tj1z UddlmZddlZddlmcmZddlmZddl m Z ddl m Z m Z ddlmZmZddlmZmZmZerdd lmZdd lmZmZnejZed d Zedd Ze ZerddlmZe j e!Z"dedefdZ#dedefdZ$de%de%de%de%de&f dZ'de%de%de%de%de&f dZ(de%defdZ)de%dejdefdZ*dejdefd Z+de%defd!Z,de%defd"Z-de%defd#Z.de%dejdefd$Z/d%ejdefd&Z0d'ejdefd(Z1d)ed*e%d+e%defd,Z2d-ejdzd.e%d+e%dejdzfd/Z3dejd-ejdzd.e%d+e%dejf d0Z4d1ej5dej5fd2Z6 dcd%ejdzd3e%d.e%d*e%d+e%d4e%dzde&fd5Z7d%ejdzd.e%d4e%dzde&fd6Z8 dcd%ejdzd.e%d4e%dzde&fd7Z9d)edefd8Z:d9ejd:ejd;ejdd d>d?f d@e%d3e%d.e%d*e%d+e%d)ed-ejdzdAe%dzdBe&dCe&dDe&dEe&dFej<e=zdejdzfdGZ>dde'dej?d>d>d?fd@e%d3e%d.e%d*e%d+e%d)ed-ejdzdHej@dCe&dEe&dFej<e=zdejfdIZAdde'dfd@e%d3e%d.e%d*e%d+e%d)ed-ejdzfdJZBdde'dd?fd@e%d3e%d.e%d*e%d+e%d)ed-ejdzdFej<e=zdefdKZCGdLdMeZDeDZEeDeFdN<dOejdejdzfdPZG dcdQe dRejd-ejezdzdSedzdOejdzdTe%dzdUejdzdeHe&ejezdze%e%ffdVZI dddQe dRejd-ejdzdSedzdOejdzdWedzdXedzdejdzdTe%dzdejezdzfdYZJ dedQe dRejd-ejdzdUejdzdSedzdWedzdXedzdejezdzfdZZK dddQe dRejd-ejdzdSedzdOejdzdWedzdXedzdejdzdTe%dzdejezdzfd[ZL dedQe dRejd-ejdzdUejdzdSedzdWedzdXedzdejezdzfd\ZM dedQe dRejd-ejdzdSedzdOejdzdWedzdXedzdTe%dzdejezdzfd]ZN dcdQe dRejd-ejdzdSedzdejdzf d^ZOeJeLeNeLeLeJeJeOeOeJeOd_eJeOd`da ZP dedQe dRejd-ejdzdSedzdOejdzdWedzdXedzdejdzfdbZQdS)f)CallableN)Cache)PreTrainedConfig)is_torch_xpu_availablelogging)GeneralInterfaceis_flash_attention_requested)is_torch_flex_attn_availableis_torch_greater_or_equal is_tracing)_DEFAULT_SPARSE_BLOCK_SIZE) BlockMaskcreate_block_maskz2.5T) accept_devz2.6)TransformGetItemToIndexmask_functionsreturnchtdDstdfd}|S)zKReturns a mask function that is the intersection of provided mask functionsc34K|]}t|VdSNcallable.0args X/var/www/html/CA-Chatbot/venv/lib/python3.11/site-packages/transformers/masking_utils.py zand_masks..2(77x}}777777.All inputs should be callable mask_functions: c|dtj}D]+}|||||||jz},|SN)dtype)new_onestorchbooltodevice batch_idxhead_idxq_idxkv_idxresultmaskrs rand_maskzand_masks..and_mask5s\%*55" Y YDdd9hvFFII&-XXXFF r all RuntimeError)rr2s` r and_masksr60sZ 77777 7 7^\N\\]]] Or chtdDstdfd}|S)zDReturns a mask function that is the union of provided mask functionsc34K|]}t|VdSrrrs rrzor_masks..@rr r!c|dtj}D]+}|||||||jz},|Sr#) new_zerosr'r(r)r*r+s ror_maskzor_masks..or_maskCs\5:66" Y YDdd9hvFFII&-XXXFF r r3)rr;s` ror_masksr<>sZ 77777 7 7^\N\\]]] Nr r,r-r.r/c||kS)z: This creates a basic lower-diagonal causal mask. r$r,r-r.r/s rcausal_mask_functionr?Ls U?r c|dkS)z This creates a full bidirectional mask. NOTE: It is important to keep an index-based version for non-vmap expansion. rr$r>s rbidirectional_mask_functionrASs A:r sliding_windowc Zdtdtdtdtdtf fd }|S)z This is an overlay depicting a sliding window pattern. Add it on top of a causal mask for a proper sliding window mask. r,r-r.r/rc||z kSrr$r,r-r.r/rBs r inner_maskz*sliding_window_overlay..inner_maskbs...r intr(rBrFs` rsliding_window_overlayrJ\sL /c/S//c/d////// r chunk_size left_paddingc ^dtdtdtdtdtf fd }|S)z This is an overlay depicting a chunked attention pattern. Add it on top of a causal mask for a proper chunked attention mask. r,r-r.r/rc@||z z||z zkSrr$)r,r-r.r/rKrLs rrFz#chunked_overlay..inner_maskns.i00Z?ELYbLcDchrCrrrr rG)rKrLrFs`` rchunked_overlayrOhs^ scsSsscsdsssssss r block_sequence_idsc Zdtdtdtdtdtf fd }|S)a This is an overlay depicting a blockwise masking pattern. Instead of a single token, each block consists of arbitrary length tokens. In causal setup, each block can attend to prev block causally and can't attend to future blocks. Within one block the attention is always bidirectional. Mostly used in MLLMs when non-text data attends bidirectionally to itself. r,r-r.r/rcF||f}||f}||k|dkzS)Nrr$)r,r-r.r/q_groupkv_grouprPs rrFz%blockwise_overlay..inner_mask}s5$Y%56%i&788#1 55r rG)rPrFs` rblockwise_overlayrUtsL6c6S66c6d666666 r cFtt|tS)zQ This return the mask_function function to create a sliding window mask. )r6rJr?rBs r#sliding_window_causal_mask_functionrXs +N;;=Q R RRr c Zdtdtdtdtdtf fd }|S)zN This is an overlay depicting a bidirectional sliding window pattern. r,r-r.r/rc0t||z kS)zA token can attend to any other token if their absolute distance is within the (inclusive) sliding window size (distance <= sliding_window).)absrEs rrFz8sliding_window_bidirectional_overlay..inner_masks56>""n44r rGrIs` r$sliding_window_bidirectional_overlayr\sL 5c5S55c5d555555 r cFtt|tS)z_ This return the mask_function function to create a bidirectional sliding window mask. )r6r\rArWs r*sliding_window_bidirectional_mask_functionr^s 9.IIKf g ggr cHtt||tS)zT This return the mask_function function to create a chunked attention mask. )r6rOr?)rKrLs rchunked_causal_mask_functionr`s _Z>>@T U UUr padding_maskc Zdtdtdtdtdtf fd }|S)zT This return the mask_function function corresponding to a 2D padding