U yh&@sUddlZddlZddlmZmZmZmZmZddlZddl m Z ddl m Z ddl mZdZeeZejed<eejjeejjeejjejjdd d Zejjed d d ZejjeeejjeegefddddZGdddejjZejjeeefd ddZ eejjedddZ!ddddddfejjeejjgefeeejjgefeegefejjdddZ"dS)N)CallableListOptionalSetTuple)make_fx)compile_fx_inner)select_decomp_table logger)inputsbodyoutputsreturnc stj}it|D]*\}}|jd|d}|j|_||<q|D]"}|j|fddd}||<qF|jtfdd|Dd| | tjj i|d S) NZarg_)namecs|SNxZnode_to_subgraph_noderW/var/www/html/venv/lib/python3.8/site-packages/torch/distributed/_spmd/partial_lower.py!z)_create_subgraph_module..)Z arg_transformc3s|]}|VqdSrr).0rrrr %sz*_create_subgraph_module..)result)rootgraph) torchfxZGraph enumerate placeholdermetaZ node_copyoutputtupleZeliminate_dead_codelint GraphModule) r r rsubgraphidxZinpZ subgraph_inpnodeZ subgraph_noderrr_create_subgraph_modules     r*r)rcCsXtdd|jDrTtdd|jDsPtdj|ddd|jDddSd S) Ncss|]}|jtjkVqdSrtargetoperatorgetitemruserrrrr,sz%_is_container_node..css|]}|jtjkVqdSrr,r0rrrr-szmMalformed graph: a container node is used as input for non-getitem nodes. Node: {fmt_node} Users: {fmt_users} css|]}|VqdSr) format_node)rurrrr1s)Zfmt_nodeZ fmt_usersTF)anyusersallAssertionErrorformatr3joinr)rrr_is_container_node+sr<)gm subgraph_namesubgraph_nodesdumperrc sgggttjjddfdd }D]\}|jdkrX||q4|jD]}|kr^||q^|kr4||q4gtjjddfdd }D]$}tfdd |j Ds||qt t tkst t t tkst t }|t |j} t||t|d j} D]}| |qF|j| `|jjd |td | id } tD]4\} } |jjdtj| | fd}| j|_| |qW5QRXtD]"}t |j dkr|j|qdS)N)argrcsl|g}t|dkrh|}t|rJfdd|jD||jq|kr||qdS)Nrc3s|]}|kr|VqdSrrr0)r?rrrLsz;_lower_subgraph_nodes..add_input..)lenpopr<extendr6appendadd)rAstackr))r prologuer?visiblerr add_inputFs  z(_lower_subgraph_nodes..add_inputZget_attr)r#rcsl|g}t|dkrh|}t|r<|j||jqtfdd|jDs|kr|qdS)Nrc3s|]}|kVqdSrrr0rIrrrosz<_lower_subgraph_nodes..add_output..)rBrCr<rDr6r7rE)r#rGr))r rrIrr add_outpuths  z)_lower_subgraph_nodes..add_outputc3s|]}|kVqdSrrr0rKrrrtsz(_lower_subgraph_nodes..Z call_moduletag)opr-argskwargsZ call_function)rOr-rPr)setrrNoderOrErFZall_input_nodesr7r6rBr8r*strrsetattr_InductorModulenextprependZinserting_beforeZ create_noder$r r.r/r"Zreplace_all_uses_withreversedZ erase_node)r=r>r?r@rJr)rArLZsubgraph_moduleZ readable_tagZinsertion_pointZ subgraph_callr(r#Z new_outputr)r r rrHr?rIr_lower_subgraph_nodes8s^            rZcsBeZdZejjddfdd Zejee ejdddZ Z S)rVN)r=rcst||_d|_dSr)super__init__r=compiled)selfr= __class__rrr\s z_InductorModule.__init__)rPrNrc Gs|jdkrNt}t|j|d|}td|t|t|dd|_td|tj |*|jdk sjt |t|W5QRSQRXdS)N)Zdecomposition_tablez%Lowering subgraph (%s) to Inductor...F)Z cudagraphsz,Completed lowering subgraph (%s) to Inductor) r]r rr=r inforlistrZprofilerZrecord_functionr8)r^rNrPZinductor_decompositionsZ decomp_gmrrrforwards    z_InductorModule.forward) __name__ __module__ __qualname__rrr&r\ZTensorrTrrc __classcell__rrr_rrVsrVcCs|jtjjjjtjjjjtjjjjtjjj jfkr6dS|jtjjj j krLdSt |jtj jr~|jtjjjs~d|jdfSdS)N)Fzfused adam is not supported yetTFz& doesn't have a meta kernel registered)r-rZopsZatenZ _fused_adam_defaultZ _fused_adamZ _foreach_add_ZScalarZ _foreach_addflattenZ using_ints isinstanceZ_opsZ OpOverloadZhas_kernel_for_dispatch_keyZ_CZ DispatchKeyZMetar;rrr_is_inductor_compatibles    rm)nodesrcCstdd|D}|tkS)NcSs g|]}t|jdr|qS)zaten.)rTr- startswith)rnrrr sz'_subgraph_predicate..)rBMIN_ATEN_OPS_TO_LOWER)rnZ num_aten_opsrrr_subgraph_predicatesrscCsdSNTrrrrrrrrcCsdSrtrrrrrrrcCsdS)Nr'rrrrrrr)r=node_predicatesubgraph_predicater@rc sgg}tt|jj}tjjttt fdfdd }|j dkr|j dkrR|j}q6||\}}|rr|d |n,t |ddkrt dt ||| g|j}q6fd d |D}t|D] \} } d | } t|| | |q|j||S) a Lower Inductor compatible portions of the graph module to Inductor. Args: node_predicate: user predicate for determining whether to consider a node for lowering. subgraph_predicate: user predicate for determining whether to consider a list of candidate nodes for lowering. dumper: a callback for dumping subgraphs for human digestion. For exmaple, it can be a function that writes to disk/blob storage and returns the path/handle. The returned path/handle for each subgraph will be made available in the subgraph call node in the parent graph, as well as the label of the profiler block for the subgraph. r+cs(t|\}}|s||fS|s$dSdS)N)Fzuser predicaterh)rm)r) should_lowerreason)rurr_node_predicates  z&partial_lower.._node_predicater#r!rMrz,partial_lower: graph break at %s. Reason: %scs g|]}|rt|r|qSr)rs)rrn)rvrrrqsz!partial_lower..Z subgraph_)rWiterrrnrrrSrboolrTrOrErBr warningr rZr%Z recompile) r=rurvr@Znodes_per_subgraphZptrryrwrxr(r?r>r)rurvr partial_lowers6        r})#loggingr.typingrrrrrrZ functorchrZtorch._inductor.compile_fxrZtorch._inductor.decompositionr rr getLoggerrdr Logger__annotations__rrSr&r*r{r<rTrZnnModulerVrmrsr}rrrrsD        b