U h-@s,ddlmZGdddZGdddZdS))ArgumentParserc@seZdZdZdZdZdZdS)AttentionMaskFormatrN)__name__ __module__ __qualname__ MaskIndexEndZMaskIndexEndAndStart AttentionMaskNoMaskr r K/tmp/pip-unpacked-wheel-socb9apf/onnxruntime/transformers/fusion_options.pyrsrc@sHeZdZdZddZdddZddZed d Zee d d d Z 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|_|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 enable_nhwc_convenable_group_normenable_skip_group_normenable_bias_splitgeluenable_packed_qkvenable_packed_kvenable_bias_add)self model_typer r r__init__s< zFusionOptions.__init__TcCs|rtj|_ntj|_dSN)rr rr )r'Z use_raw_maskr r ruse_raw_attention_maskFs z$FusionOptions.use_raw_attention_maskcCs tj|_dSr*)rr r)r'r r rdisable_attention_maskLsz$FusionOptions.disable_attention_maskcCsBt|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|_$|j%r"d|_&|j'r0d|_(|j)r>d|_*|S)NFTr)+rr( disable_gelurdisable_layer_normrZdisable_rotary_embeddingsrdisable_attentionrrdisable_skip_layer_normrdisable_embed_layer_normrdisable_bias_skip_layer_normrdisable_bias_gelurrdisable_shape_inferencerruse_mask_indexr+no_attention_maskr,use_group_norm_channels_firstrdisable_nhwc_convr disable_group_normr!disable_skip_group_normr"disable_bias_splitgelur#disable_packed_qkvr$disable_packed_kvr%disable_bias_addr&)argsoptionsr r rparseOsb    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)r0z--disable_embed_layer_normz&disable EmbedLayerNormalization fusion)r1z--disable_bias_skip_layer_normz2disable Add Bias and SkipLayerNormalization fusion)r2z--disable_bias_geluz)disable Add Bias and Gelu/FastGelu fusion)r3z--disable_layer_normz!disable LayerNormalization fusion)r.z--disable_geluzdisable Gelu fusion)r-z--enable_gelu_approximationz+enable Gelu/BiasGelu to FastGelu conversion)rz--disable_shape_inferencez disable symbolic shape inference)r4z--enable_gemm_fast_geluzenable GemmfastGelu fusion)rz--use_mask_indexzWuse mask index to activate fused attention to speed up. It requires right-side padding!)r5z--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)r6z--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. )rz--disable_group_normz9not fuse GroupNorm. Only works for model_type=unet or vae)r9z--disable_skip_group_normzPnot fuse Add + GroupNorm to SkipGroupNorm. Only works for model_type=unet or vae)r:z--disable_packed_kvz[not use packed kv for cross attention in MultiHeadAttention. Only works for model_type=unet)r=z--disable_packed_qkvz[not use packed qkv for self attention in MultiHeadAttention. Only works for model_type=unet)r<z--disable_bias_addz0not fuse BiasAdd. Only works for model_type=unet)r>z--disable_bias_splitgeluz6not fuse BiasSplitGelu. Only works for model_type=unet)r;z--disable_nhwc_convz:Do not use NhwcConv. Only works for model_type=unet or vae)r8z--use_group_norm_channels_firstznUse channels_first (NCHW) instead of channels_last (NHWC) for GroupNorm. Only works for model_type=unet or vae)r7z--disable_rotary_embeddingsz5Do not fuse rotary embeddings into RotaryEmbedding op) add_argument set_defaultsrBr r r add_argumentss@                      zFusionOptions.add_argumentsN)T) rrr __doc__r)r+r, staticmethodrArrJr r r rrs-  5rN)argparserrrr r r rs