
    <h                       S SK r S SKrS SKJr  S SKJrJrJr  S SKrS SK	J
s  Jr  S SKrS SKJ
r
  S SKJrJrJr  SSKJr  SSKJr  SS	KJr  SS
KJr  SSKJrJrJrJrJrJ r   SSK!J"r"  SSK#J$r$J%r%J&r&  SSK'J(r(  SSK)J*r*J+r+  \%" 5       (       a  S SK,J-r-  S SK.J/r/  S SK0J1r1  O\2r/\&Rf                  " \45      r5 " S S\5      r6  STS\Rn                  S\Rn                  S\\Rn                     S\\Rn                     S\8\Rn                  \Rn                  \Rn                  \9\\Rn                     \\Rn                     4   4
S jjr:S\Rn                  S\Rn                  S\9S\9S\Rn                  4
S jr; " S  S!\Rx                  Rz                  5      r>  STS"\\Rn                     S#\\9   4S$ jjr? " S% S&\/5      r@ " S' S(\
R                  5      rB " S) S*\
R                  5      rC " S+ S,\*5      rD SUS-S.S/\Rn                  S\Rn                  S0\Rn                  S\\R                     S1\8\9\94   S2\9S3\9S4\\F   S\\8\Rn                  \Rn                  4   \8\Rn                     4   4S5 jjrG\R                  4S-S.S/\Rn                  S6\@S"\Rn                  S#\9S1\8\9\94   S2\9S3\9S7\R                  S\8\Rn                     4S8 jjrJS-S.S/\Rn                  S\Rn                  S0\Rn                  S\\R                     S1\8\9\94   S2\9S3\9S\8\Rn                     4S9 jrK\J\G\KS:.rL " S; S.\
R                  5      rM " S< S=\5      rN\$ " S> S?\"5      5       rO\$ " S@ SA\O5      5       rP " SB SC\
R                  5      rQ\$" SDSE9 " SF SG\O5      5       rR\$" SHSE9 " SI SJ\O5      5       rS\$" SKSE9 " SL SM\O5      5       rT\$ " SN SO\O5      5       rU\$" SPSE9 " SQ SR\O5      5       rV/ SSQrWg)V    N)nullcontext)LiteralOptionalUnion)nn)BCEWithLogitsLossCrossEntropyLossMSELoss   )ACT2FN)PretrainedConfig)_prepare_4d_attention_mask)GradientCheckpointingLayer)BaseModelOutputMaskedLMOutputMultipleChoiceModelOutputQuestionAnsweringModelOutputSequenceClassifierOutputTokenClassifierOutput)PreTrainedModel)auto_docstringis_flash_attn_2_availablelogging)is_triton_available   )GemmaRotaryEmbeddingapply_rotary_pos_emb) flash_attn_varlen_qkvpacked_func)RotaryEmbedding)apply_rotaryc                      ^  \ rS rSrSrSrSS0rS/r                                   SS\S   4U 4S	 jjjr	U 4S
 jr
SrU =r$ )ModernBertConfig8   a  
This is the configuration class to store the configuration of a [`ModernBertModel`]. It is used to instantiate an ModernBert
model according to the specified arguments, defining the model architecture. Instantiating a configuration with the
defaults will yield a similar configuration to that of the ModernBERT-base.
e.g. [answerdotai/ModernBERT-base](https://huggingface.co/answerdotai/ModernBERT-base)

Configuration objects inherit from [`PretrainedConfig`] and can be used to control the model outputs. Read the
documentation from [`PretrainedConfig`] for more information.

Args:
    vocab_size (`int`, *optional*, defaults to 50368):
        Vocabulary size of the ModernBert model. Defines the number of different tokens that can be represented by the
        `inputs_ids` passed when calling [`ModernBertModel`]
    hidden_size (`int`, *optional*, defaults to 768):
        Dimension of the hidden representations.
    intermediate_size (`int`, *optional*, defaults to 1152):
        Dimension of the MLP representations.
    num_hidden_layers (`int`, *optional*, defaults to 22):
        Number of hidden layers in the Transformer decoder.
    num_attention_heads (`int`, *optional*, defaults to 12):
        Number of attention heads for each attention layer in the Transformer decoder.
    hidden_activation (`str` or `function`, *optional*, defaults to `"gelu"`):
        The non-linear activation function (function or string) in the decoder. Will default to `"gelu"`
        if not specified.
    max_position_embeddings (`int`, *optional*, defaults to 8192):
        The maximum sequence length that this model might ever be used with.
    initializer_range (`float`, *optional*, defaults to 0.02):
        The standard deviation of the truncated_normal_initializer for initializing all weight matrices.
    initializer_cutoff_factor (`float`, *optional*, defaults to 2.0):
        The cutoff factor for the truncated_normal_initializer for initializing all weight matrices.
    norm_eps (`float`, *optional*, defaults to 1e-05):
        The epsilon used by the rms normalization layers.
    norm_bias (`bool`, *optional*, defaults to `False`):
        Whether to use bias in the normalization layers.
    pad_token_id (`int`, *optional*, defaults to 50283):
        Padding token id.
    eos_token_id (`int`, *optional*, defaults to 50282):
        End of stream token id.
    bos_token_id (`int`, *optional*, defaults to 50281):
        Beginning of stream token id.
    cls_token_id (`int`, *optional*, defaults to 50281):
        Classification token id.
    sep_token_id (`int`, *optional*, defaults to 50282):
        Separation token id.
    global_rope_theta (`float`, *optional*, defaults to 160000.0):
        The base period of the global RoPE embeddings.
    attention_bias (`bool`, *optional*, defaults to `False`):
        Whether to use a bias in the query, key, value and output projection layers during self-attention.
    attention_dropout (`float`, *optional*, defaults to 0.0):
        The dropout ratio for the attention probabilities.
    global_attn_every_n_layers (`int`, *optional*, defaults to 3):
        The number of layers between global attention layers.
    local_attention (`int`, *optional*, defaults to 128):
        The window size for local attention.
    local_rope_theta (`float`, *optional*, defaults to 10000.0):
        The base period of the local RoPE embeddings.
    embedding_dropout (`float`, *optional*, defaults to 0.0):
        The dropout ratio for the embeddings.
    mlp_bias (`bool`, *optional*, defaults to `False`):
        Whether to use bias in the MLP layers.
    mlp_dropout (`float`, *optional*, defaults to 0.0):
        The dropout ratio for the MLP layers.
    decoder_bias (`bool`, *optional*, defaults to `True`):
        Whether to use bias in the decoder layers.
    classifier_pooling (`str`, *optional*, defaults to `"cls"`):
        The pooling method for the classifier. Should be either `"cls"` or `"mean"`. In local attention layers, the
        CLS token doesn't attend to all tokens on long sequences.
    classifier_dropout (`float`, *optional*, defaults to 0.0):
        The dropout ratio for the classifier.
    classifier_bias (`bool`, *optional*, defaults to `False`):
        Whether to use bias in the classifier.
    classifier_activation (`str`, *optional*, defaults to `"gelu"`):
        The activation function for the classifier.
    deterministic_flash_attn (`bool`, *optional*, defaults to `False`):
        Whether to use deterministic flash attention. If `False`, inference will be faster but not deterministic.
    sparse_prediction (`bool`, *optional*, defaults to `False`):
        Whether to use sparse prediction for the masked language model instead of returning the full dense logits.
    sparse_pred_ignore_index (`int`, *optional*, defaults to -100):
        The index to ignore for the sparse prediction.
    reference_compile (`bool`, *optional*):
        Whether to compile the layers of the model which were compiled during pretraining. If `None`, then parts of
        the model will be compiled if 1) `triton` is installed, 2) the model is not on MPS, 3) the model is not
        shared between devices, and 4) the model is not resized after initialization. If `True`, then the model may
        be faster in some scenarios.
    repad_logits_with_grad (`bool`, *optional*, defaults to `False`):
        When True, ModernBertForMaskedLM keeps track of the logits' gradient when repadding for output. This only
        applies when using Flash Attention 2 with passed labels. Otherwise output logits always have a gradient.

