a h#@shddlZddlZddlmZddlmZddlmZddlm Z ddl m Z GdddeZ d d ZdS) N)FakeTensorProp)CapabilityBasedPartitioner)OperatorSupport)CALLABLE_NODE_OPS)_pytreec@s eZdZejjedddZdS)CudaGraphsSupport)nodereturncs|jtvrdS|jtjjjjfvr&dS|jtj fvr8dSddd}fdd}|j D]}t |||j qVt |||j  S)NFTcSsd|vr|dS|dS)NvalZ fake_result)metar r Q/var/www/auris/lib/python3.9/site-packages/torch/fx/passes/backends/cudagraphs.pymeta_fksz4CudaGraphsSupport.is_node_supported..meta_fkcs t|tjr|jjdkrddS)NZcudaT) isinstancetorchZTensorZdevicetype)tZfound_not_cudar r find_not_cudasz:CudaGraphsSupport.is_node_supported..find_not_cuda)oprtargetropsZatenZembedding_dense_backwarddefaultoperatorgetitemZall_input_nodespytreeZ tree_map_r )selfZ submodulesrrrnr rr is_node_supporteds   z#CudaGraphsSupport.is_node_supportedN)__name__ __module__ __qualname__rZfxNodeboolrr r r r r 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)Zallows_single_node_partition)r propagaterrZpropose_partitionsZfuse_partitions)ZgminputsZ supported_opsZ partitionerZ partitionsZ fused_graphr r r partition_cudagraphs-s r&)rrZ torch.fx.passes.fake_tensor_proprZ!torch.fx.passes.infra.partitionerrZ torch.fx.passes.operator_supportrZtorch.fx.passes.tools_commonrZ torch.utilsrrrr&r r r r s     !