Dð Dús   Å?4L%Ç#4L1Ì%L.Ì1L:Úinput_lengthsc                 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   )rÄ   Údiv)Úinput_lengthr   rY   s      rB   Ú_conv_out_lengthzPHubertPreTrainedModel._get_feat_extract_output_lengths.<locals>._conv_out_lengthÞ  s"   € ô —9‘9˜\¨KÑ7¸ÈwÔWÐZ[Ñ[Ð[rI   )Úzipr@   r^   r_   )r?   ru  r|  r   rY   s        rB   Ú _get_feat_extract_output_lengthsz6HubertPreTrainedModel._get_feat_extract_output_lengthsÙ  sQ   € ò
	\ô
 $' t§{¡{×'>Ñ'>ÀÇÁ×@WÑ@WÓ#Xò 	QÑˆK˜Ù,¨]¸KÈÓP‰Mð	Qð ÐrI   Úfeature_vector_lengthr¸   c                 óä  — | j                  |j                  d«      «      j                  t        j                  «      }|j
                  d   }t        j                  ||f|j                  |j                  ¬«      }d|t        j                  |j
                  d   |j                  ¬«      |dz
  f<   |j                  dg«      j                  d«      j                  dg«      j                  «       }|S )Nrn   r   )rè   Údevicer   )r�  )r~  Úsumrî   rÄ   ÚlongrÃ   Úzerosrè   r�  ÚarangeÚflipÚcumsumrÚ   )r?   r  r¸   Úoutput_lengthsÚ
batch_sizes        rB   Ú"_get_feature_vector_attention_maskz8HubertPreTrainedModel._get_feature_vector_attention_maskè  s×   € Ø×>Ñ>¸~×?QÑ?QÐRTÓ?UÓV×YÑYÔZ_×ZdÑZdÓeˆØ#×)Ñ)¨!Ñ,ˆ
äŸ™ØÐ.Ð/°~×7KÑ7KÐTb×TiÑTiô
ˆð uvˆœŸ™ ^×%9Ñ%9¸!Ñ%<À^×EZÑEZÔ[Ð]kÐnoÑ]oÐpÑqØ'×,Ñ,¨b¨TÓ2×9Ñ9¸"Ó=×BÑBÀBÀ4ÓH×MÑMÓOˆØÐrI   N)rK   rL   rM   r’   r   Úconfig_classÚbase_model_prefixÚmain_input_nameÚsupports_gradient_checkpointingÚ_supports_flash_attn_2Ú_supports_sdpart  r   rÄ   Ú
LongTensorrØ   r~  rŠ  r)  rI   rB   rb  rb  «  sh   „ ñð
  €LØ ÐØ$€OØ&*Ð#Ø!ÐØ€Nò[ðB¸eÀE×DTÑDTÐVYÐDYÑ>Zó ð
Èð 
Ð]b×]mÑ]mô 
rI   rb  rÃ   Ú	mask_probÚmask_lengthr¸   Ú	min_masksr»   c                 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   )rØ   Úmax)r{  Únum_masked_spanÚepsilonr“  r’  r”  Úsequence_lengths     €€€€€rB   Úcompute_num_masked_spanz6_compute_mask_indices.<locals>.compute_num_masked_span  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rI   Nrn   r+  r   F)Úreplace)r   ÚnpÚrandomr<  ÚitemÚdetachr‚  Útolistr}   r„  rÚ   Úchoicer…  ÚlenÚconcatenateÚonesÚint32ÚappendÚarrayÚbroadcast_torÆ   r˜  Úput_along_axis)rÃ   r’  r“  r¸   r”  r‰  rœ  rÌ   ru  Úspec_aug_maskÚspec_aug_mask_idxsÚmax_num_masked_spanr{  r™  Úspec_aug_mask_idxÚdummy_mask_idxÚoffsetsrš  r›  s    `` `            @@rB   Ú_compute_mask_indicesr²  õ  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+a!  
    Hubert was proposed in [HuBERT: Self-Supervised Speech Representation Learning by Masked Prediction of Hidden
    Units](https://arxiv.org/abs/2106.07447) by Wei-Ning Hsu, Benjamin Bolte, Yao-Hung Hubert Tsai, Kushal Lakhotia,
    Ruslan Salakhutdinov, Abdelrahman Mohamed.

