Zjy SSKrSSKJrJr SSKJr \R "\5r\"5r S\RS\RRS\R4SjrSS\RRS\RS \RS \RS \RS-S \S \S-S\S-S\S-S\S-S\RS-S\\RS44Sjjrg)N)_flash_attention_forward!flash_attn_supports_top_left_mask)loggingquerymodulereturnclUR[R:Xa[R"S5(a[R"S5$[ UR S5(aUR R$[SUR555RR$g)ziIf the query is in float32, return a target dtype compatible with flash attention. Return None otherwise.cuda _is_quantizedc3|# UH2n[U[RR5(dM.Uv M4 g7f)N) isinstancetorchnnLinear).0layers z/root/GenerationalWealth/GenerationalWealth/venv/lib/python3.13/site-packages/transformers/integrations/flash_attention.py #get_target_dtype..s*b+;%z%QVQYQYQ`Q`?a+;s-< <N) dtyperfloat32is_autocast_enabledget_autocast_dtypehasattrconfignextmodulesweight)rrs rget_target_dtyper s| {{emm#  $ $V , ,++F3 3 V]]O 4 4==&& &b6>>+;bbiioo o keyvalueattention_maskdropoutscalingsliding_windowsoftcap is_causals_auxc  XU RSS5(a[RS5 URSn [ SUR55(a [ S5eUR SS5nUR SS5nUR SS5n[X5n U bU O URn [UUUU4U U UUUU[U URR[US5(a UROSU bU RUR 5OSS . U D6nUS4$) Noutput_attentionsFzFlash Attention does not support `output_attentions=True`. Please set your attention to `eager` if you want any of these features.rc3*# UH oS:Hv M g7f)rN)rdims rr*flash_attention_forward..0s +{!8{szTensor query has shape with a zero dimension. FlashAttention does not support inputs with dim=0. Please check your input shapes or use SDPA instead. layer_idx) query_lengthr)r% softmax_scaler'r(use_top_left_mask target_dtypeattn_implementationr2r*)getlogger warning_onceshapeany ValueError transposer r)r_use_top_left_maskr_attn_implementationrr2tor)rrr"r#r$r%r&r'r(r)r*kwargsseq_lenr6 attn_outputs rflash_attention_forwardrEs@zz%u-- W kk!nG +u{{ +++ B  OOAq !E --1 C OOAq !E$E2L'2 8H8HI*     %,!"MM>>&-fk&B&B&""  HHU[[ !%( )K.  r!)gNNNNN)rmodeling_flash_attention_utilsrrutilsr get_logger__name__r9r?TensorrModulerr floatintbooltuplerEr.r!rrPs1 h   H %68 ELL %((// ekk & !% !!%C HHOOC <<C C << C LL4' C  CT\C$JCT\Cd{C <<$ C 5<< Cr!