Examples:

```python
>>> from transformers import ModernBertModel, ModernBertConfig

>>> # Initializing a ModernBert style configuration
>>> configuration = ModernBertConfig()

>>> # Initializing a model from the modernbert-base style configuration
>>> model = ModernBertModel(configuration)

>>> # Accessing the model configuration
>>> configuration = model.config
```
modernbert
rope_thetaglobal_rope_thetapast_key_valuesclassifier_poolingclsmeanc$           	        > [         T%U ]  " SUUUUUS.U$D6  Xl        Xpl        X l        X0l        X@l        XPl        Xl        Xl	        Xl
        Xl        UU l        UU l        UU l        X`l        UU l        UU l        UU l        UU l        UU l        UU l        UU l        UU l        UU l        UU l        UU l        UU l        U U l        U!U l        U"U l        U#U l        U R.                  S;  a  [A        SU R.                   S35      eg )N)pad_token_idbos_token_ideos_token_idcls_token_idsep_token_idr)   zQInvalid value for `classifier_pooling`, should be either "cls" or "mean", but is . )!super__init__
vocab_sizemax_position_embeddingshidden_sizeintermediate_sizenum_hidden_layersnum_attention_headsinitializer_rangeinitializer_cutoff_factornorm_eps	norm_biasr&   attention_biasattention_dropouthidden_activationglobal_attn_every_n_layerslocal_attentionlocal_rope_thetaembedding_dropoutmlp_biasmlp_dropoutdecoder_biasr(   classifier_dropoutclassifier_biasclassifier_activationdeterministic_flash_attnsparse_predictionsparse_pred_ignore_indexreference_compilerepad_logits_with_grad
ValueError)&selfr6   r8   r9   r:   r;   rB   r7   r<   r=   r>   r?   r-   r/   r.   r0   r1   r&   r@   rA   rC   rD   rE   rF   rG   rH   rI   r(   rJ   rK   rL   rM   rN   rO   rP   rQ   kwargs	__class__s&                                        i/var/www/html/shao/venv/lib/python3.13/site-packages/transformers/models/modernbert/modular_modernbert.pyr5   ModernBertConfig.__init__   s>   N 	 	
%%%%%	
 	
 %'>$&!2!2#6 !2)B& "!2,!2!2*D'. 0!2 &("4"4.%:"(@%!2(@%!2&<#""/9cdhd{d{c||}~  :    c                 H   > [         TU ]  5       nUR                  SS 5        U$ )NrP   )r4   to_dictpop)rS   outputrU   s     rV   rZ   ModernBertConfig.to_dict   s#    "

&-rX   )r@   rA   rL   rK   rJ   r(   rI   rM   rF   rC   r&   rB   r8   r=   r<   r9   rD   rE   r7   rG   rH   r?   r>   r;   r:   rP   rQ   rO   rN   r6   )#i  i   i        gelui    g{Gz?       @gh㈵>Fik  j  i  rc   rb   g     AF        r           @rd   Frd   Tr*   rd   Fr`   FFiNF)__name__
__module____qualname____firstlineno____doc__
model_typeattribute_mapkeys_to_ignore_at_inferencer   r5   rZ   __static_attributes____classcell__rU   s   @rV   r"   r"   8   s    eN J!#67M#4"5   $"%"#$ 5:$!&!%$IQ8 $M29Q Qf rX   r"   inputsattention_maskposition_idslabelsreturnc                    UR                  S[        R                  S9n[        R                  " UR	                  5       SS9R	                  5       n[        UR                  5       R                  5       5      n[        R                  R                  R                  [        R                  " US[        R                  S9S5      nU R                  5       S:X  a  U R	                  5       U   nO(U R                  tpnX-  nU R                  " U/UQ76 U   nUb  UR	                  5       U   OSnUb  UR	                  5       U   OSnXXvX4$ )	aP  
Remove padding from input sequences.

Args:
    inputs: (batch, seqlen, ...) or (batch, seqlen)
    attention_mask: (batch, seqlen), bool / int, 1 means valid and 0 means not valid.
    position_ids: (batch, seqlen), int, position ids
    labels: (batch, seqlen), int, labels

Returns:
    unpadded_inputs: (total_nnz, ...), where total_nnz = number of tokens selected in attention_mask.
    indices: (total_nnz)
    cu_seqlens: (batch + 1), the cumulative sequence lengths
    max_seqlen_in_batch: int
    unpadded_position_ids: (total_nnz) or None
    unpadded_labels: (total_nnz) or None
dimdtypeF)as_tupler   )   r   r   N)sumtorchint32nonzeroflattenintmaxitemr   
functionalpadcumsumrz   shapeview)rr   rs   rt   ru   seqlens_in_batchindicesmax_seqlen_in_batch
cu_seqlensunpadded_inputsbatchseqlenrestr   unpadded_position_idsunpadded_labelss                  rV   _unpad_modernbert_inputr      s   . &))b)DmmN224uEMMOG.22499;<$$((6FAUZU`U`)acijJzz|q ..*73%|| ++e3d3G<?K?WL0027;]a393Efnn&w/4OZF[llrX   r   r   r   c                 ^   U R                  5       S:X  aC  [        R                  " X#-  U R                  U R                  S9nXU'   UR                  X#5      nU$ U R                  tpg[        R                  " X#-  /UQ7U R                  U R                  S.6nXU'   UR
                  " X#/UQ76 nU$ )a-  
Add padding to sequences.

Args:
    inputs: (total_nnz, ...) or (total_nnz,), where total_nnz = number of tokens selected in attention_mask.
    indices: (total_nnz)
    batch: int, batch size
    seqlen: int, max sequence length

Returns:
    padded_inputs: (batch, seqlen, ...) or (batch, seqlen)
r}   )r{   device)rz   r   zerosr{   r   r   r   )rr   r   r   r   r\   padded_inputs_r   s           rV   _pad_modernbert_outputr   &  s    $ zz|qU^6<<V wE2  <<U^]d]&,,v}}] wE9D9rX   c                   h    \ rS rSr\  SS\\R                     S\\   4S jj5       r	\S 5       r
Srg)	ApplyRotaryEmbUnpadiE  Nr   
max_seqlenc                     UR                  5       nUR                  u  pgpUS S 2S S24   R                  USU	5      n
[        U
UUSUUSSS9  U R	                  X#U5        XPl        U$ )Nr   rx   r   FT)seqlen_offsetsr   r   interleavedinplace)
contiguousr   r   r    save_for_backwardr   )ctxqkvcossinr   r   	total_nnz_three_nheadsheaddimqks              rV   forwardApplyRotaryEmbUnpad.forwardF  sz     nn.1ii+	7 BQBZ__YG4!!		
 	c
3#
rX   c                     U R                   u  p#nUR                  5       nUR                  u  pVpxUS S 2S S24   R                  USU5      n	[	        U	UUSUU R
                  SSSS9	  US S S S S S 4$ )Nr   rx   r   FT)r   r   r   r   r   	conjugate)saved_tensorsr   r   r   r    r   )
r   dor   r   r   r   r   r   r   dqks
             rV   backwardApplyRotaryEmbUnpad.backwarde  s    "00*]]_.0hh+	7 BQBinnYG4!~~
	