    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 ([`HubertConfig`]): 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.
aŸ  
    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)

            <Tip warning={true}>

            `attention_mask` should only be passed if the corresponding processor has `config.return_attention_mask ==
            True`. For all models whose processor has `config.return_attention_mask == False`, such as
            [hubert-base](https://huggingface.co/facebook/hubert-base-ls960), `attention_mask` should **not** be passed
            to avoid degraded performance when doing batched inference. For such models `input_values` should simply be
            padded with 0 and passed without `attention_mask`. Be aware that these models also yield slightly different
            results depending on whether `input_values` is padded or not.

            </Tip>

        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 Hubert Model transformer outputting raw hidden-states without any specific head on top.c                   óP  ‡ — 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¬«      	 	 	 	 	 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 )rp  r@   c                 ó²  •— t         ‰| �  |«       || _        t        |«      | _        t        |«      | _        |j                  dkD  s|j                  dkD  rEt        j                  t        j                  |j                  «      j                  «       «      | _        |j                   rt#        |«      | _        nt'        |«      | _        | j)                  «        y )NrÖ   )r'   r(   r@   rx   Úfeature_extractorr”   Úfeature_projectionÚmask_time_probÚmask_feature_probr)   Ú	ParameterrÄ   rÛ   r+   rq  rg  Údo_stable_layer_normr]  Úencoderr  Ú	post_initrŸ   s     €rB   r(   zHubertModel.__init__ª  s¤   ø€ Ü‰Ñ˜Ô ØˆŒÜ!5°fÓ!=ˆÔÜ"9¸&Ó"AˆÔð × Ñ  3Ò&¨&×*BÑ*BÀSÒ*HÜ%'§\¡\´%·,±,¸v×?QÑ?QÓ2R×2[Ñ2[Ó2]Ó%^ˆDÔ"à×&Ò&Ü7¸Ó?ˆD�Lä(¨Ó0ˆDŒLð 	�‰ÕrI   rG   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”  )r�  rè   )r’  r“  r”  rn   )rY  r@   rÂ   rg  rî   rè   r·  r�   r²  Úmask_time_lengthÚmask_time_min_masksrÄ   r®   r�  rÚ   r¸  Úmask_feature_lengthÚmask_feature_min_masksr;  )r?   rG   r½  r¸   r‰  r›  r+   Úmask_feature_indicess           rB   Ú_mask_hidden_stateszHubertModel._mask_hidden_states¼  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rI   )Úoutput_typer‹  r�   rº   r&  r'  r»   c                 óþ  — |�|n| j                   j                  }|�|n| j                   j                  }|�|n| j                   j                  }| j	                  |«      }|j                  dd«      }|�| j                  |j                  d   |«      }| j                  |«      }| j                  ||¬«      }| j                  |||||¬«      }	|	d   }|s	|f|	dd z   S t        ||	j                  |	j                  ¬«      S )aZ  

        Returns:

        Example:

        ```python
        >>> from transformers import AutoProcessor, HubertModel
        >>> from datasets import load_dataset
        >>> import soundfile as sf

        >>> processor = AutoProcessor.from_pretrained("facebook/hubert-large-ls960-ft")
        >>> model = HubertModel.from_pretrained("facebook/hubert-large-ls960-ft")


        >>> def map_to_array(batch):
        ...     speech, _ = sf.read(batch["file"])
        ...     batch["speech"] = speech
        ...     return batch


        >>> ds = load_dataset("hf-internal-testing/librispeech_asr_dummy", "clean", split="validation")
        >>> ds = ds.map(map_to_array)