mask. r,r-r.r/rc||fSrr$)r,r-r.r/ras rrFz)padding_mask_function..inner_masksIv-..r rG)rarFs` rpadding_mask_functionrdsL /c/S//c/d////// r packed_sequence_maskc Zdtdtdtdtdtf fd }|S)z\ This return the mask_function function corresponding to a 2D packed sequence mask. r,r-r.r/rc0||f||fkSrr$)r,r-r.r/res rrFz1packed_sequence_mask_function..inner_masks$#Iu$459MiY_N_9```r rG)rerFs` rpacked_sequence_mask_functionrhsW acaSaacadaaaaaa r mask_functionq_offset kv_offsetc bdtdtdtdtdtf fd }|S)z This function adds the correct offsets to the `q_idx` and `kv_idx` as the torch API can only accept lengths, not start and end indices. r,r-r.r/rc,|||z|zSrr$)r,r-r.r/rkrirjs rrFz0add_offsets_to_mask_function..inner_masks#}Y%(2BFYDVWWWr rG)rirjrkrFs``` radd_offsets_to_mask_functionrnse XcXSXXcXdXXXXXXXX r attention_mask kv_lengthc|}|@||z|jdz x}dkr'tjj|d|f}|S)zh From the 2D attention mask, prepare the correct padding mask to use by potentially padding it. Nr)shaper'nn functionalpad)rorprklocal_padding_maskpadding_lengths rprepare_padding_maskrysY(!')3n6J26NN NNRS S S!&!4!8!8!^I\!]!]  r cj||z|jdz x}dkrtj|d|fd}|S)z Pads the `block_sequence_ids` in case the total length is less than `kv_length`. Usually that happens with `StaticCache` generation or generating without cache. Pads to the right with `-1`. rrr)rvvalue)rsFrv)rProrprkrxs rmaybe_pad_block_sequence_idsr}sL$i/2D2J22NNNRSSSU#5A~;NVXYYY r tensorcV||kS)ziSimilar to `tensor.all()`, but uses an implementation with `tensor.sum()`, which is actually much faster.)sumnumel)r~s rfast_allrs ::<<6<<>> ))r q_lengthlocal_attention_sizecv|;|jd|kr*tj||j|z}|dd|f}t |rdS|||krdS|dks||kr|t |rdS|dkr;|7t |ddd|frt |dd|dfrdSdS)a Detects whether the causal mask can be ignored in case PyTorch's SDPA is used, rather relying on SDPA's `is_causal` argument. In case no token is masked in the 2D `padding_mask` argument, if `query_length == 1` or `key_value_length == query_length`, we rather rely on SDPA `is_causal` argument to use causal/non-causal masks, allowing to dispatch to the flash attention kernel (that can otherwise not be used if a custom `attn_mask` is passed). Nrrr*FrTr)rsr'aranger*r r)rarrprjrkr mask_indicess r_ignore_causal_mask_sdpars L$6r$:Y$F$F|Il6IJJJYV #AAA|O4 ,u'I9M,M,Mu A h..\5IXVbMcMc5It1}},qqq)8)|*D!E!E(T`abababdldmdmamTnSnJoJot 5r cht|rdS|||krdS|dS|S)z XPU-specific logic for determining if we can skip bidirectional mask creation. For XPU devices, we have special handling: - Skip if no padding and no local attention constraint FNT)r r4rarprs r _can_skip_bidirectional_mask_xpursP,u'I9M,M,Mut     r ctrt|||St|s ||r |||krdSdS)a Detects whether the bidirectional mask can be ignored in case PyTorch's SDPA is used. In case no token is masked in the 2D `padding_mask` argument and no local attention constraint applies (i.e. `local_attention_size` is None or `kv_length < local_attention_size`), we skip mask creation, allowing to dispatch to the flash attention kernel (that can otherwise not be used if a custom `attn_mask` is passed). NTF)_is_torch_xpu_availablerr r4rs r_ignore_bidirectional_mask_sdpar4sf_0 iI]^^^ | $ $  !\%5%5%7%7 ! ! )Y9M-M-Mt 5r cFgd}|D]}tj||d}|S)aJ Used to vmap our mask_functions over the all 4 dimensions (b_idx, h_idx, q_idx, kv_idx) of the inputs. Using vmap here allows us to keep the performance of vectorized ops, while having a single set of primitive functions between attention interfaces (i.e. between flex and sdpa/eager, FA2 being a bit different). ))NNNr)NNrN)NrNN)rNNNr)in_dimsout_dims)r'vmap)ri dimensionsdimss r_vmap_expansion_sdparSs?nmmJLL =$KKK r batch_indices head_indices q_indices kv_indicesc~|dddddf}|dddddf}|dddddf}|dddddf}||||fS)a Used to broadcast our mask_functions over the all 4 dimensions (b_idx, h_idx, q_idx, kv_idx) of the inputs. Allows the usage of any index-based mask function without relying on vmap. NOTE: This is limited to index based functions only and is not guaranteed to work otherwise. Reference: - https://github.com/huggingface/optimum-onnx/blob/c123e8f4fab61b54a8e0e31ce74462bcacca576e/optimum/exporters/onnx/model_patcher.py#L362-L365 Nr$)rrrrs r_non_vmap_expansion_sdpar`so"!!!T4"56Maaat 34L$aaa-.ID$aaa/0J , : ==r Fcpu batch_size local_sizeallow_is_causal_skipallow_is_bidirectional_skipallow_torch_fixuse_vmapr*c  t|||}|rt||||||rdS| rt|||rdS|t|t |}t j|| }t jd| }t j|| |z}t j|| |z}| s.