U T?hL0@sHddlmZddlmZGdddZGdddeZGdddZd S) )ArgumentParser)Enumc@seZdZdZdZdZdZdS)AttentionMaskFormatrN)__name__ __module__ __qualname__ MaskIndexEndZMaskIndexEndAndStart AttentionMaskNoMaskrrY/var/www/html/venv/lib/python3.8/site-packages/onnxruntime/transformers/fusion_options.pyr src@s4eZdZdZdZdZdZddZddZd d Z d S) AttentionOpType AttentionMultiHeadAttentionGroupQueryAttentionPagedAttentioncCs|jSNvalueselfrrr__str__szAttentionOpType.__str__cCs t|jSr)hashrrrrr__hash__!szAttentionOpType.__hash__cCs |j|jkSrr)rotherrrr__eq__$szAttentionOpType.__eq__N) rr r rrrrrrrrrrrrsrc@sVeZdZdZddZdddZddZed d d Ze d d Z e e dddZ dS) FusionOptionsz'Options of fusion in graph optimizationcCsd|_d|_d|_d|_d|_d|_d|_d|_d|_d|_ d|_ d|_ d|_ d|_ d|_|dkrhd|_tj|_|dkrtj|_n|dkrtj|_d|_|dkrd|_d|_d|_d|_d|_d|_d|_dS)NTFclipZbertZvitZunetZvaer ) enable_geluenable_layer_normenable_attentionenable_rotary_embeddingsuse_multi_head_attentionZ!disable_multi_head_attention_biasenable_skip_layer_normenable_embed_layer_normenable_bias_skip_layer_normenable_bias_geluenable_gelu_approximationZenable_qordered_matmulenable_shape_inferenceenable_gemm_fast_gelugroup_norm_channels_lastrr attention_mask_formatr r attention_op_typeenable_nhwc_convenable_group_normenable_skip_group_normenable_bias_splitgeluenable_packed_qkvenable_packed_kvenable_bias_add)r model_typerrr__init__+s> zFusionOptions.__init__TcCs|rtj|_ntj|_dSr)rr r/r )rZ use_raw_maskrrruse_raw_attention_maskZs z$FusionOptions.use_raw_attention_maskcCs tj|_dSr)rr r/rrrrdisable_attention_mask`sz$FusionOptions.disable_attention_mask) attn_op_typecCs ||_dSr)r0)rr<rrrset_attention_op_typecsz#FusionOptions.set_attention_op_typecCsBt|j}|jrd|_|jr"d|_|jr.d|_|jr:d|_ |j rFd|_ |j rRd|_ |j r^d|_|jrjd|_|jrvd|_|jrd|_|jrd|_|jrd|_|jr|d|jr|d|jr||jdkr>|jrd|_|jrd|_|jrd|_ |j!rd|_"|j#rd|_$|j%r"d|_&|j'r0d|_(|j)r>d|_*|S)NFTr!)+rr8 disable_gelur"disable_layer_normr#Zdisable_rotary_embeddingsr%disable_attentionr$r&disable_skip_layer_normr'disable_embed_layer_normr(disable_bias_skip_layer_normr)disable_bias_gelur*r+disable_shape_inferencer,r-use_mask_indexr:no_attention_maskr;use_group_norm_channels_firstr.disable_nhwc_convr1disable_group_normr2disable_skip_group_normr3disable_bias_splitgelur4disable_packed_qkvr5disable_packed_kvr6disable_bias_addr7)argsoptionsrrrparsefsb    zFusionOptions.parseparsercCs|jddddd|jdd|jddddd|jdd |jd ddd d|jdd |jd dddd|jdd|jddddd|jdd|jddddd|jdd|jddddd|jdd|jddddd|jdd|jddddd|jdd|jdddd d|jdd!|jd"ddd#d|jdd$|jd%ddd&d|jdd'|jd(ddd)d|jdd*|jd+ddd,d|jdd-|jd.ddd/d|jdd0|jd1ddd2d|jdd3|jd4ddd5d|jdd6|jd7ddd8d|jdd9|jd:ddd;d|jdd<|jd=ddd>d|jdd?|jd@dddAd|jddB|jdCdddDd|jddE|jdFdddGddS)HNz--disable_attentionF store_truezdisable Attention fusion)requiredactionhelp)r@z--disable_skip_layer_normz%disable SkipLayerNormalization fusion)rAz--disable_embed_layer_normz&disable EmbedLayerNormalization fusion)rBz--disable_bias_skip_layer_normz2disable Add Bias and SkipLayerNormalization fusion)rCz--disable_bias_geluz)disable Add Bias and Gelu/FastGelu fusion)rDz--disable_layer_normz!disable LayerNormalization fusion)r?z--disable_geluzdisable Gelu fusion)r>z--enable_gelu_approximationz+enable Gelu/BiasGelu to FastGelu conversion)r+z--disable_shape_inferencez disable symbolic shape inference)rEz--enable_gemm_fast_geluzenable GemmfastGelu fusion)r-z--use_mask_indexzWuse mask index to activate fused attention to speed up. It requires right-side padding!)rFz--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.)r:z--no_attention_maskz1no attention mask. Only works for model_type=bert)rGz--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)rJz--disable_skip_group_normzPnot fuse Add + GroupNorm to SkipGroupNorm. Only works for model_type=unet or vae)rKz--disable_packed_kvz[not use packed kv for cross attention in MultiHeadAttention. Only works for model_type=unet)rNz--disable_packed_qkvz[not use packed qkv for self attention in MultiHeadAttention. Only works for model_type=unet)rMz--disable_bias_addz0not fuse BiasAdd. Only works for model_type=unet)rOz--disable_bias_splitgeluz6not fuse BiasSplitGelu. Only works for model_type=unet)rLz--disable_nhwc_convz:Do not use NhwcConv. Only works for model_type=unet or vae)rIz--use_group_norm_channels_firstznUse channels_first (NCHW) instead of channels_last (NHWC) for GroupNorm. Only works for model_type=unet or vae)rHz--disable_rotary_embeddingsz5Do not fuse rotary embeddings into RotaryEmbedding op) add_argument set_defaultsrSrrr add_argumentss@                      zFusionOptions.add_argumentsN)T) rr r __doc__r9r:r;rr= staticmethodrRrr[rrrrr(s/  5rN)argparserenumrrrrrrrrs