        >>> input_values = processor(ds["speech"][0], return_tensors="pt").input_values  # Batch size 1
        >>> hidden_states = model(input_values).last_hidden_state
        ```Nr   r   )r½  ©r¸   rº   r&  r'  r   r4  )r@   rº   r&  Úuse_return_dictrµ  rE   rŠ  rÃ   r¶  rÅ  r»  r   rG   r6  )
r?   r�   r¸   r½  rº   r&  r'  Úextract_featuresrG   Úencoder_outputss
             rB   rH   zHubertModel.forwardê  s,  € ðL 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àÐ%à!×DÑDÐEU×E[ÑE[Ð\]ÑE^Ð`nÓoˆNà×/Ñ/Ð0@ÓAˆØ×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 	
rI   r,  ©NNNNN)rK   rL   rM   r   r(   rÄ   rU  r   r‘  rÅ  r   ÚHUBERT_INPUTS_DOCSTRINGr   r   Ú_CONFIG_FOR_DOCrÛ   rÚ   r   r   rH   rN   rO   s   @rB   rp  rp  ¥  s  ø„ ð
˜|õ ð* :>Ø59ñ	,à×(Ñ(ð,ð $ E×$5Ñ$5Ñ6ð,ð ! ×!1Ñ!1Ñ2ó	,ñ\ +Ð+BÓCÙ¨?ÈÔYð 26Ø9=Ø,0Ø/3Ø&*ñE
à˜uŸ|™|Ñ,ðE
ð ! §¡Ñ.ðE
ð $ E×$5Ñ$5Ñ6ð	E
ð
 $ D™>ðE
ð ' t™nðE
ð ˜d‘^ðE
ð 
ˆu�oÐ%Ñ	&òE
ó Zó DôE
rI   rp  z['MISTER QUILTER IS THE APOSTLE OF THE MIDDLE CLASSES AND WE ARE GLAD TO WELCOME HIS GOSPEL'g®Gáz®6@zdHubert 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HubertForCTCÚ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 zõ with a configuration that does not define the vocabulary size of the language model head. Please instantiate the model as follows: `HubertForCTC.from_pretrained(..., vocab_size=vocab_size)`. or define `vocab_size` of your model's configuration.Úadd_adapter)r'   r(   rp  rc  r)   rœ   Úfinal_dropoutrž   rÑ  Ú
vocab_sizer   rA   r3   rÓ  Úoutput_hidden_sizer+   rš   Úlm_headr¼  )r?   r@   rÑ  rÖ  rA   s       €rB   r(   zHubertForCTC.__init__@  s½   ø€ Ü‰Ñ˜Ô ä! &Ó)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rI   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.
        NrL  z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)rÑ  rY  r@   r   rì   ÚinfoÚload_adapter)r?   rÑ  s     rB   Útie_weightszHubertForCTC.tie_weightsW  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$Ø×Ñ˜k°dÐÕ;ð %rI   c                 óX   — t        j                  dt        «       | j                  «        y)ú¨
        Calling this function will disable the gradient computation for the feature encoder so that its parameter will
        not be updated during training.
        úžThe method `freeze_feature_extractor` is deprecated and will be removed in Transformers v5. Please use the equivalent `freeze_feature_encoder` method instead.N©ÚwarningsÚwarnÚFutureWarningÚfreeze_feature_encoder©r?   s    rB   Úfreeze_feature_extractorz%HubertForCTC.freeze_feature_extractorl  ó'   € ô
 	�‰ðQäô	
ð
 	×#Ñ#Õ%rI   c                 óL   — | j                   j                  j                  «        y©rÞ  N©rc  rµ  r‹   rå  s    rB   rä  z#HubertForCTC.freeze_feature_encoderx  ó   € ð
