• J
    [Eager] Optimize Grad by prune useless branch (#47827) · d1e93be1
    Jiabin Yang 提交于
    * [Eager] Fix paddle.grad interface
    
    * [Eager] Support minimum SubGraph for GeneralGrad
    
    * Add needed_nodes to prune grad graph more thoroughly
    
    * [Eager] Add grad_node_trans_mapping_ to record which grad_node has been transformed to AccumulationNode
    
    * [Eager] Fix paddle.grad interface
    
    * Polish code
    
    * remove potential_stop_node
    
    * Add endding_nodes to enhance genSugraph logic
    
    * clear endding_nodes_
    
    * polish code
    
    * rename endding_nodes to endding_nades_
    
    * Refactor grad interface
    
    * Add register_hook case to fix coverage-ci
    
    * Fix code format
    
    * Refactor general_grad
    
    * Add more code comments
    
    * call clear directly to release GradSlotMeta
    
    * fix a mistake
    
    * fix matmul/ multiply kernel logic and optional input in yaml, fill zeros logic and so on.
    
    * fix batch_norm_double_grad yaml optional config
    
    * fix tanh_triple_grad yaml and kernels
    
    * fix MultiplyTripleGradKernel optional logic
    
    * fix merge mistake
    
    * fix compile error
    
    * remove legacy attr for bn
    
    * polish code
    
    * fix some kernel
    
    * merge develop
    
    * fix error
    
    * remote log
    
    * fix kernel with full like
    
    * hide value log behind
    
    * hide value log behind
    
    * fix matmul_triple grad
    Co-authored-by: NWeilong Wu <veyron_wu@163.com>
    d1e93be1
batch_norm_grad_kernel.cu 52.3 KB