% &hsRddlZddlmZmZddlZddlZddlmZddlmZm Z ddl m Z m Z ddl mZdd lmZdd lmZmZmZmZmZmZmZmZdd lmZd d lmZej<eZ dZ!dZ"GddeZ#dZ$GddejJZ&Gdde&Z'Gdde&Z(e&e'e(dZ)GddeZ*GddeZ+GddeZ,Gd d!eZ-Gd"d#eZ.Gd$d%eZ/Gd&d'eZ0gd(Z1y))N)OptionalTuple)nn)Cache StaticCache)_flash_attention_forward!flash_attn_supports_top_left_mask)logging)GemmaForCausalLM)LlamaDecoderLayerLlamaForQuestionAnsweringLlamaForSequenceClassificationLlamaForTokenClassification LlamaModelLlamaPreTrainedModelapply_rotary_pos_emb repeat_kv) MistralMLP)DiffLlamaConfigzkajuma/DiffLlama-0.3B-handcutrc eZdZy) DiffLlamaMLPN__name__ __module__ __qualname__/var/www/pru.catia.catastroantioquia-mas.com/valormas/lib/python3.12/site-packages/transformers/models/diffllama/modular_diffllama.pyrr1r rc>ddtjd|zzz S)Ng?g333333?g333333ӿ)mathexp) layer_idxs r!lambda_init_fnr'5s txxy 011 11r cbeZdZdZddedeeffd Z ddejde ejejfdeejdeejd ee d e d e d eejd e ejeejee ejffdZxZS)DiffLlamaAttentionz=Multi-headed attention from 'Attention Is All You Need' paperconfigr&ct|||_||_|-tj d|j jd|j|_|j|_ |j|_ t|d|j|jz|_ |j|_|j|jz|_|j |_|j"|_d|_t'j(|j|j|jz|j*|_t'j(|j|j|jz|j*|_t'j(|j|j|jz|j*|_t'j(|j|jz|j|j*|_t5||_t'j8t;j<d|j>|jf|_ t'j8t;j<d|j>|jf|_!t'j8t;j<d|j>|jf|_"t'j8t;j<d|j>|jf|_#t'jHd|jz|jJd |_&y) NzInstantiating z without passing a `layer_idx` is not recommended and will lead to errors during the forward call if caching is used. Please make sure to provide a `layer_idx` when creating this class.head_dimT)biasr)sizer F)epselementwise_affine)'super__init__r*r&logger warning_once __class__rattention_dropout hidden_sizenum_attention_heads num_headsgetattrr,num_key_value_headsnum_key_value_groupsmax_position_embeddings rope_theta is_causalrLinearattention_biasq_projk_projv_projo_projr' lambda_init Parametertorchnormallambda_std_dev lambda_q1 lambda_k1 lambda_q2 lambda_k2RMSNorm rms_norm_eps groupnormselfr*r&r5s r!r2zDiffLlamaAttention.__init__<s~  "      !8!8 9:,,  "(!9!9!--33 D4D4D4VW #)#=#= $(NNd6N6N$N!'-'E'E$ ++ii 0 0$..4==2PW]WlWlm ii 0 0$2J2JT]]2Zagavavw ii 0 0$2J2JT]]2Zagavavw ii >@P@PW]WlWlm ))4ell1f6K6KSWS`S`Rb&cdell1f6K6KSWS`S`Rb&cdell1f6K6KSWS`S`Rb&cdell1f6K6KSWS`S`Rb&cdA $56;N;Nchir hidden_statesposition_embeddingsattention_mask position_idspast_key_valueoutput_attentions use_cachecache_positionreturnc  $|j\} } } | } |j|}|j|}|j|}|j | | |j |j jdd}|j | | |j|j jdd}|j | | |j|j jdd}|\}}t||||\}}|'|||d}|j|||j|\}}t||j}t||j}tjtj |ddd}|j#dddd}tj$||jddt'j(|j z }|#|ddddddd|j*df}||z}t,j.j1|dtj2j5|j6}t,j.j9||j:|j< }tj>tj@|jB|jDzdtj2j5|j6}tj>tj@|jF|jHzdtj2j5|j6}||z |jJz}tj$||}tj |dd\}}|||zz }d|jJz |jM|z}|jddjO}|jQ| | d}|jS|}|sd}||fS) Nrr sincosr[dimrrbdtype)ptraining)*r.rBrCrDviewr9r, transposer;rupdater&rr<rHcatchunkrepeatmatmulr$sqrtshaper functionalsoftmaxfloat32torfdropoutr6rhr%sumrKrLrMrNrFrQ contiguousreshaperE)rSrTrUrVrWrXrYrZr[kwargsbsz target_len_q_len query_states key_states value_statesr`r_ cache_kwargs attn_weights causal_masklambda_1lambda_2 lambda_full attn_output attn_output1 attn_output2s r!forwardzDiffLlamaAttention.forward^s+//1Z{{=1 [[/ {{=1 #((eT^^T]]S]]^_abc __S%1I1I4==Yccdeghi #((eT5M5Mt}}]gghiklm &S#7 jRUWZ#[ j  %#&snUL'5'<' %*[[aQ%G" l"[<%?? 