Upload folder using huggingface_hub

.gitattributes

@@ -1,35 +1,40 @@
-*.7z filter=lfs diff=lfs merge=lfs -text
-*.arrow filter=lfs diff=lfs merge=lfs -text
-*.bin filter=lfs diff=lfs merge=lfs -text
-*.bz2 filter=lfs diff=lfs merge=lfs -text
-*.ckpt filter=lfs diff=lfs merge=lfs -text
-*.ftz filter=lfs diff=lfs merge=lfs -text
-*.gz filter=lfs diff=lfs merge=lfs -text
+*.safetensors filter=lfs diff=lfs merge=lfs -text
 *.h5 filter=lfs diff=lfs merge=lfs -text
-*.joblib filter=lfs diff=lfs merge=lfs -text
-*.lfs.* filter=lfs diff=lfs merge=lfs -text
-*.mlmodel filter=lfs diff=lfs merge=lfs -text
-*.model filter=lfs diff=lfs merge=lfs -text
-*.msgpack filter=lfs diff=lfs merge=lfs -text
-*.npy filter=lfs diff=lfs merge=lfs -text
-*.npz filter=lfs diff=lfs merge=lfs -text
 *.onnx filter=lfs diff=lfs merge=lfs -text
-*.ot filter=lfs diff=lfs merge=lfs -text
-*.parquet filter=lfs diff=lfs merge=lfs -text
-*.pb filter=lfs diff=lfs merge=lfs -text
-*.pickle filter=lfs diff=lfs merge=lfs -text
-*.pkl filter=lfs diff=lfs merge=lfs -text
-*.pt filter=lfs diff=lfs merge=lfs -text
-*.pth filter=lfs diff=lfs merge=lfs -text
-*.rar filter=lfs diff=lfs merge=lfs -text
-*.safetensors filter=lfs diff=lfs merge=lfs -text
-saved_model/**/* filter=lfs diff=lfs merge=lfs -text
-*.tar.* filter=lfs diff=lfs merge=lfs -text
-*.tar filter=lfs diff=lfs merge=lfs -text
-*.tflite filter=lfs diff=lfs merge=lfs -text
-*.tgz filter=lfs diff=lfs merge=lfs -text
-*.wasm filter=lfs diff=lfs merge=lfs -text
-*.xz filter=lfs diff=lfs merge=lfs -text
-*.zip filter=lfs diff=lfs merge=lfs -text
-*.zst filter=lfs diff=lfs merge=lfs -text
-*tfevents* filter=lfs diff=lfs merge=lfs -text
+*.bin filter=lfs diff=lfs merge=lfs -text
+context_encoding_model/_tp0_bk0/graph.neff filter=lfs diff=lfs merge=lfs -text
+context_encoding_model/_tp0_bk0/metaneff.pb filter=lfs diff=lfs merge=lfs -text
+context_encoding_model/_tp0_bk0/model.MODULE_f4171003694760566af4+a9cd68fb.hlo_module.pb filter=lfs diff=lfs merge=lfs -text
+context_encoding_model/_tp0_bk0/model.MODULE_f4171003694760566af4+a9cd68fb.neff filter=lfs diff=lfs merge=lfs -text
+context_encoding_model/_tp0_bk1/graph.neff filter=lfs diff=lfs merge=lfs -text
+context_encoding_model/_tp0_bk1/metaneff.pb filter=lfs diff=lfs merge=lfs -text
+context_encoding_model/_tp0_bk1/model.MODULE_2914133a46cb7b4660ab+d7af8a84.hlo_module.pb filter=lfs diff=lfs merge=lfs -text
+context_encoding_model/_tp0_bk1/model.MODULE_2914133a46cb7b4660ab+d7af8a84.neff filter=lfs diff=lfs merge=lfs -text
+context_encoding_model/_tp0_bk2/graph.neff filter=lfs diff=lfs merge=lfs -text
+context_encoding_model/_tp0_bk2/metaneff.pb filter=lfs diff=lfs merge=lfs -text
+context_encoding_model/_tp0_bk2/model.MODULE_00594b8bc68e927f3dbe+1ad60ced.hlo_module.pb filter=lfs diff=lfs merge=lfs -text
+context_encoding_model/_tp0_bk2/model.MODULE_00594b8bc68e927f3dbe+1ad60ced.neff filter=lfs diff=lfs merge=lfs -text
+context_encoding_model/_tp0_bk3/graph.neff filter=lfs diff=lfs merge=lfs -text
+context_encoding_model/_tp0_bk3/metaneff.pb filter=lfs diff=lfs merge=lfs -text
+context_encoding_model/_tp0_bk3/model.MODULE_b3ddbc97e5f0d1d64c82+155de413.hlo_module.pb filter=lfs diff=lfs merge=lfs -text
+context_encoding_model/_tp0_bk3/model.MODULE_b3ddbc97e5f0d1d64c82+155de413.neff filter=lfs diff=lfs merge=lfs -text
+layout_opt/graph.neff filter=lfs diff=lfs merge=lfs -text
+layout_opt/model/graph.hlo filter=lfs diff=lfs merge=lfs -text
+model.pt filter=lfs diff=lfs merge=lfs -text
+token_generation_model/_tp0_bk0/graph.neff filter=lfs diff=lfs merge=lfs -text
+token_generation_model/_tp0_bk0/metaneff.pb filter=lfs diff=lfs merge=lfs -text
+token_generation_model/_tp0_bk0/model.MODULE_6ef5ba8b41fbbe77f080+74ae8282.hlo_module.pb filter=lfs diff=lfs merge=lfs -text
+token_generation_model/_tp0_bk0/model.MODULE_6ef5ba8b41fbbe77f080+74ae8282.neff filter=lfs diff=lfs merge=lfs -text
+token_generation_model/_tp0_bk0/wrapped_neff.hlo filter=lfs diff=lfs merge=lfs -text
+token_generation_model/_tp0_bk1/graph.neff filter=lfs diff=lfs merge=lfs -text