 4tT455rX   r3   NN)rg   rh   ri   rj   staticmethodr   r   Tensorr   r   r   ro   r3   rX   rV   r   r   E  sQ     .2$(
 U\\* SM < 6 6rX   r   r   r   c                 0    [         R                  XX#U5      $ )a  
Arguments:
    qkv: (total_nnz, 3, nheads, headdim) - input tensor for packed QKV.
    cos, sin: (seqlen_rotary, rotary_dim / 2)
    interleaved: if True, rotate pairs of even and odd dimensions (GPT-J style) instead
        of 1st half and 2nd half (GPT-NeoX style).
    inplace: if True, apply rotary embedding in-place.
    seqlen_offsets: (batch_size,) or int. Each sequence in x is shifted by this amount.
        Most commonly used in inference when we have KV cache.
    cu_seqlens: (batch + 1,) or None
    max_seqlen: int
Return:
    out: (total_nnz, dim)
rotary_dim must be <= headdim
Apply rotary embedding to the first rotary_dim of x.
)r   apply)r   r   r   r   r   s        rV   apply_rotary_unpaddedr   |  s    . $$Ss
KKrX   c                   6  ^  \ rS rSrSr    SS\S\S\\   S\\R                     S\\R                     4
U 4S jjjr SS	\R                  S
\R                  S\\   S\\R                  \\R                  \R                  4   4   4S jjrS\4S jrSrU =r$ )!ModernBertUnpaddedRotaryEmbeddingi  zH
The rotary position embeddings applied directly to unpadded sequences.
rz   baser   r   r{   c                 h   > [         TU ]  XUSS9  X0l        Ub  Ub  Ub  U R                  X4US9  gggg)z
max_seqlen: if max_seqlen, device, and dtype are provided, we precompute the cos_sin_cache
    up to max_seqlen. If the max_seqlen, device, or dtype during training/inference differ,
    the cos_sin_cache will be recomputed during the forward pass.
F)rz   r   r   r   Nr   r{   )r4   r5   r   _update_cos_sin_cache)rS   rz   r   r   r   r{   rU   s         rV   r5   *ModernBertUnpaddedRotaryEmbedding.__init__  sM     	SFN$!f&8U=N&&z&N >O&8!rX   r   r   rv   c                     Ub$  U R                  X1R                  UR                  S9  [        UU R                  U R
                  UUS9nU$ )z
Apply rotary embedding *inplace* to qkv.
qkv: (total_nnz, 3, nheads, headdim)
cu_seqlens: (batch + 1,) cumulative sequence lengths
max_seqlen: int max seq length in the batch
r   r   r   )r   r   r{   r   _cos_cached_sin_cached)rS   r   r   r   s       rV   r   )ModernBertUnpaddedRotaryEmbedding.forward  sQ     !&&z**CII&V#!!
 
rX   c                 T    SU R                    SU R                   SU R                   3$ )Nzdim=z, base=z, scale_base=)rz   r   
scale_baserS   s    rV   
extra_repr,ModernBertUnpaddedRotaryEmbedding.extra_repr  s(    dhhZwtyykt>OPPrX   )r   )rf   NNNN)rg   rh   ri   rj   rk   r   floatr   r   r   r{   r5   r   r   tupler   strr   ro   rp   rq   s   @rV   r   r     s     $()-'+OO O SM	O
 &O $O O. %)	\\ LL SM	
 
u||U5<<#=>>	?2QC Q QrX   r   c                      ^  \ rS rSrSrS\4U 4S jjr\R                  " SS9S\R                  S\R                  4S	 j5       r SS\\R                     S
\\R                     S\R                  4S jjrSrU =r$ )ModernBertEmbeddingsi  zN
Same as BertEmbeddings with a tiny tweak for positional embeddings indexing.
configc                 \  > [         TU ]  5         Xl        [        R                  " UR
                  UR                  UR                  S9U l        [        R                  " UR                  UR                  UR                  S9U l        [        R                  " UR                  5      U l        g )N)padding_idxepsbias)r4   r5   r   r   	Embeddingr6   r8   r-   tok_embeddings	LayerNormr>   r?   normDropoutrF   droprS   r   rU   s     rV   r5   ModernBertEmbeddings.__init__  su     ll6+<+<f>P>P^d^q^qrLL!3!3vO_O_`	JJv778	rX   Tdynamic	input_idsrv   c                 `    U R                  U R                  U R                  U5      5      5      $ r   )r   r   r   )rS   r   s     rV   compiled_embeddings(ModernBertEmbeddings.compiled_embeddings  s%    yy4#6#6y#ABCCrX   inputs_embedsc                    Ub"  U R                  U R                  U5      5      nU$ U R                  R                  (       a  U R	                  U5      O.U R                  U R                  U R                  U5      5      5      nU$ r   )r   r   r   rP   r   r   )rS   r   r   hidden_statess       rV   r   ModernBertEmbeddings.forward  su     $ IIdii&>?M  ;;00 ((3YYtyy)<)<Y)GHI 
 rX   )r   r   r   r   r   )rg   rh   ri   rj   rk   r"   r5   r   compile
LongTensorr   r   r   r   ro   rp   rq   s   @rV   r   r     s    9/ 9 ]]4 DU-=-= D%,, D !D ei!%"2"23KSTYT`T`Ka	 rX   r   c                   n   ^  \ rS rSrSrS\4U 4S jjrS\R                  S\R                  4S jr	Sr
U =r$ )	ModernBertMLPi  a*  Applies the GLU at the end of each ModernBERT layer.

Compared to the default BERT architecture, this block replaces :class:`~transformers.model.bert.modeling_bert.BertIntermediate`
and :class:`~transformers.model.bert.modeling_bert.SelfOutput` with a single module that has similar functionality.
r   c                   > [         TU ]  5         Xl        [        R                  " UR
                  [        UR                  5      S-  UR                  S9U l	        [        UR                     U l        [        R                  " UR                  5      U l        [        R                  " UR                  UR
                  UR                  S9U l        g )Nr   r   )r4   r5   r   r   Linearr8   r   r9   rG   Wir   rB   actr   rH   r   Wor   s     rV   r5   ModernBertMLP.__init__  s    ))F..F4L4L0MPQ0QX^XgXgh&223JJv112	))F44f6H6Hv_rX   r   rv   c                     U R                  U5      R                  SSS9u  p#U R                  U R                  U R	                  U5      U-  5      5      $ )Nr   rx   rz   )r   chunkr   r   r   )rS   r   inputgates       rV   r   ModernBertMLP.forward  sG    ggm,221"2=wwtyy%4!7899rX   )r   r   r   r   r   )rg   rh   ri   rj   rk   r"   r5   r   r   r   ro   rp   rq   s   @rV   r   r     s7    `/ `:U\\ :ell : :rX   r   c                       \ rS rSrSrg)ModernBertRotaryEmbeddingi  r3   N)rg   rh   ri   rj   ro   r3   rX   rV   r   r     s    rX   r   moduleModernBertAttentionr   sliding_window_maskrD   bsrz   output_attentionsc	                    U R                  XS9u  pUR                  SS5      R                  SS9u  pn[        XX5      u  pU R                  S-  n[
        R                  " XR                  SS5      5      U-  nUS:w  a  UnUU-   n[        R                  R                  US[
        R                  S	9R                  UR                  5      n[        R                  R                  UU R                  U R                  S
9n[
        R                  " UU5      nUR                  SS5      R!                  5       nUR#                  USU5      nU(       a  UU4$ U4$ )Nrt   r   r}   r   r         ࿩rx   rx   rx   ry   )ptraining)
rotary_emb	transposeunbindr   head_dimr   matmulr   r   softmaxfloat32tor{   dropoutrA   r	  r   r   )r   r   rs   r  rt   rD   r  rz   r  _kwargsr   r   querykeyvaluescaleattn_weightsattn_outputs                     rV   eager_attention_forwardr    s=       @HCa+22q29E%e#;JEOOT!E<<}}Q':;eCL(",.0L ==((2U]](SVVW\WbWbcL==((9Q9Q\b\k\k(lL,,|U3K''1-88:K""2r3/K\**>rX   r
  target_dtypec	           	         U" XUS9nUR                   [        R                  [        R                  4;  n
U
(       ad  UR                   nUR	                  U5      n[        UUUU R                  (       a  U R                  OSU R                  US9nUR	                  U5      nO5[        UUUU R                  (       a  U R                  OSU R                  US9nUR                  Xg5      4$ )Nr   rd   )r   r   	dropout_pdeterministicwindow_size)
r{   r   float16bfloat16r  r   r	  rA   rM   r   )r   r   r
  r   r   rD   r  rz   r  r  convert_dtype
orig_dtypeattns                rV   flash_attention_forwardr%  &  s     SJ
GCIIemmU^^%DDM YY
ff\"/!!28//f..s 99'
 wwz"/!!28//f..s 99'
 IIb  rX   c                 f   U R                  XS9u  pUR                  SS5      R                  SS9u  pn[        XX5      u  pUS:w  a  Un[        R
                  " UUUU R                  (       a  U R                  OSUS9R                  SS5      R                  5       nUR                  US	U5      nU4$ )
Nr  r   r}   r   r   r  rd   )r  	attn_maskrx   )
r
  r  r  r   Fscaled_dot_product_attentionr	  rA   r   r   )r   r   rs   r  rt   rD   r  rz   r  r   r   r  r  r  r  s                  rV   sdpa_attention_forwardr*  Q  s        @HCa+22q29E%e#;JE(", 	
&&28//f..s$	
 
1a	  ""2r3/K>rX   )flash_attention_2eagersdpac                      ^  \ rS rSrSrSS\S\\   4U 4S jjjr SS\	R                  S\\   S\	R                  4S	 jjrS
rU =r$ )r   i{  an  Performs multi-headed self attention on a batch of unpadded sequences.

