Ë
    T^(hWH ã                   óJ  — d dl Z d dlZd dlm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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#jP                  e)«      Z*dZ+dZ,e G d„ de«      «       Z- G d„ dej\                  «      Z/ G d„ dej\                  «      Z0 G d„ dej\                  «      Z1 G d„ dej\                  «      Z2 G d„ dej\                  «      Z3 G d„ dej\                  «      Z4 G d „ d!ej\                  «      Z5 G d"„ d#ej\                  «      Z6 G d$„ d%e6«      Z7 G d&„ d'e6«      Z8 G d(„ d)ej\                  «      Z9e6e8e7d*œZ: G d+„ d,ej\                  «      Z; G d-„ d.ej\                  «      Z< G d/„ d0ej\                  «      Z= G d1„ d2ej\                  «      Z> G d3„ d4ej\                  «      Z? G d5„ d6ej\                  «      Z@ G d7„ d8e«      ZA	 	 dTd9eeBeBf   d:eCd;eBd<ee
jˆ                     d=eBd>e	jŠ                  fd?„ZFg d@¢ZGdAZHdBZIeZJ e!dCeH«       G dD„ dEeA«      «       ZK e!dFeH«       G dG„ dHeA«      «       ZLdIZMdJZNdKZO e!dLeHdM«       G dN„ dOeA«      «       ZP e!dPeH«       G dQ„ dReA«      «       ZQg dS¢ZRy)Ué    N)Ú	dataclass)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ModelOutputÚSequenceClassifierOutputÚWav2Vec2BaseModelOutput)ÚPreTrainedModel)Úadd_code_sample_docstringsÚadd_start_docstringsÚ%add_start_docstrings_to_model_forwardÚloggingÚreplace_return_docstringsé   )ÚUniSpeechConfig)Ú_flash_attention_forwardz/patrickvonplaten/unispeech-large-1500h-cv-timitr   c                   ó  — e Zd ZU dZdZeej                     ed<   dZ	eej                     ed<   dZ
eej                     ed<   dZeej                     ed<   dZeeej                        ed<   dZeeej                        ed<   y)	ÚUniSpeechForPreTrainingOutputaL  
    Output type of [`UniSpeechForPreTrainingOutput`], with potential hidden states and attentions.

    Args:
        loss (*optional*, returned when model is in train mode, `torch.FloatTensor` of shape `(1,)`):
            Total loss as the sum of the contrastive loss (L_m) and the diversity loss (L_d) as stated in the [official
            paper](https://arxiv.org/pdf/2006.11477.pdf) . (classification) loss.
        projected_states (`torch.FloatTensor` of shape `(batch_size, sequence_length, config.proj_codevector_dim)`):
            Hidden-states of the model projected to *config.proj_codevector_dim* that can be used to predict the masked
            projected quantized states.
        projected_quantized_states (`torch.FloatTensor` of shape `(batch_size, sequence_length, config.proj_codevector_dim)`):
            Quantized extracted feature vectors projected to *config.proj_codevector_dim* representing the positive
            target vectors for contrastive loss.
        hidden_states (`tuple(torch.FloatTensor)`, *optional*, returned when `output_hidden_states=True` is passed or when `config.output_hidden_states=True`):
            Tuple of `torch.FloatTensor` (one for the output of the embeddings + one for the output of each layer) of
            shape `(batch_size, sequence_length, hidden_size)`.

            Hidden-states of the model at the output of each layer plus the initial embedding outputs.
        attentions (`tuple(torch.FloatTensor)`, *optional*, returned when `output_attentions=True` is passed or when `config.output_attentions=True`):
            Tuple of `torch.FloatTensor` (one for each layer) of shape `(batch_size, num_heads, sequence_length,
            sequence_length)`.

            Attentions weights after the attention softmax, used to compute the weighted average in the self-attention
            heads.
    NÚlossÚprojected_statesÚprojected_quantized_statesÚcodevector_perplexityÚhidden_statesÚ
attentions)Ú__name__Ú
__module__Ú__qualname__Ú__doc__r   r   ÚtorchÚFloatTensorÚ__annotations__r   r    r!   r"   r   r#   © ó    ún/var/www/skyplay_api_hub/venv/lib/python3.12/site-packages/transformers/models/unispeech/modeling_unispeech.pyr   r   4   s”   … ñð4 )-€Dˆ(�5×$Ñ$Ñ
%Ó,Ø48Ð�h˜u×0Ñ0Ñ1Ó8Ø>BÐ ¨×):Ñ):Ñ ;ÓBØ9=Ð˜8 E×$5Ñ$5Ñ6Ó=Ø8<€M�8˜E %×"3Ñ"3Ñ4Ñ5Ó<Ø59€J�˜˜u×0Ñ0Ñ1Ñ2Ô9r,   r   c                   ó$   ‡ — e Zd Zˆ fd„Zd„ Zˆ xZS )ÚUniSpeechSamePadLayerc                 óP   •— t         ‰| �  «        |dz  dk(  rd| _        y d| _        y )Né   r   r   )ÚsuperÚ__init__Únum_pad_remove)ÚselfÚnum_conv_pos_embeddingsÚ	__class__s     €r-   r3   zUniSpeechSamePadLayer.__init__Y   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   )r4   ©r5   r"   s     r-   ÚforwardzUniSpeechSamePadLayer.forward]   s6   € Ø×Ñ Ò"Ø)ª!ªQÐ0F°4×3FÑ3FÐ2FÐ0FÐ*FÑGˆMØÐr,   ©r$   r%   r&   r3   r;   Ú__classcell__©r7   s   @r-   r/   r/   X   s   ø„ ôKör,   r/   c                   ó$   ‡ — e Zd Zˆ fd„Zd„ Zˆ xZS )Ú UniSpeechPositionalConvEmbeddingc                 ó¦  •— t         ‰| �  «        t        j                  |j                  |j                  |j
                  |j
                  dz  |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
                  «      | _        t0        |j2                     | _        y # 1 sw Y   �Œ'xY w)	Nr1   )Úkernel_sizeÚpaddingÚgroupsÚweight_normr   )Úmodifier_rankÚweight)ÚnameÚdimÚparametrizations)r2   r3   ÚnnÚConv1dÚhidden_sizer6   Únum_conv_pos_embedding_groupsÚconvÚutilsrE   ÚhasattrrJ   r
   Ú	deepspeedÚzeroÚGatheredParametersrG   Ú	original0Ú	original1Úweight_gÚweight_vÚregister_external_parameterr/   rC   r	   Úfeat_extract_activationÚ
activation)r5   ÚconfigrE   rR   rW   rX   r7   s         €r-   r3   z)UniSpeechPositionalConvEmbedding.__init__d   s¥  ø€ Ü‰ÑÔÜ—I‘IØ×ÑØ×ÑØ×6Ñ6Ø×2Ñ2°aÑ7Ø×7Ñ7ô
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×-KÑ-KÓLˆŒÜ  ×!?Ñ!?Ñ@ˆ�÷Iñ Iús   ÄIÉIc                 ó´   — |j                  dd«      }| j                  |«      }| j                  |«      }| j                  |«      }|j                  dd«      }|S ©Nr   r1   )Ú	transposerO   rC   r[   r:   s     r-   r;   z(UniSpeechPositionalConvEmbedding.forward…   sV   € Ø%×/Ñ/°°1Ó5ˆàŸ	™	 -Ó0ˆØŸ™ ]Ó3ˆØŸ™¨Ó6ˆà%×/Ñ/°°1Ó5ˆØÐr,   r<   r>   s   @r-   r@   r@   c   s   ø„ ôAöBr,   r@   c                   ó&   ‡ — e Zd Zdˆ fd„	Zd„ Zˆ xZS )ÚUniSpeechNoLayerNormConvLayerc                 ód  •— t         ‰| �  «        |dkD  r|j                  |dz
     nd| _        |j                  |   | _        t        j                  | j                  | j                  |j                  |   |j                  |   |j                  ¬«      | _
        t        |j                     | _        y )Nr   r   ©rB   ÚstrideÚbias)r2   r3   Úconv_dimÚin_conv_dimÚout_conv_dimrK   rL   Úconv_kernelÚconv_strideÚ	conv_biasrO   r	   rZ   r[   ©r5   r\   Úlayer_idr7   s      €r-   r3   z&UniSpeechNoLayerNormConvLayer.__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 ! ×!?Ñ!?Ñ@ˆ�r,   c                 óJ   — | j                  |«      }| j                  |«      }|S ©N)rO   r[   r:   s     r-   r;   z%UniSpeechNoLayerNormConvLayer.forwardŸ   s$   € ØŸ	™	 -Ó0ˆØŸ™¨Ó6ˆØÐr,   ©r   r<   r>   s   @r-   ra   ra   �   s   ø„ õAör,   ra   c                   ó&   ‡ — e Zd Zdˆ fd„	Zd„ Zˆ xZS )ÚUniSpeechLayerNormConvLayerc                 ó°  •— t         ‰| �  «        |dkD  r|j                  |dz
     nd| _        |j                  |   | _        t        j                  | j                  | j                  |j                  |   |j                  |   |j                  ¬«      | _