 	�‰×%Ñ%×8Ñ8Õ:rI   c                 óP   — | j                   j                  «       D ]	  }d|_        Œ y©zÒ
        Calling this function will disable the gradient computation for the base model so that its parameters will not
        be updated during training. Only the classification head will be updated.
        FN©rc  r‡   rˆ   r‰   s     rB   Úfreeze_base_modelzHubertForCTC.freeze_base_model  ó(   € ð
 —[‘[×+Ñ+Ó-ò 	(ˆEØ"'ˆEÕñ	(rI   )Ú
checkpointrÆ  r‹  Úexpected_outputÚexpected_lossr�   r¸   rº   r&  r'  Úlabelsr»   c           
      ó¤  — |�|n| j                   j                  }|�I|j                  «       | j                   j                  k\  r"t	        d| j                   j                  › �«      ‚| j                  |||||¬«      }|d   }| j                  |«      }| j                  |«      }	d}
|��b|�|n$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: rÈ  r   r+  rn   )r%   rè   r   F)Úenabled)ÚblankÚ	reductionÚzero_infinity©ÚlossÚlogitsrG   r6  )r@   rÉ  r˜  rÕ  r   rc  rž   r×  rÄ   Ú	ones_likerƒ  r~  r‚  rî   Úmasked_selectr)   rÈ   Úlog_softmaxré   rE   ÚbackendsÚcudnnÚflagsÚctc_lossÚpad_token_idÚctc_loss_reductionÚctc_zero_infinityÚ_HIDDEN_STATES_START_POSITIONr   rG   r6  )r?   r�   r¸   rº   r&  r'  rô  r  rG   rü  rû  ru  Úlabels_maskÚtarget_lengthsÚflattened_targetsÚ	log_probsÚoutputs                    rB   rH   zHubertForCTC.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$ÐE¸vÐEäØ˜f°G×4IÑ4IÐV]×VhÑVhô
ð 	
÷	ð 	ús   ÆA#IÉIrd   rÌ  )rK   rL   rM   r   r  r(   rÜ  ræ  rä  rï  r   rÍ  r   Ú_CHECKPOINT_FOR_DOCr   rÎ  Ú_CTC_EXPECTED_OUTPUTÚ_CTC_EXPECTED_LOSSrÄ   rÛ   rÚ   r   r   rH   rN   rO   s   @rB   rÐ  rÐ  ;  sí   ø„ ñ
¨H°S©Mõ ò.<ò*
&ò;ò(ñ +Ð+BÓCÙØ&Ø"Ø$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$Ñ	%òD
óó DôD
rI   rÐ  zsuperb/hubert-base-superb-ksz'_unknown_'g�Âõ(\!@z•
    Hubert Model with a sequence classification head on top (a linear layer over the pooled output) for tasks like
    SUPERB Keyword Spotting.
    c                   ó  ‡ — e Zd Zˆ fd„Zd„ Zd„ Zd„ Z ee«       e	e
eedee¬«      	 	 	 	 	 ddeej"                     deej"                     d	ee   d
ee   dee   deej"                     deeef   fd„«       «       Zˆ xZS )rr  c                 óü  •— t         ‰| �  |«       t        |d«      r|j                  rt	        d«      ‚t        |«      | _        |j                  dz   }|j                  r0t        j                  t        j                  |«      |z  «      | _        t        j                  |j                  |j                   «      | _        t        j                  |j                   |j$                  «      | _        | j)                  «        y )NrÓ  z]Sequence classification does not support the use of Hubert adapters (config.add_adapter=True)r   )r'   r(   r3   rÓ  r   rp  rc  r!  Úuse_weighted_layer_sumr)   r¹  rÄ   r¦  rh  rš   r+   Úclassifier_proj_sizeÚ	projectorÚ
