`i/6*UddlmZddlZddlZddlZddlZddlZddlmZddlm Z ddl m Z m Z m Z mZmZddlmZddlmZddlmZerdd lmZdadadagad ed <gZd ed <dadaejGddZ ga!ded<da"da#da$da%dIdZ&eGddZ'dJdZ(GddZ)iZ*ded<eGddZ+e+,dgd e+,d!gd"e+,d#gd$e+,d%gd&e+,d'gd(dKd+Z-dLd.Z.dMd1Z/dNd2Z0dOd4Z1dLd5Z2dPd7Z3dQd:Z4dRd<Z5dSd?Z6dId@Z7dTdBZ8e dUdDZ9dVdGZ:dWdHZ;dS)X) annotationsN) dataclass) lru_cache)CallablecastOptional TYPE_CHECKINGUnion)config)get_benchmark_name) OrderedSet)BaseSchedulerNodez#list[tuple[BaseSchedulerNode, int]]nodes_num_elemz%list[tuple[BaseSchedulerNode, float]] node_runtimesc(eZdZUded<dZded<dS)CppOuterLoopFusedCountintinner_kernel_numberrlocal_buffer_numberN)__name__ __module__ __qualname____annotations__rk/home/jaya/work/projects/VOICE-AGENT/VIET/agent-env/lib/python3.11/site-packages/torch/_inductor/metrics.pyrr(s0       rrzlist[CppOuterLoopFusedCount]!cpp_outer_loop_fused_inner_countsreturnNonecdadadatt dadatda da da da dS)Nr) generated_kernel_countgenerated_cpp_vec_kernel_countnum_bytes_accessedrclearrir_nodes_pre_fusioncpp_to_dtype_countrnum_comprehensive_padding)num_matches_for_scatter_upon_const_tensornum_loop_reorderingparallel_reduction_countrrrresetr+;sv%&"%++--- !01- rcPeZdZUdZded<ded<ded<ded<ded<ded<d S) CachedMetricsDeltasz] The subset of metrics we want update across cache hits, e.g., the FxGraphCache. rr!r"r%r&r#r(N)rrr__doc__rrrrr-r-Usf  ''''222222rr- list[str]cHdtjtDS)Ncg|] }|j Srname).0fields r z%get_metric_fields..es L L L5EJ L L Lr) dataclassesfieldsr-rrrget_metric_fieldsr9ds" L LK$67J$K$K L L LLrc:eZdZdZd dZd dZed dZd S) CachedMetricsHelperz A helper class to help calculate and apply counter deltas for those metrics we want to save with cache entries (e.g., FxGraphCache) and apply on a cache hit. rrcni|_tD]}t||j|<dSN)cached_metricsr9globals)selfmetrics r__init__zCachedMetricsHelper.__init__osB ')) < r-)r@ delta_metricsrAs r get_deltaszCachedMetricsHelper.get_deltastsU ')) T TF$+IIf$58KF8S$SM& ! !"33]333rdeltac|tD],}t|xxt||z cc<-dSr=)r9r?getattr)rFrAs r apply_deltasz CachedMetricsHelper.apply_deltas{sN')) 8 8F IIf   !7!7 7     8 8rNrr)rr-)rFr-rr)rrrr.rBrE staticmethodrIrrrr;r;hsf <<<< 4444888\888rr;zdict[str, MetricTable]REGISTERED_METRIC_TABLEScjeZdZUded<ded<dZded<dd Zdd ZddZddZe ddZ dS) MetricTablestr table_namer/ column_namesrrnum_rows_addedrow_fn4Callable[[], dict[str, Optional[Union[str, float]]]]rrc|jtvrdS|t|jtks.Jt|jdtt |jt ks@Jt |jdt t }|gfd|jDz}td|DsJ|ttt|dS)Nz v.s. c g|] }| Srr)r4 column_namerow_dicts rr6z'MetricTable.add_row..sQQQ h{+QQQrc3@K|]}t|tVdSr=) isinstancerO)r4is r z&MetricTable.add_row..s,33!:a%%333333r) rPenabled_metric_tableslenrQr keysr all _write_rowrlistrO)r@rSbnrowrXs @radd_rowzMetricTable.add_rowsR ?"7"9"9 9 9 F6884$%%X6664$%% < > ? ? ? @ @ @ @ @ @ @ @ @ @ @ @ @ @ @ @ @ @s5A''A+.A+rdc|}|jdkr3tj|s||xjdz c_t |D]+\}}t|tr|d}n|d}n|}|||<,t|d5}tj |d}| |ddddS#1swxYwYdS)Nrz.6farkrl) rhrRospathexistsru enumeraterZfloatrorprqrr)r@rdrsidxorig_valnew_valrtrqs rrazMetricTable._write_rowsF''))  ! # #BGNN8,D,D #       q &s^^  MC(E** #%++!"CHH (C  !BZ4888F OOC  ! ! ! ! ! ! ! ! ! ! ! ! ! ! ! ! ! !s.,C''C+.C+r3c:t||}|t|<dSr=)rNrL)r3rQtables rregister_tablezMetricTable.register_tables"D,//). &&&rN)rSrTrr)rrOrJ)rdr/rr)r3rOrQr/rr) rrrrrRrerhrurarKrrrrrNrNsOOON....(5555@@@@ !!!!