If Flash Attention 2 is installed, this module uses Flash Attention to improve throughput.
If Flash Attention 2 is not installed, the implementation will use PyTorch's SDPA kernel,
which requires padding and unpadding inputs, adding some overhead.

See `forward` method for additional details.
r   layer_idc                   > [         TU ]  5         Xl        X l        UR                  UR
                  -  S:w  a&  [        SUR                   SUR
                   S35      eUR                  U l        UR                  U l        UR
                  U l	        UR                  UR
                  -  U l
        U R                  U R                  -  U l        [        R                  " UR                  SU R                  -  UR                  S9U l        X!R                   -  S:w  aU  UR"                  S-  UR"                  S-  4U l        UR$                  b  UR$                  OUR&                  nUR"                  nOSU l        UR(                  nUR&                  nUR*                  S	:X  a  [-        U R                  XCS
9U l        O*[0        R2                  " U5      nX5l        [7        US9U l        [        R                  " UR                  UR                  UR                  S9U l        UR                  S:  a   [        R:                  " UR                  5      O[        R<                  " 5       U l        [A        5       U l!        g )Nr   zThe hidden size (z6) is not a multiple of the number of attention heads ()r   r   r   r  r+  )rz   r   r   )r   rd   )"r4   r5   r   r/  r8   r;   rR   rA   rM   	num_headsr  all_head_sizer   r   r@   WqkvrC   rD   rE   r&   r7   _attn_implementationr   r
  copydeepcopyr%   r   r   r   Identityout_dropsetpruned_heads)rS   r   r/  r%   r7   config_copyrU   s         rV   r5   ModernBertAttention.__init__  s     : ::a?#F$6$6#77mnt  oI  oI  nJ  JK  L  "(!9!9(.(G(G%33**f.H.HH!]]T^^;IIf00!d6H6H2HvOdOde	7771<$*$:$:a$?AWAW[\A\#]D 4:4K4K4W00]c]u]uJ&,&<&<##+D &,&D&D#11J&&*==?MM.EDO --/K%/"7{KDO))F..0B0BI^I^_@F@X@X[^@^

6#;#;<dfdododqErX   r   r  rv   c           
         U R                  U5      nUR                  S   nU R                  R                  S:X  a)  UR	                  SSU R
                  U R                  5      nO)UR	                  USSU R
                  U R                  5      n[        U R                  R                     " U 4UU R                  U R                  UU R                  US.UD6nUS   nU R                  U R                  U5      5      nU4USS  -   $ )Nr   r+  rx   r   )r   r
  rD   r  rz   r  r}   )r4  r   r   r5  r   r2  r  MODERNBERT_ATTENTION_FUNCTIONr
  rD   r3  r9  r   )rS   r   r  rT   r   r  attn_outputss          rV   r   ModernBertAttention.forward  s     ii&  #;;++/BB((2q$..$--@C((2r1dnndmmDC4T[[5U5UV	
 00""/	
 	
 %Qdggm&<=,qr"222rX   )r   r4  r3  rA   r   rM   r  r/  rD   r2  r9  r;  r
  r   F)rg   rh   ri   rj   rk   r"   r   r   r5   r   r   boolr   ro   rp   rq   s   @rV   r   r   {  s]    %"/ %"8C= %" %"T -23||3 $D>3
 
3 3rX   c                   t  ^  \ rS rSrSS\S\\   4U 4S jjjr\R                  " SS9S\R                  S\R                  4S	 j5       r      SS\R                  S
\\R                     S\\R                     S\\R                     S\\R                     S\\   S\\   S\R                  4S jjrSrU =r$ )ModernBertEncoderLayeri  r   r/  c                   > [         TU ]  5         Xl        US:X  a  [        R                  " 5       U l        O9[        R                  " UR                  UR                  UR                  S9U l        [        XS9U l        [        R                  " UR                  UR                  UR                  S9U l        [        U5      U l        g )Nr   r   )r   r/  )r4   r5   r   r   r8  	attn_normr   r8   r>   r?   r   r$  mlp_normr   mlprS   r   r/  rU   s      rV   r5   ModernBertEncoderLayer.__init__  s    q=[[]DN\\&*<*<&//X^XhXhiDN'vI	V%7%7V__SYScScd (rX   Tr   r   rv   c                 B    U R                  U R                  U5      5      $ r   )rI  rH  rS   r   s     rV   compiled_mlp#ModernBertEncoderLayer.compiled_mlp  s    xxm455rX   rs   r  rt   r   r   r  c           
      
   U R                  U R                  U5      UUUUUUS9nXS   -   nU R                  R                  (       a  U R	                  U5      OU R                  U R                  U5      5      n	X-   nU4USS  -   $ )Nrs   r  rt   r   r   r  r   r}   )r$  rG  r   rP   rN  rI  rH  )
rS   r   rs   r  rt   r   r   r  r@  
mlp_outputs
             rV   r   ModernBertEncoderLayer.forward  s     yyNN=)) 3%!!/ ! 
 &Q7 {{,, m,$--67 	
 &2,qr"222rX   )r$  rG  r   rI  rH  r   )NNNNNF)rg   rh   ri   rj   r"   r   r   r5   r   r   r   rN  r   rC  r   ro   rp   rq   s   @rV   rE  rE    s    	)/ 	)8C= 	) 	) ]]4 6%,, 65<< 6 !6 266:37-1$(,13||3 !.3 &ell3	3
 u//03 U\\*3 SM3 $D>3 
3 3rX   rE  c                      ^  \ rS rSr% \\S'   SrSrSS/rSr	Sr
SrS\R                  4S	 jr SS
\\   S\S\4U 4S jjjrS rU 4S jrSrU =r$ )ModernBertPreTrainedModeli  r   modelTr   rE  Fr   c                   ^ U R                   R                  mTc  SmS[        R                  S[        4U4S jjnU R                   R
                  U R                   R
                  [        R                  " SU R                   R                  -  5      -  U R                   R
                  U R                   R                  S-  S.n[        U[        5      (       a  U" UR                  US   5        g [        U[        5      (       a-  U" UR                  US	   5        U" UR                  US
   5        g [        U[         5      (       a-  U" UR"                  US	   5        U" UR                  US
   5        g [        U[$        5      (       a  U" UR&                  US
   5        g [        U[(        5      (       a  U" UR*                  US
   5        g [        U[,        [.        [0        [2        45      (       a  U" UR4                  US   5        g [        U[        R6                  5      (       aX  UR8                  R:                  R=                  S5        UR>                  b%  UR>                  R:                  RA                  5         g g g )Nr   r   stdc                   > [         R                  R                  U R                  SUT* U-  TU-  S9  [	        U [         R
                  5      (       a8  U R                  b*  [         R                  R                  U R                  5        g g g )Nrd   )r+   rX  ab)r   inittrunc_normal_weight
isinstancer   r   zeros_)r   rX  cutoff_factors     rV   init_weight<ModernBertPreTrainedModel._init_weights.<locals>.init_weight  st    GG!! .3&#% "  &")),,;;*GGNN6;;/ + -rX   ra   r  )inout	embedding	final_outrf  rd  re  rg  g      ?)!r   r=   r   Moduler   r<   mathsqrtr:   r8   r_  r   r   r   r   r   r   r4  ModernBertPredictionHeaddenseModernBertForMaskedLMdecoder#ModernBertForSequenceClassificationModernBertForMultipleChoice ModernBertForTokenClassificationModernBertForQuestionAnswering
classifierr   r^  datafill_r   zero_)rS   r   rb  stdsra  s       @rV   _init_weights'ModernBertPreTrainedModel._init_weights  s   == M	0		 	0 	0 ++//;;00499S4;;C`C`=`3aa6600$6	
 f233--tK/@A..		4:.		4;/ 344T$Z0		4;/ 899d5k2 566U43+0.	
 