        t        j                  | j                  d¬«      | _        t        |j                     | _        y )Nr   r   rc   T)Úelementwise_affine)r2   r3   rf   rg   rh   rK   rL   ri   rj   rk   rO   Ú	LayerNormÚ
layer_normr	   rZ   r[   rl   s      €r-   r3   z$UniSpeechLayerNormConvLayer.__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�r,   c                 ó´   — | j                  |«      }|j                  dd«      }| j                  |«      }|j                  dd«      }| j                  |«      }|S )Néþÿÿÿéÿÿÿÿ)rO   r_   rv   r[   r:   s     r-   r;   z#UniSpeechLayerNormConvLayer.forwardµ   sV   € ØŸ	™	 -Ó0ˆà%×/Ñ/°°BÓ7ˆØŸ™¨Ó6ˆØ%×/Ñ/°°BÓ7ˆàŸ™¨Ó6ˆØÐr,   rp   r<   r>   s   @r-   rr   rr   ¥   s   ø„ õAör,   rr   c                   ó&   ‡ — e Zd Zdˆ fd„	Zd„ Zˆ xZS )ÚUniSpeechGroupNormConvLayerc                 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   rc   T)Ú
num_groupsÚnum_channelsÚaffine)r2   r3   rf   rg   rh   rK   rL   ri   rj   rk   rO   r	   rZ   r[   Ú	GroupNormrv   rl   s      €r-   r3   z$UniSpeechGroupNormConvLayer.__init__Á   s¹   ø€ Ü‰ÑÔØ<DÀqºL˜6Ÿ?™?¨8°a©<Ò8ÈaˆÔØ"ŸO™O¨HÑ5ˆÔä—I‘IØ×ÑØ×ÑØ×*Ñ*¨8Ñ4Ø×%Ñ% hÑ/Ø×!Ñ!ô
ˆŒ	ô ! ×!?Ñ!?Ñ@ˆŒäŸ,™,°$×2CÑ2CÐRV×RcÑRcÐlpÔqˆ�r,   c                 ól   — | j                  |«      }| j                  |«      }| j                  |«      }|S ro   )rO   rv   r[   r:   s     r-   r;   z#UniSpeechGroupNormConvLayer.forwardÑ   s2   € ØŸ	™	 -Ó0ˆØŸ™¨Ó6ˆØŸ™¨Ó6ˆØÐr,   rp   r<   r>   s   @r-   r{   r{   À   s   ø„ õrö r,   r{   c                   ó.   ‡ — e Zd ZdZˆ fd„Zd„ Zd„ Zˆ xZS )ÚUniSpeechFeatureEncoderz.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   )rm   r   Úlayerz`config.feat_extract_norm` is z), but has to be one of ['group', 'layer']FT)r2   r3   Úfeat_extract_normr{   ÚrangeÚnum_feat_extract_layersra   rr   Ú
ValueErrorrK   Ú
ModuleListÚconv_layersÚgradient_checkpointingÚ_requires_grad)r5   r\   ÚirŒ   r7   s       €r-   r3   z UniSpeechFeatureEncoder.__init__Û   sñ   ø€ Ü‰ÑÔà×#Ñ# wÒ.Ü6°vÈÔJÐKä˜v×=Ñ=ÀÑAÓBöOàô .¨f¸qÀ1¹uÖEòOñ ‰Kð ×%Ñ%¨Ò0äINÈv×OmÑOmÓInöØDEÔ+¨F¸QÖ?ðˆKñ ô Ø0°×1IÑ1IÐ0JÐJsÐtóð ô Ÿ=™=¨Ó5ˆÔØ&+ˆÔ#Ø"ˆÕùòOùò
s   ÁC"Â	C'c                 óJ   — | j                  «       D ]	  }d|_        Œ d| _        y ©NF)Ú
parametersÚrequires_gradrŽ   ©r5   Úparams     r-   Ú_freeze_parametersz*UniSpeechFeatureEncoder._freeze_parametersï   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)rŽ   Útrainingr“   rŒ   r�   Ú_gradient_checkpointing_funcÚ__call__)r5   Úinput_valuesr"   Ú
conv_layers       r-   r;   zUniSpeechFeatureEncoder.forwardô   s…   € Ø$¢Q¨ WÑ-ˆð ×Ò 4§=¢=Ø*.ˆMÔ'à×*Ñ*ò 	:ˆJØ×"Ò" t×'BÒ'BÀtÇ}Â}Ø $× AÑ AØ×'Ñ'Ø!ó!‘ñ
 !+¨=Ó 9‘ð	:ð Ðr,   )r$   r%   r&   r'   r3   r–   r;   r=   r>   s   @r-   rƒ   rƒ   Ø   s   ø„ Ù8ô#ò($ö
r,   rƒ   c                   ó$   ‡ — e Zd Zˆ fd„Zd„ Zˆ xZS )ÚUniSpeechFeatureProjectionc                 ó4  •— t         ‰| �  «        t        j                  |j                  d   |j
                  ¬«      | _        t        j                  |j                  d   |j                  «      | _	        t        j                  |j                  «      | _        y )Nry   ©Úeps)r2   r3   rK   ru   rf   Úlayer_norm_epsrv   ÚLinearrM   Ú
projectionÚDropoutÚfeat_proj_dropoutÚdropout©r5   r\   r7   s     €r-   r3   z#UniSpeechFeatureProjection.__init__  sf   ø€ Ü‰ÑÔÜŸ,™, v§¡°rÑ':À×@UÑ@UÔVˆŒÜŸ)™) F§O¡O°BÑ$7¸×9KÑ9KÓLˆŒÜ—z‘z &×":Ñ":Ó;ˆ�r,   c                 óp   — | j                  |«      }| j                  |«      }| j                  |«      }||fS ro   )rv   r¤   r§   )r5   r"   Únorm_hidden_statess      r-   r;   z"UniSpeechFeatureProjection.forward  s:   € à!Ÿ_™_¨]Ó;ÐØŸ™Ð(:Ó;ˆØŸ™ ]Ó3ˆØÐ0Ð0Ð0r,   r<   r>   s   @r-   rž   rž     s   ø„ ô<ö1r,   rž   c                   ó†  ‡ — e Zd ZdZ	 	 	 	 	 ddededededededee   fˆ fd	„Z	d
e
j                  dedefd„Z	 	 	 	 	 dde
j                  dee
j                     deee
j                        dee
j                     dee
j                     dedee
j                  ee
j                     eee
j                        f   fd„Zˆ xZS )ÚUniSpeechAttentionz=Multi-headed attention from 'Attention Is All You Need' paperÚ	embed_dimÚ	num_headsr§   Ú
is_decoderre   Ú	is_causalr\   c                 ó
  •— t         ‰| �  «        || _        || _        || _        ||z  | _        || _        | j
                  |z  | j                  k7  rt        d| j                  › d|› d�«      ‚| j
                  dz  | _        || _	        || _
        t        j                  |||¬«      | _        t        j                  |||¬«      | _        t        j                  |||¬«      | _        t        j                  |||¬«      | _        y )Nz;embed_dim must be divisible by num_heads (got `embed_dim`: z and `num_heads`: z).g      à¿)re   )r2   r3   r­   r®   r§   Úhead_dimr\   rŠ   Úscalingr¯   r°   rK   r£   Úk_projÚv_projÚq_projÚout_proj)	r5   r­   r®   r§   r¯   re   r°   r\   r7   s	           €r-   r3   zUniSpeechAttention.__init__  sä   ø€ ô 	‰ÑÔØ"ˆŒØ"ˆŒØˆŒØ! YÑ.ˆŒØˆŒà�M‰M˜IÑ%¨$¯.©.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_lenÚbszc                 óŽ   — |j                  ||| j                  | j                  «      j                  dd«      j	                  «       S r^   )Úviewr®   r²   r_   Ú
contiguous©r5   r¸   r¹   rº   s       r-   Ú_shapezUniSpeechAttention._shape8  s7   € Ø�{‰{˜3 ¨¯©¸¿¹ÓG×QÑQÐRSÐUVÓW×bÑbÓdÐdr,   r"   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$                  j)                  || j(                  | j*                  ¬«      }t        j                  ||«      }|j                  «       || j                  z  |	| j                  fk7  r9t!        d|| j                  z  |	| j                  f› d|j                  «       › �«      ‚|j                  || j                  |	| j                  «      }|j                  dd«      }|j                  ||	| j,                  «      }| j/                  |«      }|||fS )ú#Input shape: Batch x Time x ChannelNr   r1   r   ry   ©rI   z$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¶   r³   Úshaper¿   r´   rµ   r(   Úcatr¯   r®   r²   r¼   ÚreshapeÚbmmr_   rŠ   rK   Ú
functionalÚsoftmaxr§   r˜   r­   r·   )r5   r"   rÀ   rÁ   rÂ   rÃ   rÄ   Úis_cross_attentionrº   Útgt_lenÚ_Úquery_statesÚ
key_statesÚvalue_statesÚ
proj_shapeÚsrc_lenÚattn_weightsÚattn_weights_reshapedÚ
attn_probsÚattn_outputs                       r-   r;   zUniSpeechAttention.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)r$   r%   r&   r'   ÚintÚfloatÚboolr   r   r3   r(   ÚTensorr¿   r   r;   r=   r>   s   @r-   r¬   r¬     sM  ø„ ÙGð Ø ØØØ,0ñCàðCð ðCð ð	Cð
 ðCð ðCð ðCð ˜Ñ)õCð>e˜UŸ\™\ð e°Cð e¸có eð 48Ø8<Ø15Ø26Ø"'ñvBà—|‘|ðvBð # 5§<¡<Ñ0ðvBð !  u§|¡|Ñ!4Ñ5ð	vBð
 ! §¡Ñ.ðvBð " %§,¡,Ñ/ðvBð  ðvBð 
ˆu�|‰|˜X e§l¡lÑ3°X¸eÀEÇLÁLÑ>QÑ5RÐRÑ	S÷vBr,   r¬   c                   óV  ‡ — e Zd ZdZˆ fd„Zdej                  dedefd„Z	 	 	 	 	 ddej                  de	ej                     d	e	e
ej                        d
e	ej                     de	ej                     dede
ej                  e	ej                     e	e
ej                        f   fd„Zˆ xZS )ÚUniSpeechFlashAttention2aN  
    UniSpeech flash attention module. This module inherits from `UniSpeechAttention` as the weights of the module stays
    untouched. The only required change would be on the forward pass where it needs to correctly call the public API of
    flash attention and deal with padding tokens in case the input contains any of them.