|t||||}||d||}nXtrBt5t|||||}dddn #1swxYwYntdts| r|t j|ddz}|S)uy Create a 4D boolean mask of shape `(batch_size, 1, query_length, kv_length)` where a value of True indicates that the element should take part in the attention computation, and False that it should not. This function can only be used with torch>=2.5, as the context manager is otherwise not available. Args: batch_size (`int`): The batch size of the input sequence. q_length (`int`): The size that the query states will have during the attention computation. kv_length (`int`): The size that the key and value states will have during the attention computation. kv_offset (`int`, optional): An optional offset to indicate at which first position the key and values states will refer to. q_offset (`int`, optional): An optional offset to indicate at which first position the query states will refer to. mask_function (`Callable`): The mask factory function describing the mask pattern. attention_mask (`torch.Tensor`, optional): The 2D attention mask corresponding to padded tokens of shape (batch_size, number_of_seen_tokens+q_length) local_size (`int`, optional): The size of the local attention, if we do not use full attention. This is used only if `allow_is_causal_skip=True` to try to skip mask creation if possible. allow_is_causal_skip (`bool`, optional): Whether to allow to return `None` for the mask under conditions where we can use the `is_causal` argument in `torch.sdpa` instead. Default to `True`. allow_is_bidirectional_skip (`bool`, optional): Whether to allow to return `None` for the mask under conditions where we do not have to add any bias, i.e. full attention without any padding. Default to `False`. allow_torch_fix (`bool`, optional): Whether to update the mask in case a query is not attending to any tokens, to solve a bug in torch's older versions. We need an arg to skip it when using eager. By default `True`. use_vmap (`bool`, optional): Whether to use `vmap` during the mask construction or not. Allows powerful custom patterns that may not be index-based (for the cost of speed performance). By default `False`. device (`torch.device` or `str`, optional): An optional device to create the mask on. ## Creating a simple causal mask: To create the following causal mask: 0 ■ ⬚ ⬚ ⬚ ⬚ 1 ■ ■ ⬚ ⬚ ⬚ 2 ■ ■ ■ ⬚ ⬚ 3 ■ ■ ■ ■ ⬚ 4 ■ ■ ■ ■ ■ You can do ```python >>> sdpa_mask(batch_size=1, q_length=5, kv_length=5) >>> tensor([[[[ True, False, False, False, False], [ True, True, False, False, False], [ True, True, True, False, False], [ True, True, True, True, False], [ True, True, True, True, True]]]]) ``` ## Creating a sliding window mask: To create the following sliding window mask (`sliding_window=3`): 0 ■ ⬚ ⬚ ⬚ ⬚ 1 ■ ■ ⬚ ⬚ ⬚ 2 ■ ■ ■ ⬚ ⬚ 3 ⬚ ■ ■ ■ ⬚ 4 ⬚ ⬚ ■ ■ ■ You can do ```python >>> sdpa_mask(batch_size=1, q_length=5, kv_length=5, mask_function=sliding_window_causal_mask_function(3)) >>> tensor([[[[ True, False, False, False, False], [ True, True, False, False, False], [ True, True, True, False, False], [False, True, True, True, False], [False, False, True, True, True]]]]) ``` ## Creating a chunked attention mask To create the following chunked attention mask (`chunk_size=3`): 0 ■ ⬚ ⬚ ⬚ ⬚ 1 ■ ■ ⬚ ⬚ ⬚ 2 ■ ■ ■ ⬚ ⬚ 3 ⬚ ⬚ ⬚ ■ ⬚ 4 ⬚ ⬚ ⬚ ■ ■ You can do ```python >>> sdpa_mask(batch_size=1, q_length=5, kv_length=5, mask_function=chunked_causal_mask_function(3, torch.zeros(1, dtype=int))) >>> tensor([[[[ True, False, False, False, False], [ True, True, False, False, False], [ True, True, True, False, False], [False, False, False, True, False], [False, False, False, True, True]]]]) ``` NrrrrzThe vmap functionality for mask creation is only supported from torch>=2.6. Please update your torch version or use `use_vmap=False` with index-based masks.T)dimkeepdim)ryrrr6rdr'rrexpand#_is_torch_greater_or_equal_than_2_6rr ValueError#_is_torch_greater_or_equal_than_2_5r4)rrrprjrkrirorrrrrr*kwargsra batch_arange head_arangeq_arange kv_aranges r sdpa_maskrssp( 9MML  8h 8Y !!t"'F|U^`j'k'kt!-1F|1T1TUU < 6:::L,q000K|HV444x?H Yv666BI  &(@{\dfo(p(pq'..z2xSS -  % & & q q@1-@@{\dfoppN q q q q q q q q q q q q q q q  _    /[?['%)^OUY*Z*Z*ZZ sD))D-0D-r%c  &| dd} | dd} td|||||||d|d| | d | } | Itj|j}tj| tjd| j||} | S) a$ Create a 4D float mask of shape `(batch_size, 1, query_length, kv_length)` where a value of 0 indicates that the element should take part in the attention computation, and -inf (minimum value for the given `dtype`) that it should not. Args: batch_size (`int`): The batch size of the input sequence. q_length (`int`): The size that the query states will have during the attention computation. kv_length (`int`): The size that the key and value states will have during the attention computation. q_offset (`int`, optional): An optional offset to indicate at which first position the query states will refer to. kv_offset (`int`, optional): An optional offset to indicate at which first position the key and values states will refer to. mask_function (`Callable`): The mask factory function describing the mask pattern. attention_mask (`torch.Tensor`, optional): The 