 ))4+<=--MM$$S){{&  &&( ' .rX   attn_implementationis_init_checkrv   c                    >  Uc  U R                  5       (       a  SOUn[        TU ]  XS9$ ! [        [        4 a     Nf = f)zB
Checks and dispatches to hhe requested attention implementation.
r+  )rz  r{  )_flash_attn_2_can_dispatchrR   ImportErrorr4   %_check_and_adjust_attn_implementation)rS   rz  r{  rU   s      rV   r  ?ModernBertPreTrainedModel._check_and_adjust_attn_implementation6  s`    	 '.43R3R3T3T $(   w< 3 = 
 	
 K( 		s   , ??c                    U R                   R                  SL a  g [        U S5      (       aZ  [        U R                  5      S:  aA  U R                   R                  (       a  [
        R                  S5        SU R                   l        U R                  R                  S:X  aA  U R                   R                  (       a  [
        R                  S5        SU R                   l        U R                  R                  S:X  aA  U R                   R                  (       a  [
        R                  S5        SU R                   l        U R                   R                  c  [        5       U R                   l        g g )	NFhf_device_mapr}   zqIf `accelerate` split the model across devices, `torch.compile` will not work. Falling back to non-compiled mode.mpsz|Compiling the model with `torch.compile` and using a `torch.mps` device is not supported. Falling back to non-compiled mode.cpuz|Compiling the model with `torch.compile` and using a `torch.cpu` device is not supported. Falling back to non-compiled mode.)
r   rP   hasattrlenr  loggerwarning_oncer   typer   r   s    rV   _maybe_set_compile,ModernBertPreTrainedModel._maybe_set_compileM  s   ;;((E14))c$2D2D.E.I{{,,##9 -2DKK);;u${{,,##9 -2DKK);;u${{,,##9 -2DKK);;((0,?,ADKK) 1rX   c                    > [         TU ]  " U0 UD6nU R                  R                  S;   aA  U R                  R                  (       a  [        R                  S5        SU R                  l        U$ )N>   NTzcResizing token embeddings with `torch.compile` is not supported. Falling back to non-compiled mode.F)r4   resize_token_embeddingsr   rP   r  r  )rS   argsrT   model_embedsrU   s       rV   r  1ModernBertPreTrainedModel.resize_token_embeddingsl  s[    w6GG;;((L8{{,,##y -2DKK)rX   r3   rB  )rg   rh   ri   rj   r"   __annotations__base_model_prefixsupports_gradient_checkpointing_no_split_modules_supports_flash_attn_supports_sdpa_supports_flex_attnr   rh  rx  r   r   rC  r  r  r  ro   rp   rq   s   @rV   rU  rU    s    &*#/1IJN2)BII 2)j IN
#+C=
AE
	
 
.B>
 
rX   rU  c            !         ^  \ rS rSrS\4U 4S jjrS rS r\             SS\	\
R                     S\	\
R                     S\	\
R                     S	\	\
R                     S
\	\
R                     S\	\
R                     S\	\
R                     S\	\   S\	\   S\	\   S\	\   S\	\   S\	\   S\\\
R                  S4   \4   4S jj5       rS\
R                  S\S\
R                  4S jrSrU =r$ )ModernBertModeliy  r   c           	        > [         TU ]  U5        Xl        [        U5      U l        [
        R                  " [        UR                  5       Vs/ sH  n[        X5      PM     sn5      U l
        [
        R                  " UR                  UR                  UR                  S9U l        SU l        U R#                  5         g s  snf )Nr   F)r4   r5   r   r   
embeddingsr   
ModuleListranger:   rE  layersr   r8   r>   r?   
final_normgradient_checkpointing	post_initrJ  s      rV   r5   ModernBertModel.__init__{  s     .v6mmFKFLdLdFefFe(#F5Fef
 ,,v'9'9vU[UeUef&+#	 gs   B?c                 .    U R                   R                  $ r   r  r   r   s    rV   get_input_embeddings$ModernBertModel.get_input_embeddings  s    ---rX   c                 $    XR                   l        g r   r  )rS   r  s     rV   set_input_embeddings$ModernBertModel.set_input_embeddings  s    ).&rX   r   rs   r  rt   r   r   r   r   
batch_sizeseq_lenr  output_hidden_statesreturn_dictrv   .c                 6  ^^	^
 Ub  UOU R                   R                  nUb  UOU R                   R                  nUb  UOU R                   R                  nUSL USL-  (       a  [	        S5      eU(       a  SOSnU(       a  SOSnU R                  5         Ub  U R                  X5        T	c+  T
c(  Ub  UR                  SS u  m	m
OUR                  SS u  m	m
Ub  UR                  OUR                  nUc&  [        R                  " T	T
4U[        R                  S9nSnU R                   R                  S:X  aH  TcD  UcA  Uc>  SnUc,  [        R                  " 5          [        XS	9tnmpxnSSS5        OF[        XRS	9tnmpxnO8Uc$  [        R                  " T
US
9R!                  S5      nU R#                  X+S9u  p#U R%                  XS9nU R&                   HD  nU(       a  UU4-   nU" UUUUUUUS9nUS   nU(       d  M*  [)        U5      S:  d  M;  UUS   4-   nMF     U(       a  UU4-   nU R+                  U5      nU(       a&  [-        UTT	T
S9nUb  [/        U	UU
4S jU 5       5      nU(       d  [/        S UX4 5       5      $ [1        UUUS9$ ! , (       d  f       N= f)F  
sliding_window_mask (`torch.Tensor` of shape `(batch_size, sequence_length)`, *optional*):
    Mask to avoid performing attention on padding or far-away tokens. In ModernBert, only every few layers
    perform global attention, while the rest perform local attention. This mask is used to avoid attending to
    far-away tokens in the local attention layers when not using Flash Attention.
indices (`torch.Tensor` of shape `(total_unpadded_tokens,)`, *optional*):
    Indices of the non-padding tokens in the input sequence. Used for unpadding the output.
cu_seqlens (`torch.Tensor` of shape `(batch + 1,)`, *optional*):
    Cumulative sequence lengths of the input sequences. Used to index the unpadded tensors.
max_seqlen (`int`, *optional*):
    Maximum sequence length in the batch excluding padding tokens. Used to unpad input_ids and pad output tensors.
batch_size (`int`, *optional*):
    Batch size of the input sequences. Used to pad the output tensors.
seq_len (`int`, *optional*):
    Sequence length of the input sequences including padding tokens. Used to pad the output tensors.
Nz:You must specify exactly one of input_ids or inputs_embedsr3   r   r   Fr+  T)rr   rs   )r   r   )r  )r   r   rQ  r}   rr   r   r   r   c              3   :   >#    U H  n[        UTTTS 9v   M     g7f)r  N)r   ).0hsr  r   r  s     rV   	<genexpr>*ModernBertModel.forward.<locals>.<genexpr>  s$      */ +"gZ`gh/s   c              3   ,   #    U H  oc  M  Uv   M     g 7fr   r3   )r  vs     rV   r  r     s     m$[q$[s   	)last_hidden_stater   
attentions)r   r  r  use_return_dictrR   r  %warn_if_padding_and_no_attention_maskr   r   r   onesrC  r5  no_gradr   arange	unsqueeze_update_attention_maskr  r  r  r  r   r   r   )rS   r   rs   r  rt   r   r   r   r   r  r  r  r  r  all_hidden_statesall_self_attentionsr   repadr   r   encoder_layerlayer_outputss         `  ``           rV   r   ModernBertModel.forward  s   B 2C1N-TXT_T_TqTq$8$D $++JjJj 	 &1%<k$++B]B]-t";<YZZ"6BD$5b4! 66yQ'/(&3&9&9"1&=#
G&/oobq&9#
G%.%:!!@T@T!"ZZW(=fTYT^T^_N;;++/BB:#5*:L (I`#,JF	7JQ )
 Ja,JFM7JQ #$||GFCMMaP262M2M 3N 3/N )Y![[M#$58H$H!)-$7)%%"3M *!,M  S%7!%;&9]1=M<O&O# )"   1]4D D62$gZPWM !,$) */* %!
 m]4E$[mmm++*
 	
