o ڷi'@sLddlmZddlmZddlmZddlmZeeZ GdddeZ dS)) getLogger)Fusion)helper) OnnxModelcseZdZdeffdd ZdedefddZdededed Bfd d Zdededed Bfd d Z dededed BfddZ Z S) FusionGelumodelcst|dddS)NGeluErf)super__init__)selfr __class__Z/home/ubuntu/vllm_env/lib/python3.10/site-packages/onnxruntime/transformers/fusion_gelu.pyr szFusionGelu.__init__input_name_to_nodesoutput_name_to_nodecCs6||||r dS||||rdS||||dS)N)fuse_1fuse_2fuse_3)r erf_noderrrrrfuses zFusionGelu.fusereturnNcCs"|jd|vr dS||jd}t|dks|djdkrdS|d}|j|ds,dS|jd|vr5dS||jd}t|dksI|djdkrKdS|d}|j|dd|}|dur^dS|jj|ddd dkrkdS|jd}|jd|jdkr|dnd} ||j| kr||jd}t|dks|djdkrdS|d} |j| d sdS| jd} n$|j|d| |} | durdS|j| d sdS|| jvrdS|jd} ||||| g} |j| | g||sdS|j | t j d |g| g|j d d } d | _|j| |j|j| j<|d dS)ay This pattern is from PyTorch model Fuse Gelu with Erf into one node: Pattern 1: +-------Mul(0.5)---------------------+ | | | v [root] --> Div -----> Erf --> Add --> Mul --> (B=1.4142...) (1) Pattern 2: +------------------------------------+ | | | v [root] --> Div -----> Erf --> Add --> Mul -->Mul --> (B=1.4142...) (1) (0.5) Note that constant input for Add and Mul could be first or second input: like either A=0.5 or B=0.5 is fine. rNAddMulDiv -?MbP?delta?rinputsoutputsname com.microsoftT)outputlenop_typerhas_constant_input match_parentfind_constant_inputinputis_safe_to_fuse_nodesnodes_to_removeextendr make_nodecreate_node_namedomain nodes_to_addappendthis_graph_namenode_name_to_graph_namer%increase_counter)r rrrchildren add_after_erf mul_after_erfdivsubgraph_inputanothermul_halfsubgraph_outputsubgraph_nodes fused_noderrrrsf        zFusionGelu.fuse_1c Cs,|jd|vr dS||jd}t|dks|djdkrdS|d}|j|ds,dS|jd|vr5dS||jd}t|dksI|djdkrKdS|d}|j|dsXdS|jd|vradS||jd}t|dksu|djdkrwdS|d}|j|dd|}|durdSd} |jj|dd d dkr|j|d d|} | durdS|j| d sdS|j|d|} | durdS| jd|jvrdS|||||g} | r| | |j | |jdg||sdS|j | t jd | jdg|jdg|jd d} d| _|j | |j|j| j<|d dS)a& This pattern is from Keras model Fuse Gelu with Erf into one node: +------------------------------------------+ | | | v [root] --> Div -----> Erf --> Add --> Mul -->Mul (B=1.4142...) (A=1) (A=0.5) Note that constant input for Add and Mul could be first or second input: like either A=0.5 or B=0.5 is fine. rNrrrr!rrrrSqrtg@rr"r&T)r'r(r)rr*r+r, get_parentr-r5r.r/r0rr1r2r3r4r6r7r%r8) r rrrr9r:r;mulr< sqrt_node root_noderArBrrrrmsj      zFusionGelu.fuse_2c Cs|jd|vr dS||jd}t|dks|djdkrdS|d}|j|ds,dS|jd|vr5dS||jd}t|dksI|djdkrKdS|d}|j|dsXdS|j|dd|}|durgdS|jj|ddd }|dkrvdS|j||dkrdnd|} | durdS|jd|vrdS||jd}t|dks|djdkrdS|d} | jd| jdks| jd| jdksdS||||| g} |j | | jdg||sdS|j | t j d | jdg| jdg|jd d } d | _|j| |j|j| j<|d d S)a? This pattern is from TensorFlow model Fuse Gelu with Erf into one node: +----------------------------------------------+ | | | v [root] --> Mul -----> Erf --> Add --> Mul -->Mul (A=0.7071067690849304) (B=1) (B=0.5) Note that constant input for Add and Mul could be first or second input: like either A=0.5 or B=0.5 is fine. rNrrrr!g`?rrrr"r&T)r'r(r)rr*r+r,rDr-r.r/r0rr1r2r3r4r5r6r7r%r8) r rrrr9r:r? first_mulirGlast_mulrArBrrrrsd (     zFusionGelu.fuse_3) __name__ __module__ __qualname__rr dictrboolrrr __classcell__rrr rrs T"MrN) loggingr fusion_baseronnxr onnx_modelrrKloggerrrrrrs