(///\///rrN slow_fusion) kernel1_pathkernel1_latency kernel2_pathkernel2_latencyfused_kernel_pathfused_kernel_latencyslow_down_ratio graph_stats)graph_idnum_nodes_before_fusionnum_nodes_after_fusionpersistent_red_perf) kernel0_pathrr kernel3_pathkernel0_latencyrrkernel3_latency size_hintsreduction_hint'fusion_failure_due_to_indexing_mismatch)pre_grad_graph_idpost_grad_graph_id node1_name node2_namenode1_debug_strnode2_debug_strcommon_buffer_namesfailure_reasonkernel_metadata kernel_name kernel_pathkernel_categoryrr line_of_codenum_load num_store num_for_loopnum_atomic_addnum_argsxnumelynumelrnumelkernel_args_num_gbkernel_module_coderOcddlm}ddlm}||}||}t j|jjS)z The kernel_module_code is the python module that contains kernel function code. kernel function is the proper triton kernel function annotated with @triton.jit rw) PyCodeCache)get_triton_kernel) codecacherwrapper_benchmarkrloadinspect getsourcefn)rrrmodkernels r_parse_kernel_fn_codersc '&&&&&444444   - . .C  s # #F  VY\ * **rproper_kernel_fn_codercDt|S)zJ Return the line of code for the kernel excluding the decorators. )r^ splitlines)rs r_parse_kernel_line_of_coder&s $//11 2 22rr Optional[str]c~|dkrdStjd|}|s Jd|dS)Nforeachzsize_hints=(\[[0-9, ]*\]),zsize_hints missing!rwresearchgroup)rrms r_parse_size_hintsr-sI)##t /1CDDA #####1 771::rcz|dvrdStjd|}|s Jd|dS)N) reductionpersistent_reductionz$reduction_hint=ReductionHint\.(\w*),z/reduction_hint not found in kernel source code!rwr)rrrs r_parse_reduction_hintr6sLCCCt 9;MNNA ?????1 771::rpatternc,||Sr=)count)rrs r_count_patternr@s & &w / //rc|d}|dsJ|d}|d}||dz|}|d}t |S)Nrdef (z):rw,)r startswithindexsplitr^)rdef_line start_idxend_idxdecl_csvcompss r _count_argsrDs$//11!4H   v & &&& &s##InnT""G A /0H NN3  E u::rkernel_fn_codec@|d}||dS)z Skip decorators. rN)r)r start_poss r_parse_proper_kernel_fn_coderNs&$$V,,I )** %%rnumel_arg_name Optional[int]c~tj|d|}|r"t|dSdS)Nz = ([\d]+)rw)rrrr)rrrs r _parse_numelrVs@ ^0002GHHA1771::trOptional[float]cztjd|}|r"t|dS dS)z inductor meta looks like: inductor_meta={... 'mutated_arg_names': [], 'no_x_dim': False, 'kernel_num_gb': 2.0}, z.kernel_num_gb.:\s*([0-9.]+)rwN)rrr~r)rrrs r_parse_kernel_args_num_gbr^sB 1>BBA QWWQZZ    trrrc * ddlm}||t|t| t |t t td fddS)z An utility to log kernel metadata. We may parse metadata from kernel source code here. It's fine to parse the generated kernel code here since the logging is disabled by default. It would hurt compilation time. rw)"get_kernel_category_by_source_coderc"tdtdtdtdttdtdtdtdS) Nztl.loadztl.storezfor z tl.atomic_addrrrr)rrrr)rrkernel_line_of_coderrrrrsrz%log_kernel_metadata..s&&.$,/&'>N8HH55JKK&''//           rctD]p\}}|tvr[|}tj|rt j||qdS)z Purge the old log file at the beginning when the benchmark script runs. Should do it in the parent process rather than the child processes running each individual model. N) rLitemsr]rhrzr{r|unlinkru)r3rrss rpurge_old_log_filesrs 05577!! e (** * *,,..Hw~~h'' $ (###    !!rOrderedSet[str]c4ttjSr=)enabled_metric_tables_implr r]rrrr]r]s %f&B C CCr config_strct}|dD]E}|}|s|tvsJd|d||F|S)NrzMetric table name z is not registered)r rstriprLadd)renabledr3s rrrs)||G  %%zz||  //// 9 9 9 90//  D Nrr3boolc"|tvSr=)r]r2s ris_metric_table_enabledrs (** **rcJ|tvsJd|dt|S)Nz Metric table z is not defined)rLr2s rrrs3 + + + +-RT-R-R-R + + + #D ))rrJ)rr/)rrOrrO)rrOrr)rrOrrOrr)rrOrrOrr)rrOrrOrr)rrOrrO)rrOrrOrr)rrOrrOrr)rrOrrOrrOrr)rr)rrOrr)r3rOrr)r3rOrrN)< __future__rrpr7rrzrr functoolsrtypingrrrr r torch._inductorr torch._inductor.utilsr torch.utils._ordered_setr torch._inductor.schedulerrr!r"r#rrrr%r&rrr'r(r)r*r+r-r9r;rLrNrrrrrrrrrrrrr]rrrrrrrs"""""""  !!!!!!AAAAAAAAAAAAAA""""""444444//////<;;;;;;!"   8: 9999 !!!!!!!! CE!DDDD,-)!!!!4  3 3 3 3 3 3 3  3MMMM888888882465555 :/:/:/:/:/:/:/ :/z         " -       6 + + + + 33330000&&&&*))))X ! ! ! !DDDD      ++++******r