4+++t~~k/JJ !++Aq1<<> !))#ub9 kk+.  LL((r NNNNFFN)rrr__doc__rrintr2rHTensorr LongTensorrboolr __classcell__r5s@r!r)r)9sG j j8C= jL2637*."'59B)||B)#5<<#=>B)!. B) u//0 B) ! B) B)B)!!1!12B) u||Xell3XeELL>Q5RR SB)r r)cPeZdZdZfdZ d dej deej ej fdeejdeejdee de d e d eejd eej eej eeej ffd Z xZ S)DiffLlamaFlashAttention2aN DiffLlama flash attention module. This module inherits from `DiffLlamaAttention` as the weights of the module stays untouched. The only required change would be on the forward pass where it needs to correctly call the public API of flash attention and deal with padding tokens in case the input contains any of them. cBt||i|t|_yr)r1r2r _flash_attn_uses_top_left_mask)rSargsrzr5s r!r2z!DiffLlamaFlashAttention2.__init__s# $)&) /P.Q+r rTrUrVrWrXrYrZr[r\c  t|tr tdd}|j\} } } |j |} |j |} |j |}| j| | |j|jjdd} | j| | |j|jjdd} |j| | |j|jjdd}|+tjd|j||\}}n|\}}t| | ||\} } |'|||d}|j!| ||j"|\} }| jdd} | jdd} |jdd}|j$r |j&nd}| j(}|t*j,k(rt+j.rt+j0}nMt3|j4dr|j4j6}n |jj8j(}tjd |d | j;|} | j;|} |j;|}t+j<|dd \}}|j?dddd}|j?dddd}tA| | ||| ||tC|d d|jD|jF }tA| | ||| ||tC|d d|jD|jF }t+jH||gd }t+j<|dd \}}t+jJt+jL|jN|jPzdt*j,j;| j(}t+jJt+jL|jR|jTzdt*j,j;| j(}||z |jVz}|||zz }d|jVz |jY|z}|j[| | dj]}|j_|}|sd}|fS)Nz`static` cache implementation is not compatible with `attn_implementation==flash_attention_2` make sure to use `sdpa` in the mean time, and open an issue at https://github.com/huggingface/transformersFrr aYThe attention layers in this model are transitioning from computing the RoPE embeddings internally through `position_ids` (2D tensor with the indexes of the tokens), to using externally computed `position_embeddings` (Tuple of tensors, containing cos and sin). In v4.46 `position_ids` will be removed and `position_embeddings` will be mandatory.r^_pre_quantization_dtypezThe input hidden states seems to be silently casted in float32, this might be related to the fact you have upcasted embedding or layer norm layers in float32. We will cast back the input in .rasliding_window)rWrvruse_top_left_maskr?rcre)0 isinstancer ValueErrorr.rBrCrDrir9r,rjr;r3r4 rotary_embrrkr&rhr6rfrHrtis_autocast_enabledget_autocast_gpu_dtypehasattrr*rweightrurmrnr r:rr?rlr%rwrKrLrMrNrFrQryrxrE)rSrTrUrVrWrXrYrZr[r{r~r}rrrr`r_r dropout_rate input_dtype target_dtype value_states1 value_states2rrrrrrrs r!rz DiffLlamaFlashAttention2.forwardsh nk 2}  "%**, UA{{=1 [[/ {{=1 $((eT^^T]]S]]^_abc __S%1I1I4==Yccdeghi #((eT5M5Mt}}]gghiklm  &   G  |\BHC*HC#7 jRUWZ#[ j  %#&snUL'5'<'$  (??<8L#|4J'??