2D attention mask corresponding to padded tokens of shape (batch_size, number_of_seen_tokens+q_length) dtype (`torch.dtype`, optional): The dtype to use for the mask. By default, `torch.float32`. allow_is_bidirectional_skip (`bool`, optional): Whether to allow to return `None` for the mask under conditions where we do not have to add any bias, i.e. full attention without any padding. Default to `False`. use_vmap (`bool`, optional): Whether to use `vmap` during the mask construction or not. Allows powerful custom patterns that may not be index-based (for the cost of speed performance). By default `False`. device (`torch.device` or `str`, optional): An optional device to create the mask on. rNrF) rrrprjrkrirorrrrr*gr*r%r$)poprr'finfominwherer~r*)rrrprjrkriror%rrr*r_r1 min_dtypes r eager_maskrs\  )400A $d++A  #%"$?     D  K&&* {4c$+U!S!S!SU^__ Kr c v|6|dd| df}|jd|kr|rd}|S)a Create the attention mask necessary to use FA2. Since FA2 is un-padded by definition, here we simply return `None` if the mask is fully causal, or we return the 2D mask which will then be used to extract the seq_lens. We just slice it in case of sliding window. Args: batch_size (`int`): The batch size of the input sequence. q_length (`int`): The size that the query states will have during the attention computation. kv_length (`int`): The size that the key and value states will have during the attention computation. q_offset (`int`, optional): An optional offset to indicate at which first position the query states will refer to. kv_offset (`int`, optional): An optional offset to indicate at which first position the key and values states will refer to. mask_function (`Callable`): The mask factory function describing the mask pattern. attention_mask (`torch.Tensor`, optional): The 2D attention mask corresponding to padded tokens of shape (batch_size, number_of_seen_tokens+q_length) Nr)rsr4)rrrprjrkrirors rflash_attention_maskrfsS>!'I:;;7   "i / /N4F4F4H4H /!N r c |t|s|jdtzdztz} | |jdz } ts/| dkr)tjj|dd| f}t|||} t|t| }t|||}t||d|||t} | S)a Create a 4D block mask which is a compressed representation of the full 4D block causal mask. BlockMask is essential for performant computation of flex attention. See: https://pytorch.org/blog/flexattention/ Args: batch_size (`int`): The batch size of the input sequence. q_length (`int`): The size that the query states will have during the attention computation. kv_length (`int`): The size that the key and value states will have during the attention computation. q_offset (`int`, optional): An optional offset to indicate at which first position the query states will refer to. kv_offset (`int`, optional): An optional offset to indicate at which first position the key and values states will refer to. mask_function (`Callable`): The mask factory function describing the mask pattern. attention_mask (`torch.Tensor`, optional): The 2D attention mask corresponding to padded tokens of shape (batch_size, number_of_seen_tokens+q_length) device (`torch.device` or `str`, optional): An optional device to create the mask on. Nrr)r{rv)mask_modBHQ_LENKV_LENr*_compile) rrsflex_default_block_sizerr'rtrurvryr6rdrnr) rrrprjrkriror*rpad_lenra block_masks rflex_attention_maskrsD!(>*B*B!#(+/FF!KOffN0332 `w{{"X044^1STV]R^4__N+NIyQQ !-1F|1T1TUU 1)TTM#  4J r c eZdZeeeeeedZdS)AttentionMaskInterface)sdpaeagerflash_attention_2flash_attention_3flash_attention_4flex_attentionN)__name__ __module__ __qualname__rrrr_global_mappingr$r rrrs.111- OOOr rALL_MASK_ATTENTION_FUNCTIONS position_idsc|ddddfdz }tj||d}|dkd}t|s$|dddfdkrdS|S)a> Find the indices of the sequence to which each new query token in the sequence belongs when using packed tensor format (i.e. several sequences packed in the same batch dimension). Args: position_ids (`torch.Tensor`) A 2D tensor of shape (batch_size, query_length) indicating the positions of each token in the sequences. Returns: A 2D tensor where each similar integer indicates that the tokens belong to the same sequence. For example, if we pack 3 sequences of 2, 3 and 1 tokens respectively along a single batch dim, this will return [[0, 0, 1, 1, 1, 2]]. If the there is only one sequence in each batch item (and we don't compile), then we return `None` indicating no packed sequences. This is the same as [[0, 0, 0, 0, 0, 0]] for the example above. Nrrr)prependrr)r'diffcumsumr r4)rfirst_dummy_value position_diffres rfind_packed_sequence_indicesrs*%QQQU+a/J|5FBOOOM)Q.66r:: * + +1Eaaae1LPQ1Q0V0V0X0Xt r config inputs_embedspast_key_values layer_idxencoder_hidden_statesct|tjtfr!t |jdkr d|dddddfS|jtjvrdS|1|j dkr&| |j tj }|jd}|e| |}t|tjr| |j n|}|||\} } n'd}|| |jdn|} d} n|jd d} } d} |G|E|C|jd} | |jdkr|| d }t!