+token_generation_model/_tp0_bk1/metaneff.pb filter=lfs diff=lfs merge=lfs -text
+token_generation_model/_tp0_bk1/model.MODULE_d608453625db6ed38994+e5eecdd4.hlo_module.pb filter=lfs diff=lfs merge=lfs -text
+token_generation_model/_tp0_bk1/model.MODULE_d608453625db6ed38994+e5eecdd4.neff filter=lfs diff=lfs merge=lfs -text
+token_generation_model/_tp0_bk2/graph.neff filter=lfs diff=lfs merge=lfs -text
+token_generation_model/_tp0_bk2/metaneff.pb filter=lfs diff=lfs merge=lfs -text
+token_generation_model/_tp0_bk2/model.MODULE_0ae1021f5dbf9cbac54d+2aa9c8c9.hlo_module.pb filter=lfs diff=lfs merge=lfs -text
+token_generation_model/_tp0_bk2/model.MODULE_0ae1021f5dbf9cbac54d+2aa9c8c9.neff filter=lfs diff=lfs merge=lfs -text
+token_generation_model/_tp0_bk3/graph.neff filter=lfs diff=lfs merge=lfs -text
+token_generation_model/_tp0_bk3/metaneff.pb filter=lfs diff=lfs merge=lfs -text
+token_generation_model/_tp0_bk3/model.MODULE_d3ed4857bd8baeff8023+b05cff0a.hlo_module.pb filter=lfs diff=lfs merge=lfs -text
+token_generation_model/_tp0_bk3/model.MODULE_d3ed4857bd8baeff8023+b05cff0a.neff filter=lfs diff=lfs merge=lfs -text

context_encoding_model/_tp0_bk0/command.txt

@@ -0,0 +1 @@
+neuronx-cc compile --framework=XLA model.MODULE_f4171003694760566af4+a9cd68fb.hlo_module.pb --output model.MODULE_f4171003694760566af4+a9cd68fb.neff --target=trn1 --auto-cast=none --model-type=transformer '--tensorizer-options=--enable-ccop-compute-overlap --cc-pipeline-tiling-factor=2 --vectorize-strided-dma ' --lnc=1 -O1 '--internal-hlo2tensorizer-options= --modular-flow-mac-threshold=10  --verify-hlo=true' --logfile=log-neuron-cc.txt --verbose=35

context_encoding_model/_tp0_bk0/compile_flags.MODULE_f4171003694760566af4+a9cd68fb.json

@@ -0,0 +1 @@
+["--target=trn1", "--auto-cast=none", "--model-type=transformer", "--tensorizer-options=--enable-ccop-compute-overlap --cc-pipeline-tiling-factor=2 --vectorize-strided-dma ", "--lnc=1", "-O1", "--internal-hlo2tensorizer-options= --modular-flow-mac-threshold=10  --verify-hlo=true", "--logfile=/home/ubuntu/qwen3/context_encoding_model/_tp0_bk0/log-neuron-cc.txt"]

context_encoding_model/_tp0_bk0/global_metric_store.json

@@ -0,0 +1,1079 @@
+{
+  "Average": {
+    "tensorizer": {
+      "StaticProfiler::AverageFractalPeUtilization": 99.65389251708984,
+      "StaticProfiler::AveragePartitionUtilization": 97.55139923095703,
+      "StaticProfiler::AveragePeUtilization": 98.60253143310547,
+      "StaticProfiler::LocalizationEfficiency": 99.04553985595703,
+      "StaticProfiler::LocalizationEfficiencyIgnoreNonlocal": 100.20111846923828,
+      "TilingProfiler::AveragePartitionUtilizationAfterTiling": 0.0,
+      "TilingProfiler::AveragePeUtilizationAfterTiling": 0.0
+    }
+  },
+  "Count": {
+    "tensorizer": {
+      "StaticProfiler::AverageFractalPeUtilization": 1.0,
+      "StaticProfiler::AveragePartitionUtilization": 1.0,
+      "StaticProfiler::AveragePeUtilization": 1.0,
+      "StaticProfiler::LocalizationEfficiency": 1.0,
+      "StaticProfiler::LocalizationEfficiencyIgnoreNonlocal": 1.0,
+      "TilingProfiler::AveragePartitionUtilizationAfterTiling": 1.0,
+      "TilingProfiler::AveragePeUtilizationAfterTiling": 1.0
+    }
+  },
+  "Sum": {
+    "compiletime": {
+      "AGOrderingAnalysisPass": 0.018787622451782227,
+      "AffinePredicateResolution": 0.0011818408966064453,
+      "AliasDependencyElimination": 0.00011801719665527344,
+      "AliasDependencyInduction": 0.005483388900756836,
+      "AliasDependencyReset": 0.026019811630249023,
+      "BFComputeCutting": 0.00225830078125,
+      "BirCodeGenLoop": 0.4621126651763916,
+      "CCOpFusion": 0.01928091049194336,
+      "CanonicalizeConv": 3.7000001611886546e-05,
+      "CanonicalizeDAGForPGTiling": 0.004612922668457031,
+      "CanonicalizeForTensorizer": 4.099999932805076e-05,
+      "CanonicalizeIR": 0.0017774105072021484,
+      "Canonicalizer": 0.0009619999909773469,
+      "CoalesceCCOp": 0.0146026611328125,
+      "CommuteConcat": 0.0020241737365722656,
