" 3jb0ZSSKJr SSKJr "SS5r"SS\5r"SS5rg ) )ArgumentParser)Enumc$\rSrSrSrSrSrSrSrg)AttentionMaskFormat rN) __name__ __module__ __qualname____firstlineno__ MaskIndexEndMaskIndexEndAndStart AttentionMaskNoMask__static_attributes__r a/home/wildlama/miniconda3/lib/python3.13/site-packages/onnxruntime/transformers/fusion_options.pyrr sLMFrrc6\rSrSrSrSrSrSrSrSr Sr S r g ) AttentionOpType AttentionMultiHeadAttentionGroupQueryAttentionPagedAttentioncUR$Nvalueselfs r__str__AttentionOpType.__str__s zzrc,[UR5$r)hashr!r"s r__hash__AttentionOpType.__hash__!sDJJrc4URUR:H$rr )r#others r__eq__AttentionOpType.__eq__$s{{djj((rr N) r r rrrrrrr$r(r,rr rrrrs'I-/%N )rrcd\rSrSrSrSrS SjrSrS\4Sjr \ S5r \ S \ 4S j5r S rg ) FusionOptions(z'Options of fusion in graph optimizationcSUlSUlSUlSUlSUlSUlSUlSUlSUlSUl SUl SUl SUl SUl SUlUS;aSUl[R UlUS:Xa[R$UlOUS;a[R&UlSUlUS;a2SUlSUlSUlSUlSUlSUlSUlgg)NTF)clipqwen3bert)vitr3unetvaer2) enable_geluenable_layer_normenable_attentionenable_rotary_embeddingsuse_multi_head_attention!disable_multi_head_attention_biasenable_skip_layer_normenable_embed_layer_normenable_bias_skip_layer_normenable_bias_geluenable_gelu_approximationenable_qordered_matmulenable_shape_inferenceenable_gemm_fast_gelugroup_norm_channels_lastrrattention_mask_formatrrattention_op_typeenable_nhwc_convenable_group_normenable_skip_group_normenable_bias_splitgeluenable_packed_qkvenable_packed_kvenable_bias_add)r# model_types r__init__FusionOptions.__init__+s!% $(,% ).%16.&*#'+$+/( $).&&*#&*#%*"(,% * *+0D (&9%F%F"  )<)I)ID & + +)<)C)CD &!% 0 0$(D !%)D "*.D ')-D &%)D "$(D !#'D  1rchU(a[RUlg[RUlgr)rrrHr)r# use_raw_masks ruse_raw_attention_mask$FusionOptions.use_raw_attention_maskZs! )<)J)JD &)<)I)ID &rc.[RUlgr)rrrHr"s rdisable_attention_mask$FusionOptions.disable_attention_mask`s%8%?%?"r attn_op_typecXlgr)rI)r#r[s rset_attention_op_type#FusionOptions.set_attention_op_typecs!-rc[UR5nUR(aSUlUR(aSUlUR (aSUlUR(aSUl UR(aSUl UR(aSUl UR(aSUlUR(aSUlUR"(aSUlUR&(aSUlUR((aSUlUR,(aSUlUR.(aUR1S5 UR0(aUR1S5 UR2(aUR55 URS;aUR6(aSUlUR:(aSUlUR>(aSUl URB(aSUl"URF(aSUl$URJ(aSUl&URN(aSUl(URR(aSUl*U$)NFTr6)+r/rQ disable_gelur9disable_layer_normr:disable_rotary_embeddingsr<disable_attentionr;r=disable_skip_layer_normr?disable_embed_layer_normr@disable_bias_skip_layer_normrAdisable_bias_gelurBrCdisable_shape_inferencerErFuse_mask_indexrVno_attention_maskrYuse_group_norm_channels_firstrGdisable_nhwc_convrJdisable_group_normrKdisable_skip_group_normrLdisable_bias_splitgelurMdisable_packed_qkvrNdisable_packed_kvrOdisable_bias_addrP)argsoptionss rparseFusionOptions.parsefs0   "'G   " "(-G %  ) )/4G ,  ! !',G $  ( (/3G ,  ' '-2G *  ( (.3G +  , ,27G /  ! !',G $  ) )04G -  ' '-2G *  % %,0G )     * *5 1  & &  * *4 0  ! !  * * , ??5 511380%%+0(&&,1)++16.