|} d || || || fS) a Perform some common pre-processing of the mask arguments we get from the modeling code. Mostly determine the key-value length and offsets, and if we should early exit or not. Args: config (`PreTrainedConfig`): The model config. inputs_embeds (`torch.Tensor`): The input embeddings of shape (batch_size, query_length, hidden_dim). This is used only to infer the batch size, query length and dtype. attention_mask (`torch.Tensor`, optional): The 2D attention mask corresponding to padded tokens of shape (batch_size, number_of_seen_tokens+q_length). It can also be an already prepared 4D mask, in which case it is returned as-is. past_key_values (`Cache`, optional): The past key values, if we use a cache. position_ids (`torch.Tensor`, optional) A 2D tensor of shape (batch_size, query_length) indicating the positions of each token in the sequences. layer_idx (`int`, optional): If `past_key_values` is not None, this is the layer index of the cache from which to get the key-value length and offset. Indeed, for hybrid caches, different layers may return different lengths. encoder_hidden_states (`torch.Tensor`, optional): The input embeddings of shape (batch_size, kv_length, hidden_dim). If provided, it is used instead of `inputs_embeds` to infer the kv length. Returns: early_exit (`bool`): Whether we should early exit mask creation, and return the mask as-is. attention_mask (`torch.Tensor` or `BlockMask` or `None`): The attention mask to either return immediately, or to use in downstream mask creation. packed_sequence_mask (`torch.Tensor`, optional): In case we detected packed sequence format, this is a tensor where each similar integer indicates that the tokens belong to the same sequence. q_length (`int`): The size that the query states will have during the attention computation. kv_length (`int`): The size that the key and value states will have during the attention computation. q_offset (`int`, optional): An optional offset to indicate at which first position the query states will refer to. kv_offset (`int`): An offset to indicate at which first position the key and values states will refer to. TN)TNNNNNNrrrrrF) isinstancer'Tensorrlenrs_attn_implementationrrndimr)r*r(get_query_offsetget_mask_sizesrr) rrrorrrrrrjrprkrers r_preprocess_mask_argumentsrsf.5<";<<B^EYAZAZ^_A_A_^T4tTAA"*F*VVV77!n&9Q&>&>'**-2Fej*YY"1%H""33I>>9C8U\8Z8Zh8;;}3444`h.==h RR 99  !:O:[-3A66aiIII$2#7#;QyI N$:?V"(+ +A. . .'..z2>>L;LII ."6)XW` ``r or_mask_functionand_mask_functionc $t|ddst||||||S|6t|dr$d|jvr|jd}nd}t ||||||\} }} } } } }| r|S|jd|j|j}}}t}t|j }d}t|ddo| d k }|*tstd t||}d}d}|*tstd t||}d}d}| t|t!| }d}|1t#||| |}t|t%|}d}||| | | |||||||| }|S) a Create a standard causal mask based on the attention implementation used (stored in the config). If `past_key_values` has an hybrid cache structure, this function will return the mask corresponding to one of the "full_attention" layers (to align to what is needed in the `modeling_xxx.py` files). Args: config (`PreTrainedConfig`): The model config. inputs_embeds (`torch.Tensor`): The input embeddings of shape (batch_size, query_length, hidden_dim). This is used only to infer the batch size, query length and dtype. attention_mask (`torch.Tensor`, optional): The 2D attention mask corresponding to padded tokens of shape (batch_size, number_of_seen_tokens+q_length). It can also be an already prepared 4D mask, in which case it is returned as-is. cache_position (`torch.Tensor`): Deprecated and unused. past_key_values (`Cache`, optional): The past key values, if we use a cache. position_ids (`torch.Tensor`, optional) A 2D tensor of shape (batch_size, query_length) indicating the positions of each token in the sequences. or_mask_function (`Callable`, optional): An optional mask function to combine with the causal mask function (by doing the union of both). This is useful to easily overlay another mask on top of the causal one, for example for image tokens handling. and_mask_function (`Callable`, optional): An optional mask function to combine with the causal mask function (by doing the intersection of both). This is useful to easily overlay another mask on top of the causal one, for example for image tokens handling. block_sequence_ids (`torch.Tensor`, *optional*): A tensor of same shape as input IDs indicating to which block or group each token belongs to. Tokens from the same block will keep a bidirectional mask within the block, attending causally to the past. Index `-1` can be used for blocks that have to keep complete causality within itself. layer_idx (`int`, *optional*): The cache layer to size the mask against. By default, the first "full_attention" layer is used, which is correct whenever all layers of a given mask type have seen the same tokens. Pass it explicitly for caches where layers of the same type hold different lengths (e.g. per-depth MTP streams). is_causalTrrrN is_slidingFris_compileablerLUsing `or_mask_function` or `and_mask_function` arguments require torch>=2.6) rrrprjrkrirorr%rrr*)getattrcreate_bidirectional_maskhasattrrindexrrsr%r*r?rrrrr<r6rhr}rU)rrrorrrrrPr early_exitrerrprjrkrr%r*mask_factory_functionmask_interfacerr causal_masks rcreate_causal_maskrgsA` 6; - - (   +-/      ?L 1 1 e?Y6Y6Y'288??III #6=./[girss_J 4h 8U^ - 3A 6 8K]MavJ01&2MNN H!