    c                 óB   •— t        ‰| �  |i |¤Ž t        «       | _        y ro   )r2   r3   r   Ú_flash_attn_uses_top_left_mask)r5   ÚargsÚkwargsr7   s      €r-   r3   z!UniSpeechFlashAttention2.__init__»  s#   ø€ Ü‰Ñ˜$Ð) &Ò)ô
 /PÓ.QˆÕ+r,   r¸   r¹   rº   c                 óR   — |j                  ||| j                  | j                  «      S ro   )r¼   r®   r²   r¾   s       r-   Ú_reshapez!UniSpeechFlashAttention2._reshapeÃ  s   € Ø�{‰{˜3 ¨¯©¸¿¹ÓGÐGr,   r"   rÀ   rÁ   rÂ   rÃ   rÄ   rÅ   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 )NzEUniSpeechFlashAttention2 attention does not support output_attentionsry   r   r1   r   rÈ   rx   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.rß   )r§   r°   Úuse_top_left_mask)rŠ   rÌ   rì   r¶   rÍ   r_   r´   rµ   r(   rÎ   r¯   ÚdtypeÚfloat32Úis_autocast_enabledÚget_autocast_gpu_dtyperQ   r\   rî   rG   ÚloggerÚwarning_onceÚtor   r˜   r§   r°   rè   rÏ   r·   )r5   r"   rÀ   rÁ   rÂ   rÃ   rÄ   rÓ   rº   Úq_lenrÕ   rÖ   r×   rØ   Ú
kv_seq_lenÚinput_dtypeÚtarget_dtyperÞ   rÛ   s                      r-   r;   z UniSpeechFlashAttention2.forwardÆ  s0  € ñ ÜÐdÓeÐeð .°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$(§M¢M�D—L’L°sØ—n‘nØ"×AÑAô	
ˆð "×)Ñ)¨#¨u°bÓ9ˆØ—m‘m KÓ0ˆá ØˆLà˜L¨.Ð8Ð8r,   rà   )r$   r%   r&   r'   r3   r(   rä   rá   rì   r   r   rã   r;   r=   r>   s   @r-   ræ   ræ   ´  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,   ræ   c                   ó$  ‡ — e Zd Z	 	 	 	 	 d	dej                  deej                     deeej                        deej                     deej                     dedeej                  eej                     eeej                        f   fˆ fd„Zˆ xZ	S )
ÚUniSpeechSdpaAttentionr"   rÀ   rÁ   rÂ   rÃ   rÄ   rÅ   c                 óz  •— |s|�*t         j                  d«       t        ‰| �  ||||||¬«      S |du}|j	                  «       \  }}	}
| j                  |«      }|r0|�.|d   j                  d   |j                  d   k(  r|d   }|d   }�n
|rE| j                  | j                  |«      d|«      }| j                  | j                  |«      d|«      }nÃ|�}| j                  | j                  |«      d|«      }| j                  | j                  |«      d|«      }t        j                  |d   |gd¬«      }t        j                  |d   |gd¬«      }nD| j                  | j                  |«      d|«      }| j                  | j                  |«      d|«      }| j                  r||f}| j                  ||	|«      }| j                  r	|€|	dkD  rd	nd
}t        j                  j                  j!                  ||||| j"                  r| j$                  nd|¬«      }|j	                  «       || j&                  |	| j(                  fk7  r7t+        d|| j&                  |	| j(                  f› d|j	                  «       › �«      ‚|j-                  dd«      }|j/                  ||	| j0                  «      }| j3                  |«      }|d|fS )rÇ   Na­  UniSpeechModel is using UniSpeechSdpaAttention, 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.)rÀ   rÁ   rÂ   rÃ   rÄ   r   r1   r   ry   rÈ   TFrß   )Ú	attn_maskÚ	dropout_pr°   rË   rÉ   )rõ   rö   r2   r;   rÌ   r¶   rÍ   r¿   r´   rµ   r(   rÎ   r¯   r°   rK   rÑ   Úscaled_dot_product_attentionr˜   r§   r®   r²   rŠ   r_   rÏ   r­   r·   )r5   r"   rÀ   rÁ   rÂ   rÃ   rÄ   rÓ   rº   rÔ   rÕ   rÖ   r×   rØ   r°   rÞ   r7   s                   €r-   r;   zUniSpeechSdpaAttention.forward3  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,   rà   )
r$   r%   r&   r(   rä   r   r   rã   r;   r=   r>   s   @r-   rý   rý   2  s¿   ø„ ð 48Ø8<Ø15Ø26Ø"'ñf1à—|‘|ðf1ð # 5§<¡<Ñ0ðf1ð !  u§|¡|Ñ!4Ñ5ð	f1ð
 ! §¡Ñ.ðf1ð " %§,¡,Ñ/ðf1ð  ðf1ð 
ˆu�|‰|˜X e§l¡lÑ3°X¸eÀEÇLÁLÑ>QÑ5RÐRÑ	S÷f1ñ f1r,   rý   c                   ó$   ‡ — e Zd Zˆ fd„Zd„ Zˆ xZS )ÚUniSpeechFeedForwardc                 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 ro   )r2   r3   rK   r¥   Úactivation_dropoutÚintermediate_dropoutr£   rM   Úintermediate_sizeÚintermediate_denseÚ
isinstanceÚ
hidden_actÚstrr	   Úintermediate_act_fnÚoutput_denseÚhidden_dropoutÚoutput_dropoutr¨   s     €r-   r3   zUniSpeechFeedForward.__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 ro   )r  r  r  r  r  r:   s     r-   r;   zUniSpeechFeedForward.forwardª  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,   r<   r>   s   @r-   r  r  œ  s   ø„ ô@ör,   r  )ÚeagerÚsdpaÚflash_attention_2c                   ó&   ‡ — e Zd Zˆ fd„Zdd„Zˆ xZS )ÚUniSpeechEncoderLayerc                 óÈ  •— t         ‰| �  «        t        |j                     |j                  |j
                  |j                  d¬«      | _        t        j                  |j                  «      | _        t        j                  |j                  |j                  ¬«      | _        t        |«      | _        t        j                  |j                  |j                  ¬«      | _        y )NF©r­   r®   r§   r¯   r    )r2   r3   ÚUNISPEECH_ATTENTION_CLASSESÚ_attn_implementationrM   Únum_attention_headsÚattention_dropoutÚ	attentionrK   r¥   r  r§   ru   r¢   rv   r  Úfeed_forwardÚfinal_layer_normr¨   s     €r-   r3   zUniSpeechEncoderLayer.__init__¼  s¥   ø€ Ü‰ÑÔÜ4°V×5PÑ5PÑQØ×(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0°Ó8ˆÔÜ "§¡¨V×-?Ñ-?ÀV×EZÑEZÔ [ˆÕr,   c                 óè   — |}| j                  |||¬«      \  }}}| j                  |«      }||z   }| j                  |«      }|| j                  |«      z   }| j	                  |«      }|f}|r||fz  }|S ©N©rÂ   rÄ   )r  r§   rv   r  r  ©r5   r"   rÂ   rÄ   Úattn_residualrÛ   rÕ   Úoutputss           r-   r;   zUniSpeechEncoderLayer.forwardÊ  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,   r‘   r<   r>   s   @r-   r  r  »  s   ø„ ô\÷r,   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UniSpeechEncoderc                 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  )r2   r3   r\   r@   Úpos_conv_embedrK   ru   rM   r¢   rv   r¥   r  r§   r‹   rˆ   Únum_hidden_layersr  Úlayersr�   r  Ú_use_flash_attention_2©r5   r\   rÕ   r7   s      €r-   r3   zUniSpeechEncoder.__init__ß  s¥   ø€ Ü‰ÑÔØˆŒÜ>¸vÓFˆÔÜŸ,™, v×'9Ñ'9¸v×?TÑ?TÔUˆŒÜ—z‘z &×"7Ñ"7Ó8ˆŒÜ—m‘mÌEÐRX×RjÑRjÓLkÖ$lÀqÔ%:¸6Õ%BÒ$lÓmˆŒØ&+ˆÔ#Ø&,×&AÑ&AÐEXÑ&XˆÕ#ùò %mó   Â!Cr"   rÂ   rÄ   Úoutput_hidden_statesÚreturn_dictc                 ó4  — |rdnd }|rdnd }|�Ý|j                  d«      j                  dd|j                  d   «      }d|| <   | j                  r|�d|v r|nd }n‘d|d d …d d d d …f   j	                  |j
