rhOddlZddlmZddlZddlmZddlmZmZddlm Z m Z ddl m Z dd l mZdd lmZmZmZmZmZmZmZmZdd lmZd d lmZe j8eZdZdZ GddeZ!dZ"GddejFZ$Gdde$Z%Gdde$Z&e$e%e&dZ'GddeZ(GddeZ)GddeZ*Gd d!eZ+Gd"d#eZ,Gd$d%eZ-Gd&d'eZ.gd(Z/y))N)Optional)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/html/ai-insurance-compliance-backend/venv/lib/python3.12/site-packages/transformers/models/diffllama/modular_diffllama.pyrr0rrc>ddtjd|zzz S)Ng?g333333?g333333ӿ)mathexp) layer_idxs r lambda_init_fnr&4s txxy 011 11rc\eZdZdZddedeeffd Z ddejde ejejfdeejdeejd ee d e d eejd e ejeejee ejffd ZxZS)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%r4s r r1zDiffLlamaAttention.__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 $56;N;Nchir hidden_statesposition_embeddingsattention_mask position_idspast_key_value use_cachecache_positionreturnc |j\} } } | } |j|} |j|}|j|}| j | | |j |j jdd} |j | | |j|j jdd}|j | | |j|j jdd}|\}}t| |||\} }|'|||d}|j|||j|\}}t||j}t||j}tjtj |ddd}|j#dddd}tj$| |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$||}tj |dd\}}|||zz }d|jJz |jM|z}|jddjO}|jQ| | d}|jS|}||fS) Nrr sincosrYdimrr`dtype)ptraining)*r-rArBrCviewr8r+ transposer:rupdater%rr;rGcatchunkrepeatmatmulr#sqrtshaper functionalsoftmaxfloat32torddropoutr5rfr$sumrJrKrLrMrErP contiguousreshaperD)rRrSrTrUrVrWrXrYkwargsbsz 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+. L((rNNNNFN)rrr__doc__rrintr1rGTensortuple LongTensorrboolr __classcell__r4s@r r(r(8sG j j8C= jL2637*.59<)||<)#5<<#=><)!. <) u//0 <) ! <)<)!!1!12<) u||Xell3XeELL>Q5RR S<)rr(cJeZdZdZfdZ d dej deej ej fdeejdeejdee de 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)r0r1r _flash_attn_uses_top_left_mask)rRargsrxr4s r r1z!DiffLlamaFlashAttention2.__init__s# $)&) /P.Q+rrSrTrUrVrWrXrYrZc@ t|tr td|j\}} } |j |} |j |} |j |} | j|| |j|jjdd} | j|| |j|jjdd} | j|| |j|jjdd} |+tjd|j| |\}}n|\}}t| | ||\} } |'|||d}|j!| | |j"|\} } | jdd} | jdd} | jdd} |j$r |j&nd}| j(}| j*j,dk7r| j*j,nd}|t.j0k(rt/j2r:t5t.d rt/j6|nt/j8}nMt5|j:d r|j:j<}n |jj>j(}tjd |d | jA|} | jA|} | jA|} t/jB| dd \}}|jEdddd}|jEdddd}tG| | ||| ||tI|dd|jJ|jL }tG| | ||| ||tI|dd|jJ|jL }t/jN||gd }t/jB|dd \}}t/jPt/jR|jT|jVzdt.j0jA| j(}t/jPt/jR|jX|jZzdt.j0jA| j(}||z |j\z}|||zz }d|j\z |j_|z}|ja|| djc}|je|}|dfS)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/transformersrr 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\mpscpuget_autocast_dtype_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 .r_sliding_window)rVrtruse_top_left_maskr>rarc)3 isinstancer ValueErrorr-rArBrCrgr8r+rhr:r2r3 rotary_embrrir%rfr5rddevicetyperGrris_autocast_enabledhasattrrget_autocast_gpu_dtyper)rweightrsrkrlrr9rr>rjr$rurJrKrLrMrErPrwrvrD)rRrSrTrUrVrWrXrYryr|r{r}r~rr^r]r dropout_rate input_dtype device_type target_dtype value_states1 value_states2rrrrrrs r rz DiffLlamaFlashAttention2.forwards nk 2}  &**, UA{{=1 [[/ {{=1 $((eT^^T]]S]]^_abc __S%1I1I4==Yccdeghi #((eT5M5Mt}}]gghiklm  &   G  |\BHC*HC#7 jRUWZ#[ j  %#&snUL'5'<'2E2E2J2Je2Sl))..Y^ %-- '((*u&:;,,[9557 &?