(9I5 Q Q cV^bcVcd #2 mkll l ()>@P Q Q$$2 mkll l )*?AR S S$' )*?A^_sAtAt u u$%9:Ln^girss ()>@QRd@e@e f f$!.+%1   K r c t|dr$d|jvr|jd}nd}t||||d||\} }} } } } }| r|S||n|}|jd|j}}|j}t}t|j }d}d}|*tstdt||}d}d}|*tstdt||}d}d}||| | | |||d||||| }|S)a; Create a standard bidirectional mask based on the attention implementation used (stored in the config). Args: config (`PreTrainedConfig`): The model config. inputs_embeds (`torch.Tensor`): The input embeddings of shape (batch_size, query_length, hidden_dim). This is only used to infer metadata such as the batch size, query length, dtype, and device. attention_mask (`torch.Tensor`, optional): The 2D attention mask corresponding to padded tokens of shape (batch_size, kv_length). It can also be an already prepared 4D mask of shape (batch_size, 1, query_length, kv_length), in which case it is returned as-is. encoder_hidden_states (`torch.Tensor`, optional): The input embeddings of shape (batch_size, kv_length, hidden_dim). If provided, it is used instead of `inputs_embeds` to infer the batch size, kv length and dtype. past_key_values (`Cache`, optional): The past key values, if we use a cache. or_mask_function (`Callable`, optional): An optional mask function to combine with the base mask function (by doing the union of both). This is useful to easily overlay another mask on top, for example for image tokens handling. and_mask_function (`Callable`, optional): An optional mask function to combine with the base mask function (by doing the intersection of both). This is useful to easily overlay another mask on top, for example for image tokens handling. rFrNTr) rrrprjrkrirorrr%rrr*)rrrrrsr%r*rArrrrr<r6)rrrorrrrrrrrrrprjrkembedsrr%r*rrrrs rrrsH --%?;U2U2U#.44U;;  Oi ~iQfOOKJ8Y)&;&G " "]F QJ !F71&2MNN#'H #2 mkll l ()>@P Q Q&+#$2 mkll l )*?AR S S&+#$^+%"$?N r c zt|ddst||||||S|6t|dr$d|jvr|jd}nd}t ||||||\} }} } } } }| r|St|dd}|t d|jd|j|j }}}t|}t|j }d }t|d d o| d k }|*tst d t||}d }d}|*tst d t||}d }d}| t|t!| }d }|1t#||| |}t|t%|}d }||| | | ||||||||| }|S)aB Create a sliding window causal mask based on the attention implementation used (stored in the config). This type of attention pattern was mostly democratized by Mistral. If `past_key_values` has an hybrid cache structure, this function will return the mask corresponding to one of the "sliding_attention" layers (to align to what is needed in the `modeling_xxx.py` files). Args: config (`PreTrainedConfig`): The model config. inputs_embeds (`torch.Tensor`): The input embeddings of shape (batch_size, query_length, hidden_dim). This is used only to infer the batch size, query length and dtype. attention_mask (`torch.Tensor`, optional): The 2D attention mask corresponding to padded tokens of shape (batch_size, number_of_seen_tokens+q_length). It can also be an already prepared 4D mask, in which case it is returned as-is. cache_position (`torch.Tensor`): Deprecated and unused. past_key_values (`Cache`, optional): The past key values, if we use a cache. position_ids (`torch.Tensor`, optional) A 2D tensor of shape (batch_size, query_length) indicating the positions of each token in the sequences. or_mask_function (`Callable`, optional): An optional mask function to combine with the sliding causal mask function (by doing the union of both). This is useful to easily overlay another mask on top of the sliding causal one, for example for image tokens handling. and_mask_function (`Callable`, optional): An optional mask function to combine with the sliding causal mask function (by doing the intersection of both). This is useful to easily overlay another mask on top of the sliding causal one, for example for image tokens handling. block_sequence_ids (`torch.Tensor`, *optional*): A tensor of same shape as input IDs indicating to which block or group each token belongs to. Tokens from the same block will keep a bidirectional mask within the block, attending causally to the past. Index `-1` can be used for blocks that have to keep complete causality within itself. layer_idx (`int`, *optional*): The cache layer to size the mask against. By default, the first "full_attention" layer is used, which is correct whenever all layers of a given mask type have seen the same tokens. Pass it explicitly for caches where layers of the same type hold different lengths (e.g. per-depth MTP streams). rTrNrrrBJCould not find a `sliding_window` argument in the config, or it is not setFrrr rrrprjrkrirorrr%rrr*)r(create_bidirectional_sliding_window_maskrrrrrrsr%r*rXrrrr<r6rhr}rU)rrrorrrrrPrrrerrprjrkrBrr%r*rrrrrs r!create_sliding_window_causal_maskrJssb 6; - - 7   +-/      ?L 1 1 do>X6X6X'288>>III #6=./[girss_J 4h 8U^V%5t<@P Q Q$$2 mkll l )*?AR S S$' )*?A^_sAtAt u u$%9:Ln^girss ()>@QRd@e@e f f$!.