                  ¬«      z
  }|t        j                  |j
                  «      j                  z  }|j                  |j                  d   d|j                  d   |j                  d   «      }| j                  |«      }	||	z   }| j                  |«      }| j                  |«      }t        «       xs t        | «      }
| j                  D ]£  }|r||fz   }t        j                   g «      }| j"                  r|| j$                  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 )Nr+   ry   r   r1   r   ç      ð?©rñ   TFr!  ©NNc              3   ó&   K  — | ]	  }|€Œ|–— Œ y ­wro   r+   ©Ú.0Úvs     r-   ú	<genexpr>z+UniSpeechEncoder.forward.<locals>.<genexpr>+  ó   è ø€ Òm˜qÐ_`Ñ_lœÑmùó   ‚Š©Úlast_hidden_stater"   r#   )Ú	unsqueezeÚrepeatrÍ   r,  r÷   rñ   r(   ÚfinfoÚminÚexpandr)  rv   r§   r
   r   r+  Úrandr˜   r\   Ú	layerdropr�   r™   rš   Útupler   ©r5   r"   rÂ   rÄ   r/  r0  Úall_hidden_statesÚall_self_attentionsÚexpand_attention_maskÚposition_embeddingsÚsynced_gpusr†   Údropout_probabilityÚskip_the_layerÚlayer_outputss                  r-   r;   zUniSpeechEncoder.forwardé  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 	
r,   ©NFFT)r$   r%   r&   r3   r(   r¸   r   rä   rã   r;   r=   r>   s   @r-   r&  r&  Þ  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
r,   r&  c                   ó>   ‡ — e Zd Zˆ fd„Zdej
                  fd„Zˆ xZS )ÚUniSpeechAttnAdapterLayerc                 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)r2   r3   Úadapter_attn_dimÚ	input_dimrM   Ú
hidden_dimrK   ru   Únormr£   Úlinear_1ÚReLUÚact_fnÚlinear_2r¨   s     €r-   r3   z"UniSpeechAttnAdapterLayer.__init__4  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�r,   r"   c                 óŽ   — | j                  |«      }| j                  |«      }| j                  |«      }| j                  |«      }|S ro   )rV  rW  rY  rZ  r:   s     r-   r;   z!UniSpeechAttnAdapterLayer.forwardB  s@   € ØŸ	™	 -Ó0ˆàŸ™ mÓ4ˆØŸ™ MÓ2ˆØŸ™ mÓ4ˆàÐr,   )r$   r%   r&   r3   r(   r)   r;   r=   r>   s   @r-   rQ  rQ  3  s   ø„ ôCð U×%6Ñ%6÷ r,   rQ  c                   óf   ‡ — e Zd Zˆ fd„Z	 	 ddej
                  deej
                     defd„Zˆ xZ	S )Ú$UniSpeechEncoderLayerStableLayerNormc                 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    rS  )r2   r3   r  r  rM   r  r  r  rK   r¥   r  r§   ru   r¢   rv   r  r  r  ÚgetattrrQ  Úadapter_layerr¨   s     €r-   r3   z-UniSpeechEncoderLayerStableLayerNorm.__init__M  sÊ   ø€ Ü‰ÑÔÜ4°V×5PÑ5PÑQØ×(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0°Ó8ˆÔÜ "§¡¨V×-?Ñ-?ÀV×EZÑEZÔ [ˆÔä�6Ð-¨tÓ4Ð@Ü!:¸6Ó!BˆDÕà!%ˆDÕr,   r"   rÂ   rÄ   c                 ó$  — |}| j                  |«      }| j                  |||¬«      \  }}}| j                  |«      }||z   }|| j                  | j	                  |«      «      z   }| j
                  �|| j                  |«      z   }|f}|r||fz  }|S r   )rv   r  r§   r  r  r`  r"  s           r-   r;   z,UniSpeechEncoderLayerStableLayerNorm.forward_  s±   € ð &ˆØŸ™¨Ó6ˆØ)-¯©Ø¨.ÐL]ð *8ó *
Ñ&ˆ�| Qð Ÿ™ ]Ó3ˆØ%¨Ñ5ˆØ%¨×(9Ñ(9¸$×:OÑ:OÐP]Ó:^Ó(_Ñ_ˆà×ÑÐ)Ø)¨D×,>Ñ,>¸}Ó,MÑMˆMà Ð"ˆáØ˜�Ñ&ˆGàˆr,   r‘   )
r$   r%   r&   r3   r(   rä   r   rã   r;   r=   r>   s   @r-   r]  r]  L  s>   ø„ ô&ð* 26Ø"'ñ	à—|‘|ðð ! §¡Ñ.ðð  ÷	r,   r]  c                   ó.   ‡ — e Zd Zˆ fd„Z	 	 	 	 dd„Zˆ xZS )ÚUniSpeechEncoderStableLayerNormc                 óÀ  •— t         ‰| �  «        || _        t        |«      | _        t        j                  |j                  |j                  ¬«      | _	        t        j                  |j                  «      | _        t        j                  t        |j                  «      D �cg c]  }t!        |«      ‘Œ c}«      | _        d| _        |j&                  dk(  | _        y c c}w r(  )r2   r3   r\   r@   r)  rK   ru   rM   r¢   rv   r¥   r  r§   r‹   rˆ   r*  r]  r+  r�   r  r,  r-  s      €r-   r3   z(UniSpeechEncoderStableLayerNorm.__init__z  s©   ø€ Ü‰ÑÔØˆŒÜ>¸vÓFˆÔÜŸ,™, v×'9Ñ'9¸v×?TÑ?TÔUˆŒÜ—z‘z &×"7Ñ"7Ó8ˆŒÜ—m‘mÜCHÈ×IaÑIaÓCbÖc¸aÔ1°&Õ9Òcó
ˆŒð ',ˆÔ#Ø&,×&AÑ&AÐEXÑ&XˆÕ#ùò dr.  c                 óf  — |rdnd }|rdnd }|�ö|j                  d«      j                  dd|j                  d   «      }||j                  |j                  ¬«      z  }| j
                  r|�d|v r|nd }n‘d|d d …d d d d …f   j                  |j                  ¬«      z
  }|t        j                  |j                  «      j                  z  }|j                  |j                  d   d|j                  d   |j                  d   «      }| j                  |«      }	||	z   }| j                  |«      }t        «       xs t        | «      }
| j                  D ]£  }|r||fz   }t        j                  g «      }| j                   r|| j"                  j$                  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+   ry   r   r1   r3  r   r2  TFr!  r4  c              3   ó&   K  — | ]	  }|€Œ|–— Œ y ­wro   r+   r6  s     r-   r9  z:UniSpeechEncoderStableLayerNorm.forward.<locals>.<genexpr>Ê  r:  r;  r<  )r>  r?  rÍ   r÷   rñ   r,  r(   r@  rA  rB  r)  r§   r
   r   r+  rC  r˜   r\   rD  r�   r™   rš   rv   rE  r   rF  s                  r-   r;   z'UniSpeechEncoderStableLayerNorm.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 	
r,   rO  r<   r>   s   @r-   rc  rc  y  s   ø„ ô
Yð ØØ"Ø÷I
r,   rc  c                   ó8   ‡ — e Zd ZdZˆ fd„Zed„ «       Zd„ Zˆ xZS )ÚUniSpeechGumbelVectorQuantizerz­
    Vector quantization using gumbel softmax. See `[CATEGORICAL REPARAMETERIZATION WITH
    GUMBEL-SOFTMAX](https://arxiv.org/pdf/1611.01144.pdf) for more information.