i )s   J


Jc                 ,   U(       a  U R                   R                  S:X  a'  [        R                  S5        SU R                   l        OGU R                   R                  S:w  a-  [        R                  SU R                   R                   S35        [	        XR
                  5      n[        R                  " UR                  S   5      R                  S5      n[        R                  " XDR                  -
  5      nXPR                   R                  S-  :*  R                  S5      R                  S5      R                  UR                  5      nUR                  UR!                  5       [        R"                  " U R
                  5      R$                  5      nX74$ )Nr-  zOutputting attentions is only supported with the 'eager' attention implementation, not with "sdpa". Falling back to `attn_implementation="eager"`.r,  zZOutputting attentions is only supported with the eager attention implementation, not with zT. Consider setting `attn_implementation="eager"`. Setting `output_attentions=False`.r   r   )r   r5  r  r  r   r{   r   r  r   r  absTrD   r  r   masked_filllogical_notfinfomin)rS   rs   r  global_attention_maskrowsdistancewindow_maskr  s           rV   r  &ModernBertModel._update_attention_mask  sJ   {{//69##V 4;011W<##  $ @ @A B:: !;>:: V ||177:;EEaH99TFF]+ 4499DDQGQQRSTWWXfXmXmn 	 4??@W@W@Y[`[f[fgkgqgq[r[v[vw$99rX   )r   r  r  r  r  NNNNNNNNNNNNN)rg   rh   ri   rj   r"   r5   r  r  r   r   r   r   r   r   rC  r   r   r   r   r  ro   rp   rq   s   @rV   r  r  y  s   	/ 	./  15156:3704*.-1$($(!%,0/3&*x
E,,-x
 !.x
 &ell3	x

 u//0x
  -x
 %,,'x
 U\\*x
 SMx
 SMx
 #x
 $D>x
 'tnx
 d^x
 
uU\\3&'8	9x
 x
t:U\\ :VZ :_d_k_k : :rX   r  c                   j   ^  \ rS rSrS\4U 4S jjrS\R                  S\R                  4S jrSr	U =r
$ )rk  i'  r   c                 F  > [         TU ]  5         Xl        [        R                  " UR
                  UR
                  UR                  5      U l        [        UR                     U l
        [        R                  " UR
                  UR                  UR                  S9U l        g )Nr   )r4   r5   r   r   r   r8   rK   rl  r   rL   r   r   r>   r?   r   r   s     rV   r5   !ModernBertPredictionHead.__init__(  so    YYv1163E3EvG]G]^
&667LL!3!3vO_O_`	rX   r   rv   c                 `    U R                  U R                  U R                  U5      5      5      $ r   )r   r   rl  rM  s     rV   r    ModernBertPredictionHead.forward/  s#    yy$**]";<==rX   )r   r   rl  r   )rg   rh   ri   rj   r"   r5   r   r   r   ro   rp   rq   s   @rV   rk  rk  '  s2    a/ a>U\\ >ell > >rX   rk  zd
    The ModernBert Model with a decoder head on top that is used for masked language modeling.
    )custom_introc            "       F  ^  \ rS rSrS/rS\4U 4S jjrS rS\R                  4S jr
\R                  " SS	9S
\R                  S\R                  4S j5       r\              SS\\R"                     S\\R                     S\\R                     S\\R                     S\\R                     S\\R                     S\\R                     S\\R                     S\\   S\\   S\\   S\\   S\\   S\\   S\\\R                     \4   4S jj5       rSrU =r$ )rm  i3  zdecoder.weightr   c                 n  > [         TU ]  U5        Xl        [        U5      U l        [        U5      U l        [        R                  " UR                  UR                  UR                  S9U l        U R                  R                  U l        U R                  R                  U l        U R                  5         g )Nr   )r4   r5   r   r  rV  rk  headr   r   r8   r6   rI   rn  rN   rO   r  r   s     rV   r5   ModernBertForMaskedLM.__init__;  s     $V,
,V4	yy!3!3V5F5FVM`M`a!%!>!>(,(L(L% 	rX   c                     U R                   $ r   rn  r   s    rV   get_output_embeddings+ModernBertForMaskedLM.get_output_embeddingsH  s    ||rX   new_embeddingsc                     Xl         g r   r  )rS   r  s     rV   set_output_embeddings+ModernBertForMaskedLM.set_output_embeddingsK  s    %rX   Tr   r\   rv   c                 B    U R                  U R                  U5      5      $ r   )rn  r  )rS   r\   s     rV   compiled_head#ModernBertForMaskedLM.compiled_headN  s    ||DIIf-..rX   r   rs   r  rt   r   ru   r   r   r   r  r  r  r  r  c                 l   Ub  UOU R                   R                  nU R                  5         U R                   R                  S:X  a  Uc  Uc  U	c  U
c)  Uc&  Ub  UR                  SS u  pOUR                  SS u  pUb  UR
                  OUR
                  nUc%  [        R                  " X4U[        R                  S9nUc-  [        R                  " 5          [        XXFS9u  pppFSSS5        O[        XRXFS9u  pWppFU R                  UUUUUUUU	U
UUUUS9nUS   nU R                  (       aK  UbH  UR                  S5      nUR                  UR                  S   S5      nX`R                  :g  nUU   nUU   nU R                   R                  (       a  U R!                  U5      OU R#                  U R%                  U5      5      nSnUb*  U R&                  " UU4S	U R                   R(                  0UD6nU R                   R                  S:X  aQ  U R                   R*                  (       d  Uc
  [-        5       O[        R                  " 5          [/        UXzUS
9nSSS5        U(       d  U4nUb  U4U-   $ U$ [1        UUUR2                  UR4                  S9$ ! , (       d  f       GN= f! , (       d  f       NU= f)r  Nr+  r   r   )rr   rs   rt   ru   r   rs   r  rt   r   r   r   r   r  r  r  r  r  r   rx   r6   r  losslogitsr   r  )r   r  r  r5  r   r   r   r  rC  r  r   rV  rN   r   rO   rP   r  rn  r  loss_functionr6   rQ   r   r   r   r   r  )rS   r   rs   r  rt   r   ru   r   r   r   r  r  r  r  r  rT   r   outputsr  mask_tokensr  r  r\   s                          rV   r   ModernBertForMaskedLM.forwardR  s   F &1%<k$++B]B]!;;++/BB:#5*:L%'/$0.;.A.A"1.E+
G.7oobq.A+
-6-B))H\H\!)%*ZZ0Ef\a\f\f%gN ([r#,Zf\X	JL )
 \s,Zf\XMJL **) 3%'!!!/!5#  
 $AJ!!f&8[[_F 1 6 6v||A K !$A$AAK 1+ >K(F {{,, 01dii(9:; 	 %%ffbAWAWb[abD;;++/BB"&++"D"D\a\i\i\kk/vwipq l YF)-)9TGf$EvE!//))	
 	
m )^ lks   J
J%
J"%
J3)r   rn  r  rV  rO   rN   NNNNNNNNNNNNNN)rg   rh   ri   rj   _tied_weights_keysr"   r5   r  r   r   r  r   r   r   r  r   r   r   r   rC  r   r   r   r   ro   rp   rq   s   @rV   rm  rm  3  s    ++/ &BII & ]]4 /ELL /U\\ / !/  15156:/304)-*.-1$($(!%,0/3&*m
E,,-m
 !.m
 &ell3	m