+%1!K r c ft|dr$d|jvr|jd}nd}t||||d||\} }} } } } }| r|St |dd}|t d||n|}|jd|j|j}}}t|}t|j }d}d}|*tst dt||}d}d}|*tst dt||}d}d}||| | | |||d|||||| }|S) aJ Create a standard bidirectional sliding window mask based on the attention implementation used (stored in the config). Args: config (`PreTrainedConfig`): The model config. inputs_embeds (`torch.Tensor`): The input embeddings of shape (batch_size, query_length, hidden_dim). This is only used to infer metadata such as the batch size, query length, dtype, and device. attention_mask (`torch.Tensor`, optional): The 2D attention mask corresponding to padded tokens of shape (batch_size, kv_length). It can also be an already prepared 4D mask of shape (batch_size, 1, query_length, kv_length), in which case it is returned as-is. encoder_hidden_states (`torch.Tensor`, optional): The input embeddings of shape (batch_size, kv_length, hidden_dim). If provided, it is used instead of `inputs_embeds` to infer the batch size, kv length and dtype. past_key_values (`Cache`, optional): The past key values, if we use a cache. or_mask_function (`Callable`, optional): An optional mask function to combine with the base mask function (by doing the union of both). This is useful to easily overlay another mask on top, for example for image tokens handling. and_mask_function (`Callable`, optional): An optional mask function to combine with the base mask function (by doing the intersection of both). This is useful to easily overlay another mask on top, for example for image tokens handling. rTrNrBrFr)rrrprjrkrirorrrr%rrr*)rrrrrrrsr%r*r^rrrr<r6)rrrorrrrrrrrrrprjrkrBrrr%r*rrrrs rrrsH --$/:T2T2T#.44T::  Oi ~iQfOOKJ8Y)V%5t<@P Q Q&+#$2 mkll l )*?AR S S&+##^+%"$?!N r c|6t|dr$d|jvr|jd}nd}t||||||\}}} } } } } |r|St |dd}|t dt |r| | z|krt d|jd|j|j }}}|A| d tj |k d }ntj||t }t!||}t"|j}d }t |d d o| d k }|*t&st dt)||}d }d}|*t&st dt+||}d }d}| t+|t-| }d }||| | | | |||||||| }|S)a Create a chunked attention causal mask based on the attention implementation used (stored in the config). This type of attention pattern was mostly democratized by Llama4. If `past_key_values` has an hybrid cache structure, this function will return the mask corresponding to one of the "chunked_attention" layers (to align to what is needed in the `modeling_xxx.py` files). Args: config (`PreTrainedConfig`): The model config. inputs_embeds (`torch.Tensor`): The input embeddings of shape (batch_size, query_length, hidden_dim). This is used only to infer the batch size, query length and dtype. attention_mask (`torch.Tensor`, optional): The 2D attention mask corresponding to padded tokens of shape (batch_size, number_of_seen_tokens+q_length). It can also be an already prepared 4D mask, in which case it is returned as-is. cache_position (`torch.Tensor`): Deprecated and unused. past_key_values (`Cache`, optional): The past key values, if we use a cache. position_ids (`torch.Tensor`, optional) A 2D tensor of shape (batch_size, query_length) indicating the positions of each token in the sequences. or_mask_function (`Callable`, optional): An optional mask function to combine with the chunked causal mask function (by doing the union of both). This is useful to easily overlay another mask on top of the chunked causal one, for example for image tokens handling. and_mask_function (`Callable`, optional): An optional mask function to combine with the chunked causal mask function (by doing the intersection of both). This is useful to easily overlay another mask on top of the chunked causal one, for example for image tokens handling. layer_idx (`int`, *optional*): The cache layer to size the mask against. By default, the first "full_attention" layer is used, which is correct whenever all layers of a given mask type have seen the same tokens. Pass it explicitly for caches where layers of the same type hold different lengths (e.g. per-depth MTP streams). NrTrattention_chunk_sizezQCould not find an `attention_chunk_size` argument in the config, or it is not setzFlash attention cannot handle chunked attention, and the key-value length is larger than the chunk size so the chunked pattern cannot be respected. You should use another `attn_implementation` when instantiating the modelrr)rrFrrrr)rrrrrrr rsr%r*rr' zeros_likerzerosrHr`rrrr<r6rh)rrrorrrrrrrerrprjrkrKrr%r*left_padding_tokensrrrrrs rcreate_chunked_causal_maskr *sT ?L 1 1 do>X6X6X'288>>III #6=./[girss_J 4h 8U^!7>>Jlmmm$F++ I0E 0R0R }   !. 