+      "DMALocalityOpt": 0.005425214767456055,
+      "DMAProfiler": 0.012541055679321289,
+      "DMATilingProfiler": 0.004782676696777344,
+      "DataLocalityOpt": 0.06629562377929688,
+      "DataStreaming": 0.03773355484008789,
+      "DeConcat": 0.0006563663482666016,
+      "DeadCodeElimination": 0.002358675003051758,
+      "DeadStoreElimination": 0.0055620670318603516,
+      "DelinearIndices": 0.004741668701171875,
+      "Delinearization": 0.0036110877990722656,
+      "DoNothing": 0.00022459030151367188,
+      "DramToDramTranspose": 0.016016721725463867,
+      "DumpGraphAndMetadata": 0.0853111743927002,
+      "EliminateDivs": 0.0025675296783447266,
+      "ExpandBatchNorm": 0.002092123031616211,
+      "ExpandISAMacro": 0.011052370071411133,
+      "FactorizeBlkDims": 0.00814366340637207,
+      "FactorizeThreadAxesInFreeDims": 0.002122640609741211,
+      "FlattenMacroLoop": 0.002187013626098633,
+      "GenericAccessSimplifier": 0.0009529590606689453,
+      "HoistCompute": 6.000000212225132e-06,
+      "IdentifyCrossPassTensors": 7.700000423938036e-05,
+      "InferInitValue": 0.0242159366607666,
+      "InferIntrinsicOnCC": 0.009269952774047852,
+      "InferNeuronTensor": 0.020155906677246094,
+      "InferNonlocalTensors": 0.015646696090698242,
+      "InferPSumTensor": 0.3081786632537842,
+      "InlineNativeKernels": 0.009155511856079102,
+      "InsertIOTransposes": 0.015281438827514648,
+      "InsertLocalTransposes": 0.006501436233520508,
+      "InsertOffloadedTransposes": 0.002702474594116211,
+      "LICM": 0.002913951873779297,
+      "LateLegalizeInst": 0.014158487319946289,
+      "LateLegalizePostSplit": 0.012693405151367188,
+      "LateLowerReshapeOp": 0.0025734901428222656,
+      "LateLowerTensorOp": 0.001531362533569336,
+      "LateNeuronInstComb": 0.008838176727294922,
+      "LayoutPreprocessing": 0.026634931564331055,
+      "LayoutPreprocessingAndAnalysis": 0.5595176219940186,
+      "LayoutRequirementAnalysis": 0.005538463592529297,
+      "LegalizeCCOpLayout": 0.0022728443145751953,
+      "LegalizeOpLevelAlias": 0.001255035400390625,
+      "LegalizePartitionReduce": 0.001256704330444336,
+      "LegalizeSundaAccess": 0.07711672782897949,
+      "LegalizeSundaMacro": 0.010920286178588867,
+      "LegalizeType": 0.01314401626586914,
+      "LocalLayoutOpt": 0.012011289596557617,
+      "LoopFusion": 0.006572723388671875,
+      "LoopSplitting": 0.0003001689910888672,
+      "LowerBroadcast": 0.0018808841705322266,
+      "LowerCCOpBlockAxis": 0.0050678253173828125,
+      "LowerComplexBroadcast": 0.0025262832641601563,
+      "LowerIntrinsics": 0.3039369583129883,
+      "LowerTensorOp": 0.011744022369384766,
+      "LowerTranspose": 0.011741399765014648,
+      "MacroGeneration": 0.026911020278930664,
+      "MaskPropagation": 0.0031325817108154297,
+      "MemcastMotion": 2.2000000171829015e-05,
+      "MemcpyElimination": 0.027472257614135742,
+      "MutateDataType": 0.0015196800231933594,
+      "NeuronAliasDependencyInduction": 0.00016927719116210938,
+      "NeuronAliasDependencyReset": 0.0242006778717041,
+      "NeuronInstComb": 0.00468754768371582,
+      "NeuronLICM": 0.03664875030517578,
+      "NeuronLoopFusion": 0.00889277458190918,
+      "NeuronLoopInterchange": 0.002141237258911133,
+      "NeuronSimplifier": 0.00720524787902832,
+      "NeuronSimplifyPredicates": 0.12209796905517578,
+      "NeuronValueNumbering": 0.003449678421020508,
+      "OptimizeAliasedCopyChain": 0.0006387233734130859,
+      "OptimizeNKIKernels": 0.5260024070739746,
+      "PAGLayoutOpt": 0.5680239200592041,
+      "PComputeCutting": 0.0048143863677978516,
+      "PGLayoutTilingPipeline": 1.6304676532745361,
+      "PGTiling": 0.1616363525390625,
+      "PadElimination": 0.0003521442413330078,
+      "ParAxesAnnotation": 0.0544736385345459,
+      "PartialLoopFusion": 0.005907773971557617,
+      "PartialSimdFusion": 0.0038967132568359375,
+      "PenguinizeFunctions": 3.900000228895806e-05,
+      "PerfectLoopNest": 0.0021576881408691406,