    c                 ó0  •— t         ‰| �  «        |j                  | _        |j                  | _        |j                  | j                  z  dk7  r&t        d|j                  › d| j                  › d�«      ‚t        j                  t        j                  d| j                  | j
                  z  |j                  | j                  z  «      «      | _        t        j                  |j                  d   | j                  | j
                  z  «      | _        d| _        y )Nr   z`config.codevector_dim z5 must be divisible by `config.num_codevector_groups` z for concatenationr   ry   r1   )r2   r3   Únum_codevector_groupsr}   Únum_codevectors_per_groupÚnum_varsÚcodevector_dimrŠ   rK   Ú	Parameterr(   r)   Úcodevectorsr£   rf   Úweight_projÚtemperaturer¨   s     €r-   r3   z'UniSpeechGumbelVectorQuantizer.__init__Ø  sì   ø€ Ü‰ÑÔØ ×6Ñ6ˆŒØ×8Ñ8ˆŒà× Ñ  4§?¡?Ñ2°aÒ7ÜØ)¨&×*?Ñ*?Ð)@ð A5Ø59·_±_Ð4EÐEWðYóð ô Ÿ<™<Ü×Ñ˜a §¡°4·=±=Ñ!@À&×BWÑBWÐ[_×[jÑ[jÑBjÓkó
ˆÔô Ÿ9™9 V§_¡_°RÑ%8¸$¿/¹/ÈDÏMÉMÑ:YÓZˆÔð ˆÕr,   c           	      óÎ   — | j                  d¬«      }t        j                  t        j                  |t        j                  |dz   «      z  d¬«       «      j                  «       }|S )Nr   rÈ   gH¯¼šò×z>ry   )Úmeanr(   ÚexpÚsumÚlog)ÚprobsÚmarginal_probsÚ
perplexitys      r-   Ú_compute_perplexityz2UniSpeechGumbelVectorQuantizer._compute_perplexityì  sR   € àŸ™¨˜Ó*ˆÜ—Y‘Y¤§	¡	¨.¼5¿9¹9À^ÐVZÑEZÓ;[Ñ*[ÐacÔ dÐdÓe×iÑiÓkˆ
ØÐr,   c                 óÞ  — |j                   \  }}}| j                  |«      }|j                  ||z  | j                  z  d«      }| j                  ržt
        j                  j                  |j                  «       | j                  d¬«      j                  |«      }t        j                  |j                  ||z  | j                  d«      j                  «       d¬«      }| j                  |«      }n}|j                  d¬«      } |j                  |j                   Ž j!                  d|j                  dd«      d«      }|j                  ||z  | j                  d«      }| j                  |«      }|j                  ||z  d«      }|j#                  d«      | j$                  z  }	|	j                  ||z  | j                  | j&                  d«      }
|
j)                  d«      j                  ||d«      }
|
|fS )Nry   T)ÚtauÚhardrÈ   r   r2  rx   )rÍ   rp  r¼   r}   r˜   rK   rÑ   Úgumbel_softmaxrâ   rq  Útype_asr(   rÒ   rz  ÚargmaxÚ	new_zerosÚscatter_r>  ro  rl  ru  )r5   r"   Ú
batch_sizeÚsequence_lengthrM   Úcodevector_probsÚcodevector_soft_distry  Úcodevector_idxÚcodevectors_per_groupro  s              r-   r;   z&UniSpeechGumbelVectorQuantizer.forwardò  sÛ  € Ø3@×3FÑ3FÑ0ˆ
�O [ð ×(Ñ(¨Ó7ˆØ%×*Ñ*¨:¸Ñ+GÈ$Ï/É/Ñ+YÐ[]Ó^ˆà�=Š=ä!Ÿ}™}×;Ñ;Ø×#Ñ#Ó%¨4×+;Ñ+;À$ð  <ó  ç‰g�mÓ$ð ô
 $)§=¡=Ø×"Ñ" :°Ñ#?ÀÇÁÐRTÓU×[Ñ[Ó]Ðceô$Ð ð ×1Ñ1Ð2FÓG‰Jð +×1Ñ1°bÐ1Ó9ˆNØ6˜}×6Ñ6¸×8KÑ8KÐL×UÑUØ�N×'Ñ'¨¨AÓ.°ó Ðð  0×4Ñ4°ZÀ/Ñ5QÐSW×SbÑSbÐdfÓgÐà×1Ñ1Ð2BÓCˆJà+×0Ñ0°¸oÑ1MÈrÓRÐà 0× :Ñ :¸2Ó >À×AQÑAQÑ QÐØ+×0Ñ0°¸oÑ1MÈtÏÉÐ`d×`mÑ`mÐoqÓrˆØ!—o‘o bÓ)×.Ñ.¨z¸?ÈBÓOˆà˜JÐ&Ð&r,   )	r$   r%   r&   r'   r3   Ústaticmethodrz  r;   r=   r>   s   @r-   rh  rh  Ò  s&   ø„ ñô
ð( ñó ðö
#'r,   rh  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UniSpeechPreTrainedModelz†
    An abstract class to handle weights initialization and a simple interface for downloading and loading pretrained
    models.
    Ú	unispeechr›   Tc           
      óz  — t        |t        «      r‰|j                  j                  j                  j                  dd¬«       |j                  j                  j                  j                  «        t        j                  j                  |j                  «       yt        |t        «      r²t        j                  j                  |j                  j                  ddt        j                  d|j                  j                   d   |j                  j"                  z  z  «      z  ¬«       t        j                  j%                  |j                  j                  d«       yt        |t&        «      r›t        j                  d|j(                  j*                  z  «      }t        j                  j                  |j(                  j                  | |¬«       t        j                  j                  |j(                  j                  | |¬«       yt        |t        j,                  «      rm|j                  j                  j                  d| j.                  j0                  ¬«       |j                  �%|j                  j                  j                  «        yyt        |t        j2                  t        j4                  f«      rJ|j                  j                  j                  «        |j                  j                  j7                  d«       yt        |t        j8                  «      r t        j                  j;                  |j                  «       |j                  �jt        j                  |j<                  |j"                  |j                   d   z  z  «      }t        j                  j                  |j                  | |¬«       yyy)	zInitialize the weightsrß   r   )rs  Ústdr   r1   )ÚaÚbNr2  )r	  rh  rp  rG   ÚdataÚnormal_re   Úzero_rK   ÚinitÚuniform_ro  r@   rO   ÚmathÚsqrtrB   Úin_channelsÚ	constant_rž   r¤   Úin_featuresr£   r\   Úinitializer_rangeru   r€   Úfill_rL   Úkaiming_normal_rD   )r5   ÚmoduleÚks      r-   Ú_init_weightsz&UniSpeechPreTrainedModel._init_weights%  sŠ  € ô �fÔ<Ô=Ø×Ñ×%Ñ%×*Ñ*×2Ñ2¸ÀÐ2ÔCØ×Ñ×#Ñ#×(Ñ(×.Ñ.Ô0Ü�G‰G×Ñ˜V×/Ñ/Õ0Ü˜Ô @ÔAÜ�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.°Õ2Ü˜Ô :Ô;Ü—	‘	˜!˜f×/Ñ/×;Ñ;Ñ;Ó<ˆAÜ�G‰G×Ñ˜V×.Ñ.×5Ñ5¸!¸¸qÐÔAÜ�G‰G×Ñ˜V×.Ñ.×3Ñ3¸°r¸QÐÕ?Ü˜¤§	¡	Ô*Ø�M‰M×Ñ×&Ñ&¨C°T·[±[×5RÑ5RÐ&ÔSà�{‰{Ð&Ø—‘× Ñ ×&Ñ&Õ(ð 'ä˜¤§¡¬r¯|©|Ð <Ô=Ø�K‰K×Ñ×"Ñ"Ô$Ø�M‰M×Ñ×$Ñ$ SÕ)Ü˜¤§	¡	Ô*Ü�G‰G×#Ñ# F§M¡MÔ2à�{‰{Ð&Ü—I‘I˜fŸm™m¨v×/AÑ/AÀF×DVÑDVÐWXÑDYÑ/YÑZÓ[�Ü—‘× Ñ  §¡°°°aÐ Õ8ð 'ð +r,   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   )r(   Údiv)Úinput_lengthrB   rd   s      r-   Ú_conv_out_lengthzSUniSpeechPreTrainedModel._get_feat_extract_output_lengths.<locals>._conv_out_lengthK  s"   € ô —9‘9˜\¨KÑ7¸ÈwÔWÐZ[Ñ[Ð[r,   )Úzipr\   ri   rj   )r5   r¡  r¨  rB   rd   s        r-   Ú _get_feat_extract_output_lengthsz9UniSpeechPreTrainedModel._get_feat_extract_output_lengthsF  sQ   € ò
	\ô
 $' t§{¡{×'>Ñ'>ÀÇÁ×@WÑ@WÓ#Xò 	QÑˆK˜Ù,¨]¸KÈÓP‰Mð	Qð Ðr,   Úfeature_vector_lengthrÂ   c                 óø  — |j                  d¬«      d d …df   }| j                  |«      j                  t        j                  «      }|j