 u||,m
  -m
 &m
 %,,'m
 U\\*m
 SMm
 SMm
 #m
 $D>m
 'tnm
 d^m
" 
uU\\"N2	3#m
 m
rX   rm  z`
    The ModernBert Model with a sequence classification head on top that performs pooling.
    c            "         ^  \ rS rSrS\4U 4S jjr\              SS\\R                     S\\R                     S\\R                     S\\R                     S\\R                     S	\\R                     S
\\R                     S\\R                     S\\   S\\   S\\   S\\   S\\   S\\   S\\\R                     \4   4S jj5       rSrU =r$ )ro  i  r   c                 n  > [         TU ]  U5        UR                  U l        Xl        [	        U5      U l        [        U5      U l        [        R                  R                  UR                  5      U l        [        R                  " UR                  UR                  5      U l        U R!                  5         g r   )r4   r5   
num_labelsr   r  rV  rk  r  r   r   r   rJ   r   r   r8   rs  r  r   s     rV   r5   ,ModernBertForSequenceClassification.__init__  s      ++$V,
,V4	HH$$V%>%>?	))F$6$68I8IJ 	rX   r   rs   r  rt   r   ru   r   r   r   r  r  r  r  r  rv   c                 t   Ub  UOU R                   R                  nU R                  5         U R                  UUUUUUUU	U
UUUUS9nUS   nU R                   R                  S:X  a
  USS2S4   nOLU R                   R                  S:X  a2  UUR                  S5      -  R                  SS9UR                  SS	S
9-  nU R                  U5      nU R                  U5      nU R                  U5      nSnUGb  U R                   R                  c  U R                  S:X  a  SU R                   l
        OoU R                  S:  aN  UR                  [        R                  :X  d  UR                  [        R                  :X  a  SU R                   l
        OSU R                   l
        U R                   R                  S:X  aJ  [!        5       nU R                  S:X  a&  U" UR#                  5       UR#                  5       5      nOU" UU5      nOU R                   R                  S:X  a=  [%        5       nU" UR'                  SU R                  5      UR'                  S5      5      nO-U R                   R                  S:X  a  [)        5       nU" UU5      nU(       d  U4nUb  U4U-   $ U$ [+        UUUR,                  UR.                  S9$ )a  
sliding_window_mask (`torch.Tensor` of shape `(batch_size, sequence_length)`, *optional*):
    Mask to avoid performing attention on padding or far-away tokens. In ModernBert, only every few layers
    perform global attention, while the rest perform local attention. This mask is used to avoid attending to
    far-away tokens in the local attention layers when not using Flash Attention.
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).
indices (`torch.Tensor` of shape `(total_unpadded_tokens,)`, *optional*):
    Indices of the non-padding tokens in the input sequence. Used for unpadding the output.
cu_seqlens (`torch.Tensor` of shape `(batch + 1,)`, *optional*):
    Cumulative sequence lengths of the input sequences. Used to index the unpadded tensors.
max_seqlen (`int`, *optional*):
    Maximum sequence length in the batch excluding padding tokens. Used to unpad input_ids and pad output tensors.
batch_size (`int`, *optional*):
    Batch size of the input sequences. Used to pad the output tensors.
seq_len (`int`, *optional*):
    Sequence length of the input sequences including padding tokens. Used to pad the output tensors.
Nr  r   r*   r+   rx   r}   r   Trz   keepdim
regressionsingle_label_classificationmulti_label_classificationr  )r   r  r  rV  r(   r  r~   r  r   rs  problem_typer  r{   r   longr   r
   squeezer	   r   r   r   r   r  )rS   r   rs   r  rt   r   ru   r   r   r   r  r  r  r  r  rT   r  r  pooled_outputr  r  loss_fctr\   s                          rV   r   +ModernBertForSequenceClassification.forward  s   N &1%<k$++B]B]!**) 3%'!!!/!5#  
 $AJ;;))U2 1!Q$ 7[[++v5!2^5M5Mb5Q!Q V V[\ V ]`n`r`rt as a ! 		"34		-0/{{''/??a'/;DKK,__q(fllejj.HFLL\a\e\eLe/LDKK,/KDKK,{{''<7"9??a'#FNN$4fnn6FGD#FF3D))-JJ+-B @&++b/R))-II,./YF)-)9TGf$EvE'!//))	
 	
rX   )rs  r   r   r  rV  r  r  )rg   rh   ri   rj   r"   r5   r   r   r   r   r   r   rC  r   r   r   r   ro   rp   rq   s   @rV   ro  ro    sk   /   15156:/304)-*.-1$($(!%,0/3&*e
E,,-e
 !.e
 &ell3	e

 u||,e
  -e
 &e
 %,,'e
 U\\*e
 SMe
 SMe
 #e
 $D>e
 'tne
 d^e
" 
uU\\"$<<	=#e
 e
rX   ro  zv
    The ModernBert Model with a token classification head on top, e.g. for Named Entity Recognition (NER) tasks.
    c            "         ^  \ rS rSrS\4U 4S jjr\              SS\\R                     S\\R                     S\\R                     S\\R                     S\\R                     S	\\R                     S
\\R                     S\\R                     S\\   S\\   S\\   S\\   S\\   S\\   S\\\R                     \4   4S jj5       rSrU =r$ )rq  i?  r   c                 b  > [         TU ]  U5        UR                  U l        [        U5      U l        [        U5      U l        [        R                  R                  UR                  5      U l        [        R                  " UR                  UR                  5      U l        U R                  5         g r   r4   r5   r  r  rV  rk  r  r   r   r   rJ   r   r   r8   rs  r  r   s     rV   r5   )ModernBertForTokenClassification.__init__E  s{      ++$V,
,V4	HH$$V%>%>?	))F$6$68I8IJ 	rX   r   rs   r  rt   r   ru   r   r   r   r  r  r  r  r  rv   c                    Ub  UOU R                   R                  nU R                  5         U R                  UUUUUUUU	U
UUUUS9nUS   nU R	                  U5      nU R                  U5      nU R                  U5      nSnUb<  [        5       nU" UR                  SU R                  5      UR                  S5      5      nU(       d  U4USS -   nUb  U4U-   $ U$ [        UUUR                  UR                  S9$ )a  
sliding_window_mask (`torch.Tensor` of shape `(batch_size, sequence_length)`, *optional*):
    Mask to avoid performing attention on padding or far-away tokens. In ModernBert, only every few layers
    perform global attention, while the rest perform local attention. This mask is used to avoid attending to
    far-away tokens in the local attention layers when not using Flash Attention.
labels (`torch.LongTensor` of shape `(batch_size, sequence_length)`, *optional*):
    Labels for computing the token classification loss. Indices should be in `[0, ..., config.num_labels - 1]`.
indices (`torch.Tensor` of shape `(total_unpadded_tokens,)`, *optional*):
    Indices of the non-padding tokens in the input sequence. Used for unpadding the output.
cu_seqlens (`torch.Tensor` of shape `(batch + 1,)`, *optional*):
    Cumulative sequence lengths of the input sequences. Used to index the unpadded tensors.
max_seqlen (`int`, *optional*):
    Maximum sequence length in the batch excluding padding tokens. Used to unpad input_ids and pad output tensors.
batch_size (`int`, *optional*):
    Batch size of the input sequences. Used to pad the output tensors.
seq_len (`int`, *optional*):
    Sequence length of the input sequences including padding tokens. Used to pad the output tensors.
Nr  r   rx   r}   r  )r   r  r  rV  r  r   rs  r	   r   r  r   r   r  )rS   r   rs   r  rt   r   ru   r   r   r   r  r  r  r  r  r  r  r  r  r  r\   s                        rV   r   (ModernBertForTokenClassification.forwardQ  s#   H &1%<k$++B]B]!**) 3%'!!!/!5#  
 $AJ II&78 II&78!23')HFKKDOO<fkk"oNDY,F)-)9TGf$EvE$!//))	
 	
rX   rs  r   r  rV  r  r  )rg   rh   ri   rj   r"   r5   r   r   r   r   r   r   rC  r   r   r   r   ro   rp   rq   s   @rV   rq  rq  ?  sk   
/ 
  15156:/304)-*.-1$($(!%,0/3&*I
E,,-I
 !.I
 &ell3	I