num_labelsÚ
classifierr¼  )r?   r@   Ú
num_layersrA   s      €rB   r(   z(HubertForSequenceClassification.__init__ã  sÀ   ø€ Ü‰Ñ˜Ô ä�6˜=Ô)¨f×.@Ò.@ÜØoóð ô " &Ó)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rI   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.
        rß  Nrà  rå  s    rB   ræ  z8HubertForSequenceClassification.freeze_feature_extractorô  rç  rI   c                 óL   — | j                   j                  j                  «        yré  rê  rå  s    rB   rä  z6HubertForSequenceClassification.freeze_feature_encoder   rë  rI   c                 óP   — | j                   j                  «       D ]	  }d|_        Œ yrí  rî  r‰   s     rB   rï  z1HubertForSequenceClassification.freeze_base_model  rð  rI   Úaudio)rñ  rÆ  r‹  Úmodalityrò  ró  r�   r¸   rº   r&  r'  rô  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).
        NTrÈ  r   r¾   rn   r   r   rÖ   rú  )r@   rÉ  r  rc  r  rÄ   Ústackr)   rÈ   rÉ   rh  r²   r‚  r  re  rŠ  rÃ   r7  r8  r  r   r  r   rG   r6  )r?   r�   r¸   rº   r&  r'  rô  r  rG   Únorm_weightsÚpooled_outputÚpadding_maskÚexpand_padding_maskrü  rû  Úloss_fctr  s                    rB   rH   z'HubertForSequenceClassification.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$ÐE¸vÐEä'ØØØ!×/Ñ/Ø×)Ñ)ô	
ð 	
rI   rÌ  )rK   rL   rM   r(   ræ  rä  rï  r   rÍ  r   Ú_SEQ_CLASS_CHECKPOINTr   rÎ  Ú_SEQ_CLASS_EXPECTED_OUTPUTÚ_SEQ_CLASS_EXPECTED_LOSSr   rÄ   rÛ   rÚ   r   r   rH   rN   rO   s   @rB   rr  rr  Û  sØ   ø„ ôò"
&ò;ò(ñ +Ð+BÓCÙØ(Ø,Ø$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 Dô<
rI   rr  )rÐ  rr  rp  rb  rT   )Lrá  Útypingr   r   r   Únumpyrž  rÄ   Útorch.nnr)   r   Úactivationsr   Úintegrations.deepspeedr	   Úintegrations.fsdpr
   Úmodeling_flash_attention_utilsr   r   Úmodeling_outputsr   r   r   Úmodeling_utilsr   r2   r   r   r   r   r   Úconfiguration_hubertr   r   Ú
get_loggerrK   rì   r  rÎ  ÚModuler   r<   rV   rg   rp   rx   r”   r¢   rÝ   rô   rú   r  r  r  rJ  rW  r]  rb  rØ   rÙ   r‘  Úndarrayr²  ÚHUBERT_START_DOCSTRINGrÍ  rp  r  r  r  rÐ  r$  r%  r&  rr  Ú__all__r)  rI   rB   ú<module>r6     sÊ  ðó ß )Ñ )ã Û Ý Ý %å !Ý @Ý 7ß hß YÑ YÝ -÷õ õ /ñ ÔÝJð 
ˆ×	Ñ	˜HÓ	%€ð 7Ð ð !€ô/ B§I¡Iô /ôd˜Ÿ™ô ô §¡ô ô*˜rŸy™yô ô6˜rŸy™yô ô0)˜2Ÿ9™9ô )ôX˜bŸi™iô ô$[B�b—i‘iô [Bô|{9˜Oô {9ô|g1˜/ô g1ôT˜Ÿ	™	ô ð2 ØØ.ñÐ ô ˜Ÿ™ô  ôFR
�B—I‘Iô R
ôj˜RŸY™Yô ô2*¨¯	©	ô *ôZV
 2§9¡9ô V
ôrG˜Oô G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ðnÐ ð&#Ð ñL ØfØóôH
Ð'ó H
ó	ðH
ðV !"Ð ð uÐ ØÐ ñ ØnØóôT
Ð(ó T
ó	ðT
ðn 7Ð Ø*Ð ØÐ ñ ðð óôr
Ð&;ó r
óðr
òj f�rI   