+      "PruneFunctions": 3.5000000934815034e-05,
+      "RecognizeOpIdiom": 0.0039520263671875,
+      "Recompute": 0.0002884864807128906,
+      "RelaxPredicates": 0.013870716094970703,
+      "Rematerialization": 0.0024657249450683594,
+      "RemoveOptimizationBarriers": 6.500000017695129e-05,
+      "ReshapeWeights": 0.0006930828094482422,
+      "ResolveAccessConflict": 0.0038983821868896484,
+      "ResolveComplicatePredicates": 0.0012950897216796875,
+      "RewriteReplicationMatmul": 0.002060413360595703,
+      "RewriteWeights": 0.0028791427612304688,
+      "SFKVectorizer": 0.2904393672943115,
+      "ScatterMotion": 2.8000000384054147e-05,
+      "SimpleAllReduceTiling": 0.008909463882446289,
+      "Simplifier": 0.003449678421020508,
+      "SimplifyMacroPredicates": 0.010317325592041016,
+      "SimplifyNeuronTensor": 1.038323163986206,
+      "SimplifySlice": 0.0008852481842041016,
+      "SimplifyTensor": 0.005218982696533203,
+      "SpillPSum": 0.010073423385620117,
+      "SplitAPUnionSets": 0.10591006278991699,
+      "SplitAccGrp": 0.0011169910430908203,
+      "StaticProfiler": 0.01290583610534668,
+      "StaticTransposeLocalTensor": 0.003824472427368164,
+      "SundaISel": 0.041872262954711914,
+      "TCTransform": 0.0008666515350341797,
+      "TensorInitialization": 0.013058185577392578,
+      "TensorOpSimplifier": 0.0061550140380859375,
+      "TensorOpTransform": 0.020328521728515625,
+      "TensorizerLegalizationPass": 6.900000153109431e-05,
+      "TileCCOps": 0.006834983825683594,
+      "TilingProfiler": 0.0072863101959228516,
+      "TransformConvOp": 0.0032320022583007813,
+      "TritiumFusion": 0.03062152862548828,
+      "ValueNumbering": 0.0023603439331054688,
+      "VectorizeDMA": 0.004430294036865234,
+      "VectorizeMatMult": 0.0021605491638183594,
+      "VerifySupportedOps": 3.300000025774352e-05,
+      "WeightCoalescing": 0.00846409797668457,
+      "ZeroSizeTensorElimination": 0.00011014938354492188,
+      "algsimp": 0.004399999976158142,
+      "batchnorm_expander": 3.600000127335079e-05,
+      "boundary-marker-removal": 1.2000000424450263e-05,
+      "call-inliner": 0.0007670000777579844,
+      "canonicalize-boundary-marker": 1.4999999621068127e-05,
+      "collective-stream-id-checker": 7.300000288523734e-05,
+      "comparison-expander": 0.0006099999882280827,
+      "computation-deduplicator": 5.8999998145736754e-05,
+      "conditional-to-select": 1.5999999959603883e-05,
+      "config-lowering": 8.70000003487803e-05,
+      "constant-statistics": 0.0005649999948218465,
+      "constant_folding": 0.0005520000122487545,
+      "cse": 3.600000127335079e-05,
+      "dce": 0.00014599999121855944,
+      "dot_decomposer": 0.0013859999598935246,
+      "dynamic-slice-transpose": 1.2000000424450263e-05,
+      "eliminate-redundant-compare": 0.0004949999856762588,
+      "emit-offloaded-dropout": 3.80000019504223e-05,
+      "flatten-call-graph": 0.0009339999523945153,
+      "fuse-send-recv": 7.100000220816582e-05,
+      "hilo::LegalizeAlias": 1.1999999514955562e-05,
+      "hilo::NeuronInstCombine": 0.00010099999781232327,
+      "hilo::NeuronOpFusion": 4.7999998059822246e-05,
+      "hilo::ReplaceTokenTypeWithU8Pass": 3.899999865097925e-05,
+      "hilo::ScheduleFusion": 1.9999999949504854e-06,
+      "hilo::SixtyFourHack": 6.200000643730164e-05,
+      "hilo::VerifyAliasing": 4.999999873689376e-06,
+      "hlo-mac-count": 0.0018479999853298068,
+      "hlo-verifier": 0.007563999388366938,
+      "instruction-histogram": 0.002553999889642,
+      "io-con-pipe-begin": 4.999999873689376e-06,
+      "io-con-pipe-end": 9.999999974752427e-07,
+      "io-layout-normalization": 0.0013040000339969993,
+      "io-statistics": 6.500000017695129e-05,
+      "legalize-ccops": 3.000000106112566e-06,
+      "legalize-compare": 1.1000000085914508e-05,
+      "lower-argminmax-custom-call": 1.2000000424450263e-05,
+      "map-inline": 0.0008759999764151871,
+      "metadata-naming": 6.0999998822808266e-05,
+      "mlir::detail::OpToOpPassAdaptor": 7.200000254670158e-05,