 u||,I
  -I
 &I
 %,,'I
 U\\*I
 SMI
 SMI
 #I
 $D>I
 'tnI
 d^I
  
uU\\"$99	:!I
 I
rX   rq  c            "         ^  \ rS rSrS\4U 4S jjr\             SS\\R                     S\\R                     S\\R                     S\\R                     S\\R                     S	\\R                     S
\\R                     S\\R                     S\\
   S\\
   S\\
   S\\   S\\   S\\   S\\\R                     \4   4S jj5       rSrU =r$ )rr  i  r   c                 b  > [         TU ]  U5        UR                  U l        [        U5      U l        [        U5      U l        [        R                  R                  UR                  5      U l        [        R                  " UR                  UR                  5      U l        U R                  5         g r   r  r   s     rV   r5   'ModernBertForQuestionAnswering.__init__  sy      ++$V,
,V4	HH$$V%>%>?	))F$6$68I8IJrX   r   rs   r  rt   start_positionsend_positionsr   r   r   r  r  r  r  r  rv   c                 R   Ub  UOU R                   R                  nU R                  5         U R                  UUUUUUU	U
UUUUS9nUS   nU R	                  U5      nU R                  U5      nU R                  U5      nUR                  SSS9u  nnUR                  S5      R                  5       nUR                  S5      R                  5       nSnUb  Ub  U R                  " UUXV40 UD6nU(       d  UU4USS -   nUb  U4U-   $ U$ [        UUUUR                  UR                  S9$ )r  N)rs   r  rt   r   r   r   r  r  r  r  r  r   r}   rx   r   )r  start_logits
end_logitsr   r  )r   r  r  rV  r  r   rs  splitr   r   r  r   r   r  )rS   r   rs   r  rt   r  r  r   r   r   r  r  r  r  r  rT   r  r  r  r  r  r  r\   s                          rV   r   &ModernBertForQuestionAnswering.forward  sc   F &1%<k$++B]B]!**) 3%!!!/!5#  
 $AJ II&78 II&78!23#)<<r<#: j#++B/::<''+668
&=+D%%lJibhiD"J/'!"+=F)-)9TGf$EvE+%!!//))
 	
rX   r
  r  )rg   rh   ri   rj   r"   r5   r   r   r   r   r   rC  r   r   r   r   ro   rp   rq   s   @rV   rr  rr    sf   	/ 	  266:/32604*.-1$($(!%,0/3&*K
ELL)K
 !.K
 &ell3	K

 u||,K
 "%,,/K
  -K
 %,,'K
 U\\*K
 SMK
 SMK
 #K
 $D>K
 'tnK
 d^K
" 
uU\\"$@@	A#K
 K
rX   rr  z
    The ModernBert Model with a multiple choice classification head on top (a linear layer on top of the pooled output and a softmax) e.g. for RocStories/SWAG tasks.
    c            "         ^  \ rS rSrS\4U 4S jjr\              SS\\R                     S\\R                     S\\R                     S\\R                     S\\R                     S	\\R                     S
\\R                     S\\R                     S\\   S\\   S\\   S\\   S\\   S\\   S\\\R                     \4   4S jj5       rSrU =r$ )rp  i  r   c                 8  > [         TU ]  U5        Xl        [        U5      U l        [        U5      U l        [        R                  R                  UR                  5      U l        [        R                  " UR                  S5      U l        U R                  5         g )Nr}   )r4   r5   r   r  rV  rk  r  r   r   r   rJ   r   r   r8   rs  r  r   s     rV   r5   $ModernBertForMultipleChoice.__init__   sm     $V,
,V4	HH$$V%>%>?	))F$6$6: 	rX   r   rs   r  rt   r   ru   r   r   r   r  r  r  r  r  rv   c                 P   Ub  UOU R                   R                  nUb  UR                  S   OUR                  S   nUb!  UR                  SUR	                  S5      5      OSnUb!  UR                  SUR	                  S5      5      OSnUb!  UR                  SUR	                  S5      5      OSnUb1  UR                  SUR	                  S5      UR	                  S5      5      OSnU R                  5         U R                  UUUUUUUU	U
UUUUS9nUS   nU R                   R                  S:X  a
  USS2S4   nOLU R                   R                  S:X  a2  UUR                  S5      -  R                  SS	9UR                  SS
S9-  nU R                  U5      nU R                  U5      nU R                  U5      nUR                  SU5      nSnUb  [        R                  " 5       nU" UU5      nU(       d  U4USS -   nUb  U4U-   $ U$ [        UUUR                   UR"                  S9$ )aK  
sliding_window_mask (`torch.Tensor` of shape `(batch_size, sequence_length)`, *optional*):
    Mask to avoid performing attention on padding or far-away tokens. In ModernBert, only every few layers
    perform global attention, while the rest perform local attention. This mask is used to avoid attending to
    far-away tokens in the local attention layers when not using Flash Attention.
labels (`torch.LongTensor` of shape `(batch_size,)`, *optional*):
    Labels for computing the multiple choice classification loss. Indices should be in `[0, ...,
    num_choices-1]` where `num_choices` is the size of the second dimension of the input tensors.
indices (`torch.Tensor` of shape `(total_unpadded_tokens,)`, *optional*):
    Indices of the non-padding tokens in the input sequence. Used for unpadding the output.
cu_seqlens (`torch.Tensor` of shape `(batch + 1,)`, *optional*):
    Cumulative sequence lengths of the input sequences. Used to index the unpadded tensors.
max_seqlen (`int`, *optional*):
    Maximum sequence length in the batch excluding padding tokens. Used to unpad input_ids and pad output tensors.
batch_size (`int`, *optional*):
    Batch size of the input sequences. Used to pad the output tensors.
seq_len (`int`, *optional*):
    Sequence length of the input sequences including padding tokens. Used to pad the output tensors.
Nr}   rx   r  r   r*   r+   r   Tr  r  )r   r  r   r   sizer  rV  r(   r  r~   r  r   rs  r   r	   r   r   r  )rS   r   rs   r  rt   r   ru   r   r   r   r  r  r  r  r  rT   num_choicesr  r  r  r  reshaped_logitsr  r  r\   s                            rV   r   #ModernBertForMultipleChoice.forward  sa   L &1%<k$++B]B],5,Aiooa(}GZGZ[\G]>G>SINN2y~~b'9:Y]	M[Mg,,R1D1DR1HImqGSG_|((\->->r-BCei ( r=#5#5b#9=;M;Mb;QR 	 	!**) 3%'!!!/!5#  
 $AJ;;))U2 1!Q$ 7[[++v5!2^5M5Mb5Q!Q V V[\ V ]`n`r`rt as a ! 		"34		-0/ ++b+6**,HOV4D%''!"+5F)-)9TGf$EvE("!//))	
 	
rX   )rs  r   r   r  rV  r  )rg   rh   ri   rj   r"   r5   r   r   r   r   r   r   rC  r   r   r   r   ro   rp   rq   s   @rV   rp  rp    sk   
/ 
  15156:/304)-*.-1$($(!%,0/3&*_
E,,-_
 !._
 &ell3	_

 u||,_
  -_
 &_
 %,,'_
 U\\*_
 SM_
 SM_
 #_
 $D>_
 'tn_
 d^_
" 
uU\\"$==	>#_
 _
rX   rp  )r"   r  rU  rm  ro  rq  rr  rp  r   rB  )Xr6  ri  
contextlibr   typingr   r   r   r   torch.nn.functionalr   r   r(  torch.utils.checkpointtorch.nnr   r	   r
   activationsr   configuration_utilsr   modeling_attn_mask_utilsr   modeling_layersr   modeling_outputsr   r   r   r   r   r   modeling_utilsr   utilsr   r   r   utils.import_utilsr   gemma.modeling_gemmar   r   flash_attn.flash_attn_interfacer   flash_attn.layers.rotaryr   flash_attn.ops.triton.rotaryr    object
get_loggerrg   r  r"   r   r   r   r   r   autogradFunctionr   r   r   rh  r   r   r   r   rC  r  r!  r{   r%  r*  r?  r   rE  rU  r  rk  rm  ro  rq  rr  rp  __all__r3   rX   rV   <module>r4     s      " + +      A A ! 3 B 9  . G G 5 M P89O 
		H	%B' BP ,0%)	&mLL&mLL&m 5<<(&m U\\"	&m
 5<<u||S(5<<:PRZ[`[g[gRhhi&mRLL\\  	
 \\>46%..11 46v *. $L &	L
 L42Q 2Qj299 <:BII :(	 4 	 )."!"	" LL" 	"
 5++," 38_" 	" 
"  ~" 5u||+,eELL.AAB"\ !&(!!(!	(! 2(! 	(!
 (! 38_(! 	(! 
(! ++(! 5<<(!V ! 	  LL  	 
 5++,  38_  	  
  5<< H 1$"! L3")) L3^+37 +3\ } } }@ j:/ j: j:Z	>ryy 	> 
H
5 H

H
V 
t
*C t

t
n 
W
'@ W

W
t X
%> X
 X
v 
m
"; m

m
`	rX   