+ G-jb0\^RIHt^RIHt!RR4t!RR]4t!RR4tR #) ArgumentParser)Enumc&]tRt^ t^t^t^t^tRtR#)AttentionMaskFormatN) __name__ __module__ __qualname____firstlineno__ MaskIndexEndMaskIndexEndAndStart AttentionMaskNoMask__static_attributes__rx/Volumes/fast/ai/experiments/nudenet-smoke/.venv/lib/python3.14/site-packages/onnxruntime/transformers/fusion_options.pyrr sLMFrrcBa]tRt^toRtRtRtRtRtRt Rt Rt Vt R #) AttentionOpType AttentionMultiHeadAttentionGroupQueryAttentionPagedAttentioncVP#Nvalueselfs&r__str__AttentionOpType.__str__s zzrc,\VP4#r)hashrrs&r__hash__AttentionOpType.__hash__!sDJJrc4VPVP8H#rr)rothers&&r__eq__AttentionOpType.__eq__$s{{djj((rrN) r r r r rrrrr r$r(r__classdictcell__ __classdict__s@rrrs/I-/%N ))rrcxa]tRt^(toRtRtR RltRtV3RlRlt] R4t ] V3RlR l4t R t Vt R #) FusionOptionsz'Options of fusion in graph optimizationcRVnRVnRVnRVnRVnRVnRVnRVnRVnRVn RVn RVn RVn RVn RVnVR9dRVn\P VnVR8Xd\P$VnMVR9d\P&VnRVnVR9d4RVnRVnRVnRVnRVnRVnRVnR#R#)TFbertN)clipqwen3)vitr2unetvaer1) 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_formatr rattention_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 "$(D !#'D  1rcnV'd\PVnR#\PVnR#r)rrrFr )r use_raw_masks&&ruse_raw_attention_mask$FusionOptions.use_raw_attention_maskZs! )<)J)JD &)<)I)ID &rc0\PVnR#r)rrrFrs&rdisable_attention_mask$FusionOptions.disable_attention_mask`s%8%?%?"rc <V^8dQhRS[/#) attn_op_type)r)formatr,s"r __annotate__FusionOptions.__annotate__cs../.rcWnR#r)rG)rr[s&&rset_attention_op_type#FusionOptions.set_attention_op_typecs!-rc \VP4pVP'dRVnVP'dRVnVP 'dRVnVP'dRVn VP'dRVn VP'dRVn VP'dRVnVP'dRVnVP"'dRVnVP&'dRVnVP('dRVnVP,'dRVnVP.'dVP1R4VP0'dVP1R4VP2'dVP54VPR9dVP6'dRVnVP:'dRVnVP>'dRVn VPB'dRVn"VPF'dRVn$VPJ'dRVn&VPN'dRVn(VPR'dRVn*V#)FTr4)+r.rO disable_gelur7disable_layer_normr8disable_rotary_embeddingsr:disable_attentionr9r;disable_skip_layer_normr=disable_embed_layer_normr>disable_bias_skip_layer_normr?disable_bias_gelur@rAdisable_shape_inferencerCrDuse_mask_indexrTno_attention_maskrWuse_group_norm_channels_firstrEdisable_nhwc_convrHdisable_group_normrIdisable_skip_group_normrJdisable_bias_splitgelurKdisable_packed_qkvrLdisable_packed_kvrMdisable_bias_addrN)argsoptionss& rparseFusionOptions.parsefs0    "'G   " " "(-G %  ) ) )/4G ,  ! ! !',G $  ( ( (/3G ,  ' ' '-2G *  ( ( (.3G +  , , ,27G /  ! ! !',G $  ) ) )04G -  ' ' '-2G *  % % %,0G )      * *5 1  & & &  * *4 0  ! ! !  * * , ??5 5111380%%%+0(&&&,1)+++16.***05-&&&,1)%%%+0($$$*/'rc <V^8dQhRS[/#)rZparserr)r\r,s"rr]r^sw w nw rcVPRRRRR7VPRR7VPRRRRR7VPRR7VPR RRR R7VPRR 7VPR RRR R7VPRR7VPRRRRR7VPRR7VPRRRRR7VPRR7VPRRRRR7VPRR7VPRRRRR7VPRR7VPRRRRR7VPRR7VPRRRRR7VPRR 7VPR!RRR"R7VPRR#7VPR$RRR%R7VPRR&7VPR'RRR(R7VPRR)7VPR*RRR+R7VPRR,7VPR-RRR.R7VPRR/7VPR0RRR1R7VPRR27VPR3RRR4R7VPRR57VPR6RRR7R7VPRR87VPR9RRR:R7VPRR;7VPR7VPR?RRR@R7VPRRA7VPRBRRRCR7VPRRD7VPRERRRFR7RG#)Hz--disable_attentionF store_truezdisable Attention fusion)requiredactionhelp)rfz--disable_skip_layer_normz%disable SkipLayerNormalization fusion)rgz--disable_embed_layer_normz&disable EmbedLayerNormalization fusion)rhz--disable_bias_skip_layer_normz2disable Add Bias and SkipLayerNormalization fusion)riz--disable_bias_geluz)disable Add Bias and Gelu/FastGelu fusion)rjz--disable_layer_normz!disable LayerNormalization fusion)rdz--disable_geluzdisable Gelu fusion)rcz--enable_gelu_approximationz+enable Gelu/BiasGelu to FastGelu conversion)rAz--disable_shape_inferencez disable symbolic shape inference)rkz--enable_gemm_fast_geluzenable GemmfastGelu fusion)rDz--use_mask_indexzWuse mask index to activate fused attention to speed up. It requires right-side padding!)rlz--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.)rTz--no_attention_maskz1no attention mask. Only works for model_type=bert)rmz--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)rpz--disable_skip_group_normzPnot fuse Add + GroupNorm to SkipGroupNorm. Only works for model_type=unet or vae)rqz--disable_packed_kvz[not use packed kv for cross attention in MultiHeadAttention. Only works for model_type=unet)rtz--disable_packed_qkvz[not use packed qkv for self attention in MultiHeadAttention. Only works for model_type=unet)rsz--disable_bias_addz0not fuse BiasAdd. Only works for model_type=unet)ruz--disable_bias_splitgeluz6not fuse BiasSplitGelu. Only works for model_type=unet)rrz--disable_nhwc_convz:Do not use NhwcConv. Only works for model_type=unet or vae)roz--use_group_norm_channels_firstznUse channels_first (NCHW) instead of channels_last (NHWC) for GroupNorm. Only works for model_type=unet or vae)rnz--disable_rotary_embeddingsz5Do not fuse rotary embeddings into RotaryEmbedding opN) add_argument set_defaults)r{s&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)rFrGr<r9rNr@r?rKr>r7rArDrIr8rHrMrLrBr:rCrJr=rEr;N)T)r r r r __doc__rPrTrWr` staticmethodrxrrr*r+s@rr.r.(sO1-(^J @..33jw w rr.N)argparserenumrrrr.rrrrs/ $  )d)"l l r