+      "mlir::hlo::MhloToPyPenguin": 0.002776999957859516,
+      "mlir::mhlo::LowerComplexExtraPass": 0.00023499999952036887,
+      "mlir::mhlo::LowerComplexPass": 0.00032500000088475645,
+      "native-to-custom-softmax": 0.0007319999858736992,
+      "native-to-custom-softmax-dx": 0.000678999989759177,
+      "operand_upcaster": 4.900000203633681e-05,
+      "opt-barrier-removal": 0.0005629999795928597,
+      "post-par-pipe-begin": 9.000000318337698e-06,
+      "post-par-pipe-end": 0.0,
+      "post-partition-simplification": 0.001663000090047717,
+      "pre-par-pipe-begin": 9.999999974752427e-07,
+      "pre-par-pipe-end": 9.999999974752427e-07,
+      "pre-partition-simplification": 0.13888800144195557,
+      "replace-minimum-constant": 0.0007169999880716205,
+      "reshape-mover": 0.00021499999274965376,
+      "simplify-concat": 0.00014099999680183828,
+      "simplify-while-loops": 0.00017800000205170363,
+      "transform-variadic-reduce": 6.70000008540228e-05,
+      "tuple-simplifier": 0.0005469999159686267,
+      "unpack-nested-aws-ntwsr": 0.00046300000394694507,
+      "unroll-while-loop": 3.099999958067201e-05,
+      "zero_sized_hlo_elimination": 0.0008880000095814466
+    },
+    "hilo": {
+      "ConstantSize": 304437.0,
+      "HloInputCount": 475.0,
+      "HloMacCount": 25141444608.0,
+      "HloOutputCount": 73.0,
+      "IfmapSize": 8266542080.0,
+      "OfmapSize": 75497472.0,
+      "OutputsReadFromCount": 0.0,
+      "PassthroughTensorsCount": 0.0,
+      "RedundantOutputCount": 0.0,
+      "Traffic": 1649111936.0
+    },
+    "tensorizer": {
+      "DMATilingProfiler::TotalInstructionsAfterTiling": 42834.0,
+      "StaticProfiler::AifUb": 129.43267822265625,
+      "StaticProfiler::ArithmeticIntensityTensorizer": 128.19729614257813,
+      "StaticProfiler::AverageDmaLength": 4810.17578125,
+      "StaticProfiler::DDRTransferBytes": 782946624.0,
+      "StaticProfiler::InternalTransferBytes": 629086720.0,
+      "StaticProfiler::LoadExpanded": 97814.0,
+      "StaticProfiler::StoreExpanded": 1757.0,
+      "StaticProfiler::TotalDMAExpanded": 99571.0,
+      "StaticProfiler::TotalDynamicInstancesCount": 50031.0,
+      "StaticProfiler::TotalDynamicInstancesWithMmPackedCount": 49585.0,
+      "StaticProfiler::TotalLNCComm": 0.0,
+      "StaticProfiler::TotalLNCCommTransfer": 0.0,
+      "TilingProfiler::BatchnormInstructionsAfterTiling": 0.0,
+      "TilingProfiler::DmaInstructionsAfterTiling": 0.0,
+      "TilingProfiler::GenericInstructionsAfterTiling": 4.0,
+      "TilingProfiler::MatMultInstructionsAfterTiling": 22464.0,
+      "TilingProfiler::NumPfTransposes": 5.0,
+      "TilingProfiler::NumPfTransposesForIo": 1.0,
+      "TilingProfiler::NumPfTransposesForLocal": 1.0,
+      "TilingProfiler::NumPfTransposesForNonlocal": 3.0,
+      "TilingProfiler::PfTransposeInstructions": 19105.0,
+      "TilingProfiler::PfTransposeInstructionsForIo": 19008.0,
+      "TilingProfiler::PfTransposeInstructionsForLocal": 1.0,
+      "TilingProfiler::PfTransposeInstructionsForNonlocal": 96.0,
+      "TilingProfiler::ReduceInstructionsAfterTiling": 4.0,
+      "TilingProfiler::SimdInstructionsAfterTiling": 158.0,
+      "TilingProfiler::TotalInstructionsAfterTiling": 0.0,
+      "TransformConvOp::Conv1d_depthwise_bf01_oi01_bf01": 0.0,
+      "TransformConvOp::Conv2d_dw_fb01_io01_01bf_rep_nhwc_Pcinh": 0.0,
+      "TransformConvOp::Conv2d_pbp_0f1b_0i1o_01fb_experimental_1": 0.0,
+      "TransformConvOp::Conv2d_pbp_fb01_io01_01bf_experimental_1": 0.0,
+      "TransformConvOp::conv2d_column_packing": 0.0,
+      "TransformConvOp::conv2d_column_packing_1": 0.0,
+      "TransformConvOp::conv2d_column_packing_io10": 0.0,
+      "TransformConvOp::conv2d_depthwise_f01b_o01i_bf01": 0.0
+    }
+  },
+  "all": {
+    "compiletime": {
+      "algsimp": 0.004207999911159277,
+      "call-inliner": 0.0007350000087171793,
+      "collective-stream-id-checker": 6.399999983841553e-05,
+      "comparison-expander": 0.0005949999904260039,
+      "constant-statistics": 0.0005649999948218465,