**05-&&,1)%%+0($$*/'rparsercURSSSSS9 URSS9 URSSSSS9 URSS 9 URS SSS S9 URSS 9 URS SSSS9 URSS9 URSSSSS9 URSS9 URSSSSS9 URSS9 URSSSSS9 URSS9 URSSSSS9 URSS9 URSSSSS9 URSS9 URSSSS S9 URSS!9 URS"SSS#S9 URSS$9 URS%SSS&S9 URSS'9 URS(SSS)S9 URSS*9 URS+SSS,S9 URSS-9 URS.SSS/S9 URSS09 URS1SSS2S9 URSS39 URS4SSS5S9 URSS69 URS7SSS8S9 URSS99 URS:SSS;S9 URSS<9 URS=SSS>S9 URSS?9 URS@SSSAS9 URSSB9 URSCSSSDS9 URSSE9 URSFSSSGS9 g)HNz--disable_attentionF store_truezdisable Attention fusion)requiredactionhelp)rcz--disable_skip_layer_normz%disable SkipLayerNormalization fusion)rdz--disable_embed_layer_normz&disable EmbedLayerNormalization fusion)rez--disable_bias_skip_layer_normz2disable Add Bias and SkipLayerNormalization fusion)rfz--disable_bias_geluz)disable Add Bias and Gelu/FastGelu fusion)rgz--disable_layer_normz!disable LayerNormalization fusion)raz--disable_geluzdisable Gelu fusion)r`z--enable_gelu_approximationz+enable Gelu/BiasGelu to FastGelu conversion)rCz--disable_shape_inferencez disable symbolic shape inference)rhz--enable_gemm_fast_geluzenable GemmfastGelu fusion)rFz--use_mask_indexzWuse mask index to activate fused attention to speed up. It requires right-side padding!)riz--use_raw_attention_maskzuse raw attention mask. Use this option if your input is not right-side padding. This might deactivate fused attention and get worse performance.)rVz--no_attention_maskz1no attention mask. Only works for model_type=bert)rjz--use_multi_head_attentionzUse MultiHeadAttention instead of Attention operator for testing purpose. Note that MultiHeadAttention might be slower than Attention when qkv are not packed. )r=z--disable_group_normz9not fuse GroupNorm. Only works for model_type=unet or vae)rmz--disable_skip_group_normzPnot fuse Add + GroupNorm to SkipGroupNorm. Only works for model_type=unet or vae)rnz--disable_packed_kvz[not use packed kv for cross attention in MultiHeadAttention. Only works for model_type=unet)rqz--disable_packed_qkvz[not use packed qkv for self attention in MultiHeadAttention. Only works for model_type=unet)rpz--disable_bias_addz0not fuse BiasAdd. Only works for model_type=unet)rrz--disable_bias_splitgeluz6not fuse BiasSplitGelu. Only works for model_type=unet)roz--disable_nhwc_convz:Do not use NhwcConv. Only works for model_type=unet or vae)rlz--use_group_norm_channels_firstznUse channels_first (NCHW) instead of channels_last (NHWC) for GroupNorm. Only works for model_type=unet or vae)rkz--disable_rotary_embeddingsz5Do not fuse rotary embeddings into RotaryEmbedding op) add_argument set_defaults)rws r add_argumentsFusionOptions.add_argumentssv !+  e4 '8  E: (9  U; ,E  ? !<  e4 "4  u5 &  / )>  e< '3  E: %-  %8 j  51 &e  59 !D  e4 (d   U; "L  u5 'c  E: !n  e4 "n  u5 C  U3 &I  59 !M  e4 -B  %@ )H  r)rHrIr>r;rPrBrArMr@r9rCrFrKr:rJrOrNrDr<rErLr?rGr=N)T)r r rr__doc__rRrVrYrr] staticmethodrurrrr rrr/r/(sV1-(^J @./.33jw nw w rr/N)argparserenumrrrr/r rrrs/ $  )d)"l l r