o h@shddlZddlmZddlmZddlmZddlmZddl m Z ddl Z GdddeZ d d Z dS) N)CapabilityBasedPartitioner)OperatorSupport)CALLABLE_NODE_OPS)FakeTensorProp)tree_mapc@s"eZdZdejjdefddZdS)CudaGraphsSupportnodereturncs|jtvrdS|jtjjjjfvrdS|jtj fvrdSddd}fdd}|j D] }t |||j q+t |||j  S)NFTcSsd|vr|dS|dS)Nval fake_result)metar r Y/var/www/html/ai/venv/lib/python3.10/site-packages/torch/fx/passes/backends/cudagraphs.pymeta_fksz4CudaGraphsSupport.is_node_supported..meta_fkcs(t|tjr|jjdkrddSdSdS)NcudaT) isinstancetorchTensordevicetype)tfound_not_cudar r find_not_cudasz:CudaGraphsSupport.is_node_supported..find_not_cuda) oprtargetropsatenembedding_dense_backwarddefaultoperatorgetitemall_input_nodesrr )self submodulesrrrnr rris_node_supported s   z#CudaGraphsSupport.is_node_supportedN)__name__ __module__ __qualname__rfxNodeboolr&r r r rr srcCs8t|j|t}t||dd}|}||}|S)z Partition an FX graph into sub-GraphModules that can be validly run under CUDA graphs. For a subgraph to be runnable under CUDA, all of the operations must involve CUDA tensors only/ T)allows_single_node_partition)r propagaterrpropose_partitionsfuse_partitions)gminputs supported_ops partitioner partitions fused_graphr r rpartition_cudagraphs*s  r7)r!torch.fx.passes.infra.partitionerr torch.fx.passes.operator_supportrtorch.fx.passes.tools_commonr torch.fx.passes.fake_tensor_proprtorch.utils._pytreerr rr7r r r rs