+      "constant_folding": 0.0005249999812804163,
+      "dce": 0.0001429999974789098,
+      "dot_decomposer": 0.0013859999598935246,
+      "eliminate-redundant-compare": 0.0004839999892283231,
+      "flatten-call-graph": 0.000901999999769032,
+      "hlo-mac-count": 0.0016270000487565994,
+      "hlo-verifier": 0.007029999978840351,
+      "instruction-histogram": 0.002553999889642,
+      "io-con-pipe-begin": 4.999999873689376e-06,
+      "io-con-pipe-end": 9.999999974752427e-07,
+      "io-layout-normalization": 0.0013040000339969993,
+      "io-statistics": 6.500000017695129e-05,
+      "map-inline": 0.000838999985717237,
+      "native-to-custom-softmax": 0.0007050000131130219,
+      "native-to-custom-softmax-dx": 0.0005089999758638442,
+      "opt-barrier-removal": 0.0005629999795928597,
+      "pre-par-pipe-begin": 9.999999974752427e-07,
+      "pre-par-pipe-end": 9.999999974752427e-07,
+      "pre-partition-simplification": 0.13888800144195557,
+      "replace-minimum-constant": 0.0006949999951757491,
+      "reshape-mover": 0.00020500000391621143,
+      "simplify-while-loops": 0.0001720000000204891,
+      "tuple-simplifier": 0.0005319999763742089,
+      "unpack-nested-aws-ntwsr": 0.000450999999884516,
+      "unroll-while-loop": 2.9000000722589903e-05,
+      "zero_sized_hlo_elimination": 0.0008880000095814466
+    }
+  },
+  "cumsum": {
+    "compiletime": {
+      "CoalesceCCOp": 0.00020933151245117188,
+      "DMALocalityOpt": 0.0001666545867919922,
+      "DMAProfiler": 0.0008401870727539063,
+      "DataStreaming": 0.0002658367156982422,
+      "DoNothing": 0.00014090538024902344,
+      "ExpandISAMacro": 0.0004999637603759766,
+      "FactorizeBlkDims": 0.00046062469482421875,
+      "InferPSumTensor": 0.0004820823669433594,
+      "LateLegalizeInst": 0.0004343986511230469,
+      "LateNeuronInstComb": 0.0004832744598388672,
+      "LegalizeSundaAccess": 0.002238750457763672,
+      "LegalizeType": 0.0002429485321044922,
+      "LowerBroadcast": 0.0002453327178955078,
+      "LowerIntrinsics": 0.00021791458129882813,
+      "LowerTranspose": 0.00022292137145996094,
+      "NeuronInstComb": 0.0005400180816650391,
+      "NeuronLICM": 0.0003840923309326172,
+      "NeuronSimplifyPredicates": 0.0028014183044433594,
+      "NeuronValueNumbering": 0.00042724609375,
+      "SFKVectorizer": 0.0028204917907714844,
+      "SimpleAllReduceTiling": 0.0002048015594482422,
+      "SimplifyNeuronTensor": 0.00043082237243652344,
+      "SpillPSum": 0.0005221366882324219,
+      "WeightCoalescing": 0.00020456314086914063
+    }
+  },
+  "sg00": {
+    "compiletime": {
+      "CanonicalizeConv": 1.4000000192027073e-05,
+      "CanonicalizeForTensorizer": 1.4999999621068127e-05,
+      "Canonicalizer": 0.0003440000000409782,
+      "HoistCompute": 3.000000106112566e-06,
+      "IdentifyCrossPassTensors": 3.099999958067201e-05,
+      "MemcastMotion": 1.2000000424450263e-05,
+      "PenguinizeFunctions": 1.5999999959603883e-05,
+      "PruneFunctions": 1.2999999853491317e-05,
+      "RemoveOptimizationBarriers": 2.4000000848900527e-05,
+      "ScatterMotion": 1.1000000085914508e-05,
+      "TensorizerLegalizationPass": 2.9000000722589903e-05,
+      "VerifySupportedOps": 1.1000000085914508e-05,
+      "algsimp": 6.70000008540228e-05,
+      "batchnorm_expander": 1.2999999853491317e-05,
+      "boundary-marker-removal": 3.999999989900971e-06,
+      "call-inliner": 1.1000000085914508e-05,
+      "canonicalize-boundary-marker": 6.000000212225132e-06,
+      "collective-stream-id-checker": 3.000000106112566e-06,
+      "comparison-expander": 4.999999873689376e-06,
+      "computation-deduplicator": 1.700000029813964e-05,
+      "conditional-to-select": 4.999999873689376e-06,
+      "config-lowering": 3.199999991920777e-05,
+      "constant_folding": 9.000000318337698e-06,
+      "cse": 1.2999999853491317e-05,
+      "dce": 9.999999974752427e-07,
+      "dynamic-slice-transpose": 3.999999989900971e-06,
+      "eliminate-redundant-compare": 3.999999989900971e-06,