@#{{BB #{{1177    >$  (??<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+. D  rr)rrrrr1rGrrrrrrrrrs@r rrs R6:37*.59B!||B!#5<<#=>B!!!1!12 B! u//0 B! ! B!B!!!1!12B! u||Xell3XeELL>Q5RR SB!rrc8eZdZdZ d dej deej ej fdeej deejdee de d eejd eej eej eeej ffd Z y) 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. NrSrTrUrVrWrXrYrZc |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}|}||ddddddd| j$df}| j&j(dk(r2|0| j+} | j+} |j+}|duxr| dkD}tj,j.j1| | |||j2r |j4nd| }tj |dd\}}tj6tj8|j:|j<zdtj> jA| jB}tj6tj8|jD|jFzdtj> jA| jB}||z |jHz}|||zz }d|jHz |jK|z}|jddj+}|j | | d}|jM|}|dfS) Nrr r\r_rarbcudar) attn_mask dropout_pr>rc)'r-rArBrCrgr8r+rhr:rrir%rr;rGrjrkrlrorrrvrrpscaled_dot_product_attentionrfr5r$rurJrKrrrsrdrLrMrErPrD)rRrSrTrUrVrWrXrYrxryr|r{r}r~rr^r]rrr>rrrrrrs r rzDiffLlamaSdpaAttention.forward8sm&**, UA{{=1 [[/ {{=1 #((eT^^T]]S]]^_abc __S%1I1I4==Yccdeghi #((eT5M5Mt}}]gghiklm &S#7 jRUWZ#[ j  %#&snUL'5'<' !&&sE26 kk+. D  rr) rrrrrGrrrrrrrrrr rr0s2637*.59I!||I!#5<<#=>I!!. I! u//0 I! ! I!I!!!1!12I! u||Xell3XeELL>Q5RR SI!rr)eagerflash_attention_2sdpac(eZdZdedeffd ZxZS)DiffLlamaDecoderLayerr)r%cdt|||t|j|||_y)N)r)r%)r0r1DIFFLLAMA_ATTENTION_CLASSES_attn_implementation self_attnrQs r r1zDiffLlamaDecoderLayer.__init__s- +4V5P5PQY_ktur)rrrrrr1rrs@r rrsvv3vvrrceZdZdZdZdZy)DiffLlamaPreTrainedModelFc tj|t|tr|jj j d|jj|jj j d|jj|jj j d|jj|jj j d|jjyy)Nr) r _init_weightsrr(rJdatanormal_r)rIrKrLrM)rRmodules r rz&DiffLlamaPreTrainedModel._init_weightss**62 f0 1    ! ! ) )!T[[-G-G H    ! ! ) )!T[[-G-G H    ! ! ) )!T[[-G-G H    ! ! ) )!T[[-G-G H 2rN)rrr_supports_flex_attn_supports_attention_backendrrrr rrs"'Irrc eZdZy)DiffLlamaModelNrrrr rrr!rrc eZdZy)DiffLlamaForCausalLMNrrrr rrr!rrc eZdZy)"DiffLlamaForSequenceClassificationNrrrr rrr!rrc eZdZy)DiffLlamaForQuestionAnsweringNrrrr rrr!rrc eZdZy)DiffLlamaForTokenClassificationNrrrr rrr!rr)rrrrrr)0r#typingrrGr cache_utilsrrmodeling_flash_attention_utilsrr utilsr gemma.modeling_gemmar llama.modeling_llamar rrrrrrrmistral.modeling_mistralrconfiguration_diffllamar get_loggerrr2_CHECKPOINT_FOR_DOC_CONFIG_FOR_DOCrr&Moduler(rrrrrrrrrr__all__rrr rs$  -i3   24   H %5# : 2a)a)HQ!1Q!hQ!/Q!j 1 "v-v I3 I Z  +  )G  $=  &A  r