3A 6 8K]MavJ!-444<<@PQ_@`@``eejlemm#k*V3OOO8EXYY1&2MNN H!(9I5 Q Q cV^bcVcd #2 mkll l ()>@P Q Q$$2 mkll l )*?AR S S$' )*?A^_sAtAt u u$!.+%1K r c | |jdkrdS||rdSt|stj|dkrdS|dd|jd dfS)uReturn the 2D padding mask for mamba / linear-attention layers, sized to the local sequence. Returns ``None`` (so the consumer skips masking entirely) when any of: - the input mask is missing or is already a custom 4D attention mask (no 2D padding signal); - the recurrent state already covers past tokens (cached forwards); - the mask is all-ones (un-padded batch — the masking multiply would be a no-op), skipped only outside trace/compile so the graph specialisation stays stable. Otherwise we trim the mask to the trailing ``inputs_embeds.shape[1]`` positions so it aligns with the current forward's local sequence and the consumer can multiply directly without further slicing. Nrr)rhas_previous_stater r'r4rs contiguous)rrrorrs rcreate_recurrent_attention_maskrs&!4!9!9t"'I'I'K'K"t n % %%)Na4G*H*Ht !!!m1!44666 7 B B D DDr )full_attentionlinear_attention)sliding_attentionr) rrchunked_attentioncompressed_sparse_attentionheavily_compressed_attentionminimax_m3_sparsedeepseek_sparse_attentionrconvhybridhybrid_slidingc |} | |||||||d} t| drt| j} t d| Dr|Si} | D]Y} t | }t |tr*|D]\}}|| vr |di| | |<N|di| | | <Z| St| dd tdi| St| dd tdi| Stdi| S)a This function mimics how we create the masks in the `modeling_xxx.py` files, and is used in places like `generate` in order to easily create the masks in advance, when we compile the forwards with Static caches. Args: config (`PreTrainedConfig`): The model config. inputs_embeds (`torch.Tensor`): The input embeddings of shape (batch_size, query_length, hidden_dim). This is used only to infer the batch size, query length and dtype. attention_mask (`torch.Tensor`, optional): The 2D attention mask corresponding to padded tokens of shape (batch_size, number_of_seen_tokens+q_length). It can also be an already prepared 4D mask, in which case it is returned as-is. past_key_values (`Cache`, optional): The past key values, if we use a cache. position_ids (`torch.Tensor`, optional) A 2D tensor of shape (batch_size, query_length) indicating the positions of each token in the sequences. or_mask_function (`Callable`, optional): An optional mask function to combine with the other mask function (by doing the union of both). This is useful to easily overlay another mask on top of the causal one, for example for image tokens handling. and_mask_function (`Callable`, optional): An optional mask function to combine with the other mask function (by doing the intersection of both). This is useful to easily overlay another mask on top of the causal one, for example for image tokens handling. block_sequence_ids (`torch.Tensor`, *optional*): A tensor of same shape as input IDs indicating to which block or group each token belongs to. Tokens from the same block will keep a bidirectional mask within the block, attending causally to the past. Index `-1` can be used for blocks that have to keep complete causality within itself. )rrrorrrrrP layer_typesc3(K|] }|tvVdSr)&LAYER_PATTERN_TO_MASK_FUNCTION_MAPPING)r layer_types rrz,create_masks_for_generate.. s(iiJz!GGiiiiiir rBNrr$) get_text_configrsetranyrrdictitemsrrr r)rrrorrrrrPreffective_config mask_kwargslayer_patterns causal_masks layer_patternriactual_patternactual_functions rcreate_masks_for_generater+sP--//#&(*$,.0  K//9-9:: iiZhiii i i "! ! + K KMB=QM-.. K7D7J7J7L7LVV3NO%\997F7U7U7U7U ^4V /r<s%$$$$$$ 11111122222222IIIIIIII !!ggggggNNNNNNNNN I&?&?RV&W&W&W#&?&?RV&W&W&W#0022&UTTTTTT  H % % x H     h 8    C3sCTX3#cSV[_ 3 8      5< H    %,8$SSSSSS      hshxhhhhVSV VQYVVVV         C TW \d     )<  Y\ afamptat       6;lT6I VY fi  \    *U%*%*:****(, ++,%+++ +  + * + ++++\,%* <(,,%* >  X    ><>/4|>HM >bgbn>>>>.2*.!!%(- !&iiiii i  i  iL4'id ii"&iii L3 i \Diiii`2*.(-!&DDDDD D  D  DL4'D ;D"&DD L3 D \DDDDV2*.((((( (  (  (L4'((((^2*.!&::::: :  :  :L4': L3 :::::z     -   8N7M7O7O4OOO u|  t@S    N26eaea ea<eaL9,t3eaT\ ea ,% ea Tz ea!<$.ea 4 )D0#s :;eaeaeaeaZ)-(,)-.2 || |<|L4'|T\ | ,% | o | $| t+|Tz| \I$||||F26$((,)-aa a<aL4'a!<$. a T\ a o a $a \I$aaaaR)-(,)-.2 AA A<AL4'AT\ A ,% A o A $A t+ATzA \I$AAAAP26$((,)-YY Y<YL4'Y!<$. Y T\ Y o Y $Y \I$YYYYB)-(,)- zz z<zL4'zT\ z ,% z o z $zTzz \I$zzzzB%) EE E<EL4'ET\ E  \D EEEE<):3#D$E+!37 +!3Ihii,>Tstt * *&()-(,)-.2N-N- N-<N-L4'N-T\ N- ,% N- o N- $N- t+N-N-N-N-N-N-r