+      "emit-offloaded-dropout": 1.2999999853491317e-05,
+      "flatten-call-graph": 1.1000000085914508e-05,
+      "fuse-send-recv": 2.4000000848900527e-05,
+      "hilo::LegalizeAlias": 4.999999873689376e-06,
+      "hilo::NeuronInstCombine": 6.0999998822808266e-05,
+      "hilo::NeuronOpFusion": 7.000000096013537e-06,
+      "hilo::ReplaceTokenTypeWithU8Pass": 1.5999999959603883e-05,
+      "hilo::ScheduleFusion": 9.999999974752427e-07,
+      "hilo::SixtyFourHack": 1.4999999621068127e-05,
+      "hilo::VerifyAliasing": 1.9999999949504854e-06,
+      "hlo-mac-count": 2.5999999706982635e-05,
+      "hlo-verifier": 0.00018699999782256782,
+      "legalize-ccops": 9.999999974752427e-07,
+      "legalize-compare": 3.999999989900971e-06,
+      "lower-argminmax-custom-call": 3.999999989900971e-06,
+      "map-inline": 1.2000000424450263e-05,
+      "metadata-naming": 2.099999983329326e-05,
+      "mlir::detail::OpToOpPassAdaptor": 2.300000051036477e-05,
+      "mlir::hlo::MhloToPyPenguin": 0.0010349999647587538,
+      "mlir::mhlo::LowerComplexExtraPass": 8.600000001024455e-05,
+      "mlir::mhlo::LowerComplexPass": 0.0001740000006975606,
+      "native-to-custom-softmax": 1.4000000192027073e-05,
+      "native-to-custom-softmax-dx": 0.0001340000017080456,
+      "operand_upcaster": 1.8000000636675395e-05,
+      "post-par-pipe-begin": 6.000000212225132e-06,
+      "post-par-pipe-end": 0.0,
+      "post-partition-simplification": 0.0005830000154674053,
+      "replace-minimum-constant": 7.000000096013537e-06,
+      "reshape-mover": 3.999999989900971e-06,
+      "simplify-concat": 4.8000001697801054e-05,
+      "simplify-while-loops": 1.9999999949504854e-06,
+      "transform-variadic-reduce": 9.000000318337698e-06,
+      "tuple-simplifier": 4.999999873689376e-06,
+      "unpack-nested-aws-ntwsr": 3.999999989900971e-06,
+      "unroll-while-loop": 0.0
+    },
+    "hilo": {
+      "ArithmeticIntensity": 8.198826789855957,
+      "ConstantSize": 304437.0,
+      "HloInputCount": 475.0,
+      "HloMacCount": 2751463424.0,
+      "HloOutputCount": 73.0,
+      "IfmapSize": 8266542080.0,
+      "OfmapSize": 75497472.0,
+      "OutputsReadFromCount": 0.0,
+      "PassthroughTensorsCount": 0.0,
+      "RedundantOutputCount": 0.0,
+      "Traffic": 671184704.0
+    }
+  },
+  "sg0000": {
+    "compiletime": {
+      "AGOrderingAnalysisPass": 0.04074835777282715,
+      "AffinePredicateResolution": 0.002183198928833008,
+      "AliasDependencyElimination": 0.00012922286987304688,
+      "AliasDependencyInduction": 0.008634567260742188,
+      "AliasDependencyReset": 0.03679013252258301,
+      "BFComputeCutting": 0.0019538402557373047,
+      "BirCodeGenLoop": 0.04571366310119629,
+      "CCOpFusion": 0.01575756072998047,
+      "CanonicalizeDAGForPGTiling": 0.003149271011352539,
+      "CanonicalizeIR": 0.002719879150390625,
+      "CoalesceCCOp": 0.0047032833099365234,
+      "CommuteConcat": 0.0013585090637207031,
+      "DMALocalityOpt": 0.001116037368774414,
+      "DMAProfiler": 0.0047032833099365234,
+      "DMATilingProfiler": 0.004144191741943359,
+      "DataLocalityOpt": 0.10100674629211426,
+      "DataStreaming": 0.0033788681030273438,
+      "DeConcat": 0.0007069110870361328,
+      "DeadCodeElimination": 0.0010058879852294922,
+      "DeadStoreElimination": 0.031080961227416992,
+      "DelinearIndices": 0.007829427719116211,
+      "Delinearization": 0.003365039825439453,
+      "DoNothing": 7.033348083496094e-05,
+      "DramToDramTranspose": 0.024500370025634766,
+      "DumpGraphAndMetadata": 0.005262136459350586,
+      "EliminateDivs": 0.005412578582763672,
+      "ExpandBatchNorm": 0.0019643306732177734,
+      "ExpandISAMacro": 0.002582550048828125,
+      "FactorizeBlkDims": 0.00794839859008789,
+      "FactorizeThreadAxesInFreeDims": 0.0020449161529541016,
+      "FlattenMacroLoop": 0.0028934478759765625,
+      "GenericAccessSimplifier": 0.0009298324584960938,
+      "InferInitValue": 0.026146411895751953,
+      "InferIntrinsicOnCC": 0.010050058364868164,