<8L',{{<'J$ }%,,Q1a8 %,,Q1a8 /     % "4)94@"AAnn  0     % "4)94@"AAnn  ii| <"E %*[[aQ%G" l99UYYt~~'FBV[VcVcdehh    99UYYt~~'FBV[VcVcdehh    )D,<,<< "[<%?? 4+++t~~k/JJ !))#ub9DDF kk+.  LL((r r)rrrrr2rHrrrrrrrrrs@r!rrs R6:37*."'59D)||D)#5<<#=>D)!!1!12 D) u//0 D) ! D) D)D)!!1!12D) u||Xell3XeELL>Q5RR SD)r rcJeZdZdZ d dej deej ej fdeej deejdee de de d eejd eej eej eeej fffd Z xZ S) DiffLlamaSdpaAttentiona DiffLlama attention module using torch.nn.functional.scaled_dot_product_attention. This module inherits from `DiffLlamaAttention` as the weights of the module stays untouched. The only changes are on the forward pass to adapt to SDPA API. rTrUrVrWrXrYrZr[r\c |r,tjdt| ||||||||S|j \} } } |j |} |j |}|j|}| j| | |j|jjdd} |j| | |j|jjdd}|j| | |j|jjdd}|\}}t| |||\} }|'|||d}|j|||j|\}}t!||j"}t!||j"}t%j&t%j(|ddd}|j+dddd}|}||ddddddd|j,df}| j.j0d k(r2|0| j3} |j3}|j3}|| dkDrd nd }t$j4j6j9| ||||j:r |j<nd | }t%j(|dd\}}t%j>t%j@|jB|jDzdt$jFjI| jJ}t%j>t%j@|jL|jNzdt$jFjI| jJ}||z |jPz}|||zz }d|jPz |jS|z}|jddj3}|j| | d}|jU|}|dfS)NaDiffLlamaModel is using DiffLlamaSdpaAttention, but `torch.nn.functional.scaled_dot_product_attention` does not support `output_attentions=True`. Falling back to the manual attention implementation, but specifying the manual implementation will be required from Transformers version v5.0.0 onwards. This warning can be removed using the argument `attn_implementation="eager"` when loading the model.)rTrVrWrXrYrZr[rUrr r^rarcrdcudaTFr) attn_mask dropout_pr?re)+r3r4r1rr.rBrCrDrir9r,rjr;rrkr&rr<rHrlrmrnrqdevicetyperxrrrscaled_dot_product_attentionrhr6r%rwrKrLrtrurfrMrNrFrQrE)rSrTrUrVrWrXrYrZr[rzr{r~r}rrrr`r_rrr?rrrrrrr5s r!rzDiffLlamaSdpaAttention.forwardAs    [ 7?+-)-"3#-$7#  &**, UA{{=1 [[/ {{=1 #((eT^^T]]S]]^_abc __S%1I1I4==Yccdeghi #((eT5M5Mt}}]gghiklm &S#7 jRUWZ#[ j  %#&snUL'5'<' !&&sE26 kk+. D  r r) rrrrrHrrrrrrrrrs@r!rr9s2637*."'59\!||\!#5<<#=>\!!. \! u//0 \! ! \! \!\!!!1!12\! u||Xell3XeELL>Q5RR S\!\!r r)eagerflash_attention_2sdpac(eZdZdedeffd ZxZS)DiffLlamaDecoderLayerr*r&cdt|||t|j|||_y)N)r*r&)r1r2DIFFLLAMA_ATTENTION_CLASSES_attn_implementation self_attnrRs r!r2zDiffLlamaDecoderLayer.__init__s- +4V5P5PQY_ktur )rrrrrr2rrs@r!rrsvv3vvr rceZdZdZdZy)DiffLlamaPreTrainedModelFN)rrr_supports_flex_attn_supports_attention_backendrr r!rrs"'r rc eZdZy)DiffLlamaModelNrrr r!rrr"r rc eZdZy)DiffLlamaForCausalLMNrrr r!rrr"r rc eZdZy)"DiffLlamaForSequenceClassificationNrrr r!rrr"r rc eZdZy)DiffLlamaForQuestionAnsweringNrrr r!rrr"r rc eZdZy)DiffLlamaForTokenClassificationNrrr r!rrr"r r)rrrrrr)2r$typingrrrHtorch.utils.checkpointr cache_utilsrrmodeling_flash_attention_utilsr r utilsr gemma.modeling_gemmar llama.modeling_llamarrrrrrrrmistral.modeling_mistralrconfiguration_diffllamar get_loggerrr3_CHECKPOINT_FOR_DOC_CONFIG_FOR_DOCrr'Moduler)rrrrrrrrrr__all__rr r!rs$ " -i3   24   H %5# : 2g)g)TS)1S)ld!/d!P 1 "v-v(3(  Z  +  )G  $=  &A  r