qi ddlZddlmZmZddlmZej eZeZ dejdejjdejfdZ ddejjdejd ejd ejd ejdzd ed edzdedzdedzdedzdeejdffdZy)N)_flash_attention_forward!flash_attn_supports_top_left_mask)loggingquerymodulereturncR|jtjk(rtjdrtjdSt |j dr|j jStd|jDjjSy)ziIf the query is in float32, return a target dtype compatible with flash attention. Return None otherwise.cuda _is_quantizedc3jK|]+}t|tjjs(|-yw)N) isinstancetorchnnLinear).0layers [/opt/pipecat/venv/lib/python3.12/site-packages/transformers/integrations/flash_attention.py z#get_target_dtype..s&b%z%QVQYQYQ`Q`?abs)33N) dtyperfloat32is_autocast_enabledget_autocast_dtypehasattrconfignextmodulesweight)rrs rget_target_dtyper sv {{emm#  $ $V ,++F3 3 V]]O 4==&& &b6>>+;bbiioo o keyvalueattention_maskdropoutscalingsliding_windowsoftcap is_causalc  | jddrtjd|jd} t d|jDr t d|j dd}|j dd}|j dd}t||} | | n |j} t||||f| | ||||t| |jjt|dr |jndd | } | dfS) Noutput_attentionsFzFlash Attention does not support `output_attentions=True`. Please set your attention to `eager` if you want any of these features.rc3&K|] }|dk( yw)rN)rdims rrz*flash_attention_forward../s +3!8 +szTensor query has shape with a zero dimension. FlashAttention does not support inputs with dim=0. Please check your input shapes or use SDPA instead. layer_idx) query_lengthr(r$ softmax_scaler&r'use_top_left_mask target_dtypeattn_implementationr/)getlogger warning_onceshapeany ValueError transposerr(r_use_top_left_maskr_attn_implementationrr/)rrr!r"r#r$r%r&r'r(kwargsseq_lenr3 attn_outputs rflash_attention_forwardrAszz%u- W kk!nG +u{{ ++ B  OOAq !E --1 C OOAq !E$E62L'2 8H8HI*     %,!"MM>>&-fk&B&"" K$  r )gNNNN)rmodeling_flash_attention_utilsrrutilsr get_logger__name__r6r<TensorrModulerrfloatintbooltuplerAr,r rrLs h   H %68 ELL %((// ekk & !% != HHOO= <<= = << = LL4' =  =T\=$J=T\=d{= 5<< =r