+      "InferNeuronTensor": 0.03407764434814453,
+      "InferNonlocalTensors": 0.06189298629760742,
+      "InferPSumTensor": 0.03060150146484375,
+      "InlineNativeKernels": 0.0014431476593017578,
+      "InsertIOTransposes": 0.009805679321289063,
+      "InsertLocalTransposes": 0.007609128952026367,
+      "InsertOffloadedTransposes": 0.004189968109130859,
+      "LICM": 0.0029850006103515625,
+      "LateLegalizeInst": 0.004921674728393555,
+      "LateLegalizePostSplit": 0.0025641918182373047,
+      "LateLowerReshapeOp": 0.002185821533203125,
+      "LateLowerTensorOp": 0.00531768798828125,
+      "LateNeuronInstComb": 0.02812671661376953,
+      "LayoutPreprocessing": 0.11982965469360352,
+      "LayoutPreprocessingAndAnalysis": 0.24928760528564453,
+      "LayoutRequirementAnalysis": 0.007187366485595703,
+      "LegalizeCCOpLayout": 0.0035941600799560547,
+      "LegalizeOpLevelAlias": 0.0022826194763183594,
+      "LegalizePartitionReduce": 0.002084970474243164,
+      "LegalizeSundaAccess": 0.03499269485473633,
+      "LegalizeSundaMacro": 0.00858449935913086,
+      "LegalizeType": 0.0038924217224121094,
+      "LocalLayoutOpt": 0.015146255493164063,
+      "LoopFusion": 0.00600433349609375,
+      "LoopSplitting": 0.0003192424774169922,
+      "LowerBroadcast": 0.0030934810638427734,
+      "LowerCCOpBlockAxis": 0.0053822994232177734,
+      "LowerComplexBroadcast": 0.0017805099487304688,
+      "LowerIntrinsics": 0.03145861625671387,
+      "LowerTensorOp": 0.013553142547607422,
+      "LowerTranspose": 0.008147954940795898,
+      "MacroGeneration": 0.10158348083496094,
+      "MaskPropagation": 0.004988193511962891,
+      "MemcpyElimination": 0.1091456413269043,
+      "MutateDataType": 0.002095937728881836,
+      "NeuronAliasDependencyInduction": 0.00023055076599121094,
+      "NeuronAliasDependencyReset": 0.036977291107177734,
+      "NeuronInstComb": 0.01214146614074707,
+      "NeuronLICM": 0.007807016372680664,
+      "NeuronLoopFusion": 0.014447927474975586,
+      "NeuronLoopInterchange": 0.0015079975128173828,
+      "NeuronSimplifier": 0.009031057357788086,
+      "NeuronSimplifyPredicates": 0.0026018619537353516,
+      "NeuronValueNumbering": 0.00443577766418457,
+      "OptimizeAliasedCopyChain": 0.0012700557708740234,
+      "OptimizeNKIKernels": 0.00177764892578125,
+      "PAGLayoutOpt": 0.3914484977722168,
+      "PComputeCutting": 0.005900144577026367,
+      "PGLayoutTilingPipeline": 1.2139532566070557,
+      "PGTiling": 0.2603449821472168,
+      "PadElimination": 0.00040340423583984375,
+      "ParAxesAnnotation": 0.2578258514404297,
+      "PartialLoopFusion": 0.010677099227905273,
+      "PartialSimdFusion": 0.011437177658081055,
+      "PerfectLoopNest": 0.001963376998901367,
+      "RecognizeOpIdiom": 0.004378318786621094,
+      "Recompute": 0.0002574920654296875,
+      "RelaxPredicates": 0.003600597381591797,
+      "Rematerialization": 0.004474163055419922,
+      "ReshapeWeights": 0.0006759166717529297,
+      "ResolveAccessConflict": 0.003798246383666992,
+      "ResolveComplicatePredicates": 0.002101421356201172,
+      "RewriteReplicationMatmul": 0.0012481212615966797,
+      "RewriteWeights": 0.004036903381347656,
+      "SFKVectorizer": 0.09602093696594238,
+      "SimpleAllReduceTiling": 0.0017740726470947266,
+      "Simplifier": 0.004450559616088867,
+      "SimplifyMacroPredicates": 0.010053157806396484,
+      "SimplifyNeuronTensor": 0.00724029541015625,
+      "SimplifySlice": 0.001153707504272461,
+      "SimplifyTensor": 0.005860805511474609,
+      "SpillPSum": 0.011501789093017578,
+      "SplitAPUnionSets": 0.03104996681213379,
+      "SplitAccGrp": 0.002181529998779297,
+      "StaticProfiler": 0.004481792449951172,
+      "StaticTransposeLocalTensor": 0.006117343902587891,
+      "SundaISel": 0.041422128677368164,
+      "TCTransform": 0.0022428035736083984,
+      "TensorInitialization": 0.00680994987487793,
+      "TensorOpSimplifier": 0.008346796035766602,
+      "TensorOpTransform": 0.030104398727416992,
+      "