o ߗi0@s0ddlmZddlmZmZmZmZddlmZm Z GdddZ GdddeZ Gdd d eZ Gd d d eZ Gd d d eZGdddeZde dee defddZde dee e fde fddZdee dee e fdee e ffddZde de defdd Zdee d!ee e fdefd"d#Zd$S)%)Enum)DictList NamedTupleSet)map_argNodec@sBeZdZdZdeddfddZddZd d Zd d Zd dZ dS) PartitionzPartition class contains all the information about an individual partition. It also provides necessary methods for manipulation the partition. partition_idreturnNcCs4t|_||_t|_t|_d|_d|_g|_dS)Nr)setnodesr parentschildren bfs_levelused_mem_byteslogical_device_ids)selfr re/home/ubuntu/transcripts/venv/lib/python3.10/site-packages/torch/fx/experimental/partitioner_utils.py__init__ s zPartition.__init__cCs t|jSN)strr rrrr__str__s zPartition.__str__cCs,d|_|jD] }|jt||j7_qdS)Nr)rrget_extra_size_of)rnoderrrrecalculate_mem_sizes zPartition.recalculate_mem_sizecCsXi}t|j|jt|j|j|D] }|jdvr|j|q|j||dS)N>get_attr placeholder)rargs setdefaultkwargsopraddr)rr input_nodesnrrradd_nodes    zPartition.add_nodecs~|jvr=j|i}t|j|jt|j|j|D]}tfdd|jDr6|jdvr6j|q dSdS)Nc3s|]}|jvVqdSr)r.0r'rrr 5s  z(Partition.remove_node..>rr ) rremoverr!r"r#allusersr$r)rrr& input_noderrr remove_node)s      zPartition.remove_node) __name__ __module__ __qualname____doc__intrrrr(r0rrrrr s  r c@s&eZdZUeed<eed<eed<dS)Devicenameavailable_mem_bytes logical_idN)r1r2r3r__annotations__r5rrrrr6<s  r6c@seZdZUeed<eed<dS) NodeLatencymem_latency_seccomputer_latency_secNr1r2r3floatr:rrrrr;Bs  r;c@s&eZdZUeed<eed<eed<dS)PartitionLatencyr<r=overall_latency_secNr>rrrrr@Is  r@c@s eZdZdZdZdZdZdZdS) PartitionModerN)r1r2r3 size_based sparse_nn cost_awarekl_based aot_basedrrrrrBRs rBc@seZdZUeeed<ejZeed<dZ e ed<iZ e e efed<iZe e efed<iZe eeefed<dZeed <d S) PartitionerConfigdevicesmodetransfer_rate_bytes_per_secnode_to_latency_mappingnode_to_partition_mapping#partition_to_logical_device_mappingF saturate_hostN)r1r2r3rr6r:rBrGrNrPr?rQrrr;rRr5rSrTboolrrrrrLZs  rLrrr cCsi}t|j|jt|j|jd}|D]}||vr,t|dd}|r(||j7}qtdqt|dd}|r<||j7}|Std)zGiven a node and a set of nodes, this function return the extra size that needed if this node is included in this set. r size_bytesNznode has no size_bytes attr)rr!r"r#getattr output_size RuntimeError total_size)rrr&total_size_of_input_nodesr'rVrrrres     r partitionrQcsxdtdttfdd}dtdtffdd |}tdddd }|D]}|tdddd }|j|jkr9|}q%|S) zVGiven a partition and its nodes' latency, return a PartitionLatency for this partitionr\r cs`g}jD](}|jdvr qi}t|j|jt|j|jtfdd|Ds-||q|S)z>Given a partition, return a list of nodes on the top bfs level>rr c3s$|] }|jvo |jdvVqdS)>rr N)rr$r)r\rrr+s  zFget_latency_of_one_partition..get_top_nodes..)rr$rr!r"r#anyappend)r\ top_nodesrr&rr]r get_top_nodess    z3get_latency_of_one_partition..get_top_nodesrc s|}|jt|j|j}|j|j}|j|j}t|jj}|rDtdddd}|D]}|t|||} | j|jkrA| }q.|St|||S)zyGiven a top node of a partition, this function returns the latency of the critical path in the partition rOr<r=rA) rAmaxr=r<r r. intersectionrr@) rpartition_latency node_latencyrAr<r=r. max_latencyr'new_partition_latency dfs_helperrQr\rrrjs:  z0get_latency_of_one_partition..dfs_helperrOrb)r rrr@rA)r\rQrar`critical_path_latencyrrerrirget_latency_of_one_partitions&,rl partitionscCs$i}|D] }t||}|||<q|S)zGiven all the partitions and node_to_latency_mapping dictionary, return a mapping dictionary of each partition to its overall latency )rl)rmrQpartition_to_latency_mappingr\rerrr get_partition_to_latency_mappings roparent_partitionchild_partitionrPc Cs|jgkr|jgkr|j|jkrdSd}t}|jD]4}i}t|j|jt|j|j|D]}||jvrM||vrMt|dd}|durH||j7}| |q.q||S)zfGiven two partitions (parent and child), calculate the communication latency between the two. rOrrVN) rr rrr!r"r#rWrXr%) rprqrP comm_size visited_nodesrr&r'rVrrrget_comm_latency_betweens&      rtrncshdtdtdtffdd dttdttfdd}||}d }|D] }|d }||kr1|}q$|S) zGiven all partitions in a graph, find the critical path among all partitions and return its latency as the latency of the whole graph r\latency_so_far_secr csR||j7}|jr'd}|jD]}t||}|||}||kr$|}q|S|S)zJThis function helps to recursively get the latency of a path of partitionsrO)rArrt)r\rumax_latency_secchildcomm_latency_secnew_latency_secrjrnrPrrrjs& z4get_latency_of_partitioned_graph..dfs_helperrmcSsdd|D}|S)zvThis function is to return all the partitions without parents as the starting points of all the paths cSsg|] }t|jdkr|qS)r)lenr)r*r\rrr 2szPget_latency_of_partitioned_graph..get_top_partitions..r)rmtop_partitionsrrrget_top_partitions-sz.get_top_partitionsrO)r r?r)rmrnrPr~r}critical_path_latency_secr\ latency_secrrzr get_latency_of_partitioned_graphs   rN)enumrtypingrrrr torch.fx.noderrr r6r;r@rBrLr5rrlror?rtrrrrrsL 4    Y    $