                  d   }t        j                  ||f|j                  |j                  ¬«      }d|t        j                  |j
                  d   |j                  ¬«      |dz
  f<   |j                  dg«      j                  d«      j                  dg«      j                  «       }|S )Nry   rÈ   r   )rñ   Údevicer   )r­  )Úcumsumrª  r÷   r(   ÚlongrÍ   Úzerosrñ   r­  ÚarangeÚfliprã   )r5   r«  rÂ   Únon_padded_lengthsÚoutput_lengthsrƒ  s         r-   Ú"_get_feature_vector_attention_maskz;UniSpeechPreTrainedModel._get_feature_vector_attention_maskU  së   € ð ,×2Ñ2°rÐ2Ó:º1¸b¸5ÑAÐØ×>Ñ>Ð?QÓR×UÑUÔV[×V`ÑV`ÓaˆØ#×)Ñ)¨!Ñ,ˆ
äŸ™ØÐ.Ð/°~×7KÑ7KÐTb×TiÑTiô
ˆð uvˆœŸ™ ^×%9Ñ%9¸!Ñ%<À^×EZÑEZÔ[Ð]kÐnoÑ]oÐpÑqØ'×,Ñ,¨b¨TÓ2×9Ñ9¸"Ó=×BÑBÀBÀ4ÓH×MÑMÓOˆØÐr,   N)r$   r%   r&   r'   r   Úconfig_classÚbase_model_prefixÚmain_input_nameÚsupports_gradient_checkpointingÚ_supports_flash_attn_2Ú_supports_sdpar   r   r(   Ú
LongTensorrá   rª  rµ  r+   r,   r-   r‹  r‹    sg   „ ñð
 #€LØ#ÐØ$€OØ&*Ð#Ø!ÐØ€Nò9ðB¸eÀE×DTÑDTÐVYÐDYÑ>Zó ðÈð Ð]b×]mÑ]mô r,   r‹  rÍ   Ú	mask_probÚmask_lengthrÂ   Ú	min_masksrÅ   c                 óà  ‡‡‡‡‡— | \  }Š‰dk  rt        d«      ‚‰‰kD  rt        d‰› d‰› d�«      ‚t        j                  j                  d«      j	                  «       Šˆˆˆˆˆfd„}|�-|j                  «       j                  d«      j                  «       nt        |«      D �cg c]  }‰‘Œ c}}t        j                  |‰ft        ¬	«      }	g }
 |‰«      }|d
k(  r|	S |D ]¯  } ||«      }t        j                  j                  t        j                  |‰dz
  z
  «      |d¬«      }t        |«      d
k(  r‰dz
  }n|d
   }t        j                  |t        j                  ||z
  t        j                   ¬	«      |z  g«      }|
j#                  |«       Œ± t        j$                  |
«      }
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   )rá   Úmax)r§  Únum_masked_spanÚepsilonr¾  r½  r¿  r„  s     €€€€€r-   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r,   Nry   r3  r   F)Úreplace)rŠ   ÚnpÚrandomrC  ÚitemÚdetachru  Útolistrˆ   r°  rã   Úchoicer±  ÚlenÚconcatenateÚonesÚint32ÚappendÚarrayÚbroadcast_torÏ   rÃ  Úput_along_axis)rÍ   r½  r¾  rÂ   r¿  rƒ  rÆ  rÕ   r¡  Úspec_aug_maskÚspec_aug_mask_idxsÚmax_num_masked_spanr§  rÄ  Úspec_aug_mask_idxÚdummy_mask_idxÚoffsetsrÅ  r„  s    `` `            @@r-   Ú_compute_mask_indicesrÜ  e  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+)r   i$  i   a  
    UniSpeech was proposed in [UniSpeech: Unified Speech Representation Learning with Labeled and Unlabeled
    Data](https://arxiv.org/abs/2101.07597) by Chengyi Wang, Yu Wu, Yao Qian, Kenichi Kumatani, Shujie Liu, Furu Wei,
    Michael Zeng, Xuedong Huang.

    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 ([`UniSpeechConfig`]): 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.
aI  
    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`, `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.
zcThe bare UniSpeech 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UniSpeechModelr\   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 )Nrß   )r2   r3   r\   rƒ   Úfeature_extractorrž   Úfeature_projectionÚmask_time_probÚmask_feature_probrK   rn  r(   rä   rM   r•  Úmasked_spec_embedÚdo_stable_layer_normrc  Úencoderr&  Ú	post_initr¨   s     €r-   r3   zUniSpeechModel.__init__  s¢   ø€ Ü‰Ñ˜Ô ØˆŒÜ!8¸Ó!@ˆÔÜ"<¸VÓ"DˆÔà× Ñ  3Ò&¨&×*BÑ*BÀSÒ*HÜ%'§\¡\´%·,±,¸v×?QÑ?QÓ2R×2[Ñ2[Ó2]Ó%^ˆDÔ"à×&Ò&Ü:¸6ÓBˆD�Lä+¨FÓ3ˆDŒLð 	�‰Õr,   r"   Úmask_time_indicesrÂ   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   )r½  r¾  rÂ   r¿  )r­  rñ   )r½  r¾  r¿  ry   )r_  r\   rÌ   rä  r÷   rñ   râ  r˜   rÜ  Úmask_time_lengthÚmask_time_min_masksr(   r¸   r­  rã   rã  Úmask_feature_lengthÚmask_feature_min_masksrB  )r5   r"   rè  rÂ   rƒ  r„  rM   Úmask_feature_indicess           r-   Ú_mask_hidden_statesz"UniSpeechModel._mask_hidden_states/  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_outputr›   rÄ   r/  r0  rÅ   c                 ó
  — |�|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 )Nr   r1   )rè  rÂ   ©rÂ   rÄ   r/  r0  r   )r=  Úextract_featuresr"   r#   )r\   rÄ   r/  Úuse_return_dictrà  r_   rµ  rÍ   rá  rð  ræ  ÚUniSpeechBaseModelOutputr"   r#   )
r5   r›   rÂ   rè  rÄ   r/  r0  rø  r"   Úencoder_outputss
             r-   r;   zUniSpeechModel.forward]  s@  € 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àÐ%à!×DÑDÐEU×E[ÑE[Ð\]ÑE^Ð`nÓoˆNà*.×*AÑ*AÐBRÓ*SÑ'ˆÐ'Ø×0Ñ0ØÐ->È~ð 1ó 
ˆð Ÿ,™,ØØ)Ø/Ø!5Ø#ð 'ó 
ˆð (¨Ñ*ˆáØ!Ð#3Ð4°ÀqÀrÐ7JÑJÐJä'Ø+Ø-Ø)×7Ñ7Ø&×1Ñ1ô	
ð 	
r,   r4  ©NNNNN)r$   r%   r&   r   r3   r(   r)   r   r¼  rð  r   ÚUNISPEECH_INPUTS_DOCSTRINGr   Ú_CHECKPOINT_FOR_DOCrú  Ú_CONFIG_FOR_DOCÚ_EXPECTED_OUTPUT_SHAPErä   rã   r   r   r;   r=   r>   s   @r-   rÞ  rÞ    s  ø„ ð
˜õ ð( :>Ø59ñ	,à×(Ñ(ð,ð $ E×$5Ñ$5Ñ6ð,ð ! ×!1Ñ!1Ñ2ó	,ñ\ +Ð+EÓFÙØ&Ø,Ø$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 
ˆuÐ.Ð.Ñ	/ò-
óó Gô-
r,   rÞ  zPUniSpeech Model with a vector-quantization module and ctc loss for pre-training.c                   óH  ‡ — e Zd Zdefˆ fd„Zdefd„Zd„ Zd„ Ze		 dde
j                  de
j                  d	e
j                  defd
„«       Z 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ef   fd„«       «       Zˆ xZS )ÚUniSpeechForPreTrainingr\   c                 ó.  •— t         ‰| �  |«       t        |«      | _        t	        j
                  |j                  «      | _        t        |«      | _	        t	        j                  |j                  |j                  «      | _        t	        j                  |j                  |j                  «      | _        t	        j                  |j                  |j                   «      | _        t	        j
                  |j$                  «      | _        | j)                  «        y ro   )r2   r3   rÞ  rŒ  rK   r¥   Úfeat_quantizer_dropoutÚdropout_featuresrh  Ú	quantizerr£   rm  Úproj_codevector_dimÚ	project_qrM   Úproject_hidÚnum_ctc_classesÚctc_projÚfinal_dropoutr§   rç  r¨   s     €r-   r3   z UniSpeechForPreTraining.__init__™  s¼   ø€ Ü‰Ñ˜Ô Ü'¨Ó/ˆŒÜ "§
¡
¨6×+HÑ+HÓ IˆÔä7¸Ó?ˆŒÜŸ™ 6×#8Ñ#8¸&×:TÑ:TÓUˆŒÜŸ9™9 V×%?Ñ%?À×ASÑASÓTˆÔäŸ	™	 &×"4Ñ"4°f×6LÑ6LÓMˆŒÜ—z‘z &×"6Ñ"6Ó7ˆŒð 	�‰Õr,   rq  c                 ó&   — || j                   _        y)zb
        Set the Gumbel softmax temperature to a given value. Only necessary for training
        N)r  rq  )r5   rq  s     r-   Úset_gumbel_temperaturez.UniSpeechForPreTraining.set_gumbel_temperature¨  s   € ð &1ˆ�‰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.
        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©r5   s    r-   Úfreeze_feature_extractorz0UniSpeechForPreTraining.freeze_feature_extractor®  ó'   € ô
 	�‰ðQäô	
ð
 	×#Ñ#Õ%r,   c                 óL   — | j                   j                  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.
        N©rŒ  rà  r–   r  s    r-   r  z.UniSpeechForPreTraining.freeze_feature_encoderº  ó   € ð
 	�‰×(Ñ(×;Ñ;Õ=r,   Útarget_featuresÚnegative_featuresÚpredicted_featuresc                 óÌ   — t        j                  | |gd¬«      } t        j                  |j                  «       | j                  «       d¬«      }|j	                  | «      }||z  }|S )zé
        Compute logits for contrastive loss based using cosine similarity as the distance measure between
        `[positive_feature, negative_features]` and `[predicted_features]`. Additionally, temperature can be applied.
        r   rÈ   ry   )r(   rÎ   Úcosine_similarityrâ   r  )r  r   r!  rq  Úlogitss        r-   Úcompute_contrastive_logitsz2UniSpeechForPreTraining.compute_contrastive_logitsÁ  sa   € ô  Ÿ)™) _Ð6GÐ$HÈaÔPˆä×(Ñ(Ð);×)AÑ)AÓ)CÀ_×EZÑEZÓE\ÐbdÔeˆØ—‘ Ó0ˆð ˜+Ñ%ˆØˆr,   )ró  r¶  r›   rÂ   rÄ   r/  r0  rÅ   c                 óþ  — |�|n| j                   j                  }| j                  |||||¬«      }|d   }| j                  |d   «      }| j	                  |«      \  }	}
| j                  |	j                  | j
                  j                  j                  «      «      }	| j                  |	«      }	t        j                  |j                  d«      |j                  d«      «      j                  | j                   j                  «      }|j                  dd«      }t        j                   |«      j#                  «       j                  |j$                  «      }|j                  dd«      }|j'                  d«      }|j)                  |d«      |	j)                  | d«      z   }| j+                  |«      }| j-                  |«      }d}|s|�|||	|
f|dd z   S ||	|
f|dd z   S t/        |||	|
|j0                  |j2                  ¬«      S )	aô  
        mask_time_indices (`torch.BoolTensor` of shape `(batch_size, sequence_length)`, *optional*):
            Indices to mask extracted features for contrastive loss. When in training mode, model learns to predict
            masked extracted features in *config.proj_codevector_dim* space.
        sampled_negative_indices (`torch.BoolTensor` of shape `(batch_size, sequence_length, num_negatives)`, *optional*):
            Indices indicating which quantized target vectors are used as negative sampled vectors in contrastive loss.
            Required input for pre-training.

        Returns:

        Example:

        ```python
        >>> import torch
        >>> from transformers import AutoFeatureExtractor, UniSpeechForPreTraining

        >>> feature_extractor = AutoFeatureExtractor.from_pretrained("microsoft/unispeech-large-1500h-cv")
        >>> model = UniSpeechForPreTraining.from_pretrained("microsoft/unispeech-large-1500h-cv")
        >>> # TODO: Add full pretraining example
        ```Nr÷  r   r   ry   rß   r1   )r   r   r    r!   r"   r#   )r\   rù  rŒ  r  r  r  r÷   rG   rñ   r	  r(   ÚemptyrÌ   rœ  Úreplace_probr_   Ú	bernoullirã   r­  r>  Úmasked_fillr§   r  r   r"   r#   )r5   r›   rÂ   rÄ   r/  r0  r$  Útransformer_featuresrø  Úquantized_featuresr!   Úprob_replace_matrixÚsampled_replace_matrixr$  r   s                  r-   r;   zUniSpeechForPreTraining.forwardÕ  s  € ð> &1Ð%<‘kÀ$Ç+Á+×B]ÑB]ˆà—.‘.ØØ)Ø/Ø!5Ø#ð !ó 
ˆð  ' q™zÐð  ×0Ñ0°¸±Ó<ÐØ48·N±NÐCSÓ4TÑ1ÐÐ1ð "Ÿ^™^Ð,>×,AÑ,AÀ$Ç.Á.×BWÑBW×B]ÑB]Ó,^Ó_ÐØ!×-Ñ-Ð.@ÓAÐä#Ÿk™kÐ*>×*CÑ*CÀAÓ*FÐH\×HaÑHaÐbcÓHdÓe×kÑkØ�K‰K×$Ñ$ó
Ðð 2×;Ñ;¸A¸qÓAÐÜ!&§¡Ð1DÓ!E×!JÑ!JÓ!L×!OÑ!OÐPd×PkÑPkÓ!lÐØ!7×!AÑ!AÀ!ÀQÓ!GÐØ!7×!AÑ!AÀ"Ó!EÐØ%×1Ñ1Ð2HÈ#ÓNØ×*Ñ*Ð,BÐ+BÀCÓHñ
ˆð
 —‘˜fÓ%ˆØ—‘˜vÓ&ˆð ˆÙØÐØÐ2Ð4FÐH]Ð^ÐahÐijÐikÐalÑlÐlØ(Ð*<Ð>SÐTÐW^Ð_`Ð_aÐWbÑbÐbä,ØØ1Ø'9Ø"7Ø!×/Ñ/Ø×)Ñ)ô
ð 	
r,   )r   )NNNN)r$   r%   r&   r   r3   rá   r  r  r  r‰  r(   r)   r%  r   rý  r   r   rÿ  r   rä   rã   r   r   r;   r=   r>   s   @r-   r  r  •  s  ø„ ð˜õ ð1°#ó 1ò
&ò>ð ð
 ñ	Ø×*Ñ*ðà ×,Ñ,ðð "×-Ñ-ðð ò	ó ðñ& +Ð+EÓFÙÐ+HÐWfÔgð 26Ø,0Ø/3Ø&*ñM
à˜uŸ|™|Ñ,ðM
ð ! §¡Ñ.ðM
ð $ D™>ð	M
ð
 ' t™nðM
ð ˜d‘^ðM
ð 
ˆuÐ3Ð3Ñ	4òM
ó hó GôM
r,   r  r1   zW'mister quilter is the apposl of the midle classes and weare glad to welcom his gosepl'gìQ¸…+1@zgUniSpeech Model with a `language modeling` head on top for Connectionist Temporal Classification (CTC).a/  
        target_lang (`str`, *optional*):
            Language id of adapter weights. Adapter weights are stored in the format adapter.<lang>.safetensors or
            adapter.<lang>.bin. Only relevant when using an instance of [`UniSpeechForCTC`] with adapters. Uses 'eng'
            by default.
    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UniSpeechForCTCÚ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 zø with a configuration that does not define the vocabulary size of the language model head. Please instantiate the model as follows: `UniSpeechForCTC.from_pretrained(..., vocab_size=vocab_size)`. or define `vocab_size` of your model's configuration.Úadd_adapter)r2   r3   rÞ  rŒ  rK   r¥   r  r§   r1  Ú
vocab_sizerŠ   r7   rQ   r3  Úoutput_hidden_sizerM   r£   Úlm_headrç  )r5   r\   r1  r5  r7   s       €r-   r3   zUniSpeechForCTC.__init__9  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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.
        NrS  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)r1  r_  r\   rŠ   rõ   ÚinfoÚload_adapter)r5   r1  s     r-   Útie_weightszUniSpeechForCTC.tie_weightsP  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$Ø×Ñ˜k°dÐÕ;ð %r,   c                 óX   — t        j                  dt        «       | j                  «        y)r  r  Nr  r  s    r-   r  z(UniSpeechForCTC.freeze_feature_extractore  r  r,   c                 óL   — | j                   j                  j                  «        yr  r  r  s    r-   r  z&UniSpeechForCTC.freeze_feature_encoderq  r  r,   c                 óP   — | j                   j                  «       D ]	  }d|_        Œ y©zÒ
        Calling this function will disable the gradient computation for the base model so that its parameters will not
        be updated during training. Only the classification head will be updated.
        FN©rŒ  r’   r“   r”   s     r-   Úfreeze_base_modelz!UniSpeechForCTC.freeze_base_modelx  ó(   € ð
 —^‘^×.Ñ.Ó0ò 	(ˆEØ"'ˆEÕñ	(r,   )rò  ró  r¶  rõ  Úexpected_lossr›   rÂ   rÄ   r/  r0  ÚlabelsrÅ   c           
      ó¤  — |�|n| j                   j                  }|�I|j                  «       | j                   j                  k\  r"t	        d| j                   j                  › �«      ‚| j                  |||||¬«      }|d   }| j                  |«      }| j                  |«      }	d}
|��b|�|n$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   r3  ry   )rI   rñ   r   F)Úenabled)ÚblankÚ	reductionÚzero_infinity©r   r$  r"   r#   )r\   rù  rÃ  r4  rŠ   rŒ  r§   r6  r(   Ú	ones_liker¯  rª  ru  r÷   Úmasked_selectrK   rÑ   Úlog_softmaxrò   r_   ÚbackendsÚcudnnÚflagsÚctc_lossÚpad_token_idÚctc_loss_reductionÚctc_zero_infinityÚ_HIDDEN_STATES_START_POSITIONr   r"   r#   )r5   r›   rÂ   rÄ   r/  r0  rD  r$  r"   r$  r   r¡  Úlabels_maskÚtarget_lengthsÚflattened_targetsÚ	log_probsÚoutputs                    r-   r;   zUniSpeechForCTC.forward€  s'  € ð0 &1Ð%<‘kÀ$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ro   rü  )r$   r%   r&   r   r  r3   r;  r  r  rA  r   rý  r   rþ  r   rÿ  Ú_CTC_EXPECTED_OUTPUTÚ_CTC_EXPECTED_LOSSr(   rä   rã   r   r   r;   r=   r>   s   @r-   r0  r0  .  sí   ø„ ñ¨H°S©Mõ ò.<ò*
&ò>ò(ñ +Ð+EÓFÙØ&Ø"Ø$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$Ñ	%òD
óó GôD
r,   r0  z˜
    UniSpeech Model with a sequence classification head on top (a linear layer over the pooled output) for tasks like
    SUPERB Keyword Spotting.
    c                   óþ   ‡ — e Zd Zˆ fd„Zd„ Zd„ Zd„ Z ee«       e	e
eed¬«      	 	 	 	 	 ddeej                     deej                     d	ee   d
ee   dee   deej                     deeef   fd„«       «       Zˆ xZS )Ú"UniSpeechForSequenceClassificationc                 óü  •— t         ‰| �  |«       t        |d«      r|j                  rt	        d«      ‚t        |«      | _        |j                  dz   }|j                  r0t        j                  t        j                  |«      |z  «      | _        t        j                  |j                  |j                   «      | _        t        j                  |j                   |j$                  «      | _        | j)                  «        y )Nr3  z`Sequence classification does not support the use of UniSpeech adapters (config.add_adapter=True)r   )r2   r3   rQ   r3  rŠ   rÞ  rŒ  r*  Úuse_weighted_layer_sumrK   rn  r(   rÐ  Úlayer_weightsr£   rM   Úclassifier_proj_sizeÚ	projectorÚ
num_labelsÚ
classifierrç  )r5   r\   Ú
num_layersr7   s      €r-   r3   z+UniSpeechForSequenceClassification.__init__×  sÀ   ø€ Ü‰Ñ˜Ô ä�6˜=Ô)¨f×.@Ò.@ÜØró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r,   c                 óX   — t        j                  dt        «       | j                  «        yr  r  r  s    r-   r  z;UniSpeechForSequenceClassification.freeze_feature_extractorè  r  r,   c                 óL   — | j                   j                  j                  «        yr  r  r  s    r-   r  z9UniSpeechForSequenceClassification.freeze_feature_encoderô  r  r,   c                 óP   — | j                   j                  «       D ]	  }d|_        Œ yr?  r@  r”   s     r-   rA  z4UniSpeechForSequenceClassification.freeze_base_modelû  rB  r,   rñ  )rò  ró  r¶  rô  r›   rÂ   rÄ   r/  r0  rD  rÅ   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   rÈ   ry   r   r1   rß   rJ  )r\   rù  r`  rŒ  rU  r(   ÚstackrK   rÑ   rÒ   ra  r¼   ru  rc  rs  rµ  rÍ   r>  r?  re  r   rd  r   r"   r#   )r5   r›   rÂ   rÄ   r/  r0  rD  r$  r"   Únorm_weightsÚpooled_outputÚpadding_maskÚexpand_padding_maskr$  r   Úloss_fctrZ  s                    r-   r;   z*UniSpeechForSequenceClassification.forward  s  € ð. &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$ÐE¸vÐEä'ØØØ!×/Ñ/Ø×)Ñ)ô	
ð 	
r,   rü  )r$   r%   r&   r3   r  r  rA  r   rý  r   rþ  r   rÿ  r   r(   rä   rã   r   r   r;   r=   r>   s   @r-   r^  r^  Ï  sÒ   ø„ ôò"
&ò>ò(ñ +Ð+EÓFÙØ&Ø,Ø$Ø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 Gô<
r,   r^  )r0  r  r^  rÞ  r‹  r9   )Sr–  r  Údataclassesr   Útypingr   r   r   ÚnumpyrÈ  r(   Útorch.nnrK   r   Úactivationsr	   Úintegrations.deepspeedr
   Úintegrations.fsdpr   Úmodeling_flash_attention_utilsr   r   Úmodeling_outputsr   r   r   r   r   Úmodeling_utilsr   rP   r   r   r   r   r   Úconfiguration_unispeechr   r   Ú
get_loggerr$   rõ   rþ  rÿ  r   ÚModuler/   r@   ra   rr   r{   rƒ   rž   r¬   ræ   rý   r  r  r  r&  rQ  r]  rc  rh  r‹  rá   râ   r¼  ÚndarrayrÜ  r   ÚUNISPEECH_START_DOCSTRINGrý  rú  rÞ  r  rU  r[  r\  r0  r^  Ú__all__r+   r,   r-   ú<module>r�     s3  ðó Û Ý !ß )Ñ )ã Û Ý Ý %å !Ý @Ý 7ß h÷õ õ .÷õ õ 5ñ ÔÝJð 
ˆ×	Ñ	˜HÓ	%€ð HÐ ð $€ð ô : Kó  :ó ð :ôF˜BŸI™Iô ô* r§y¡yô *ôZ B§I¡Iô ô* "§)¡)ô ô6 "§)¡)ô ô0,˜bŸi™iô ,ô^1 §¡ô 1ô[B˜Ÿ™ô [Bô|{9Ð1ô {9ô|g1Ð/ô g1ôT˜2Ÿ9™9ô ð2  Ø"Ø1ñÐ ô ˜BŸI™Iô  ôFR
�r—y‘yô R
ôj §	¡	ô ô2*¨2¯9©9ô *ôZV
 b§i¡iô V
ôrC' R§Y¡Yô C'ôLJ˜ô Jðb 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òn (Ð ðÐ ð$"Ð ðJ 3Ð ñ ØiØóôu
Ð-ó u
ó	ðu
ñp ØZÐ\uóôL
Ð6ó L
óðL
ð^ !"Ð ð qÐ ØÐ ñ ØqØðó	ôT
Ð.ó T
ó	ðT
ñn ðð óôp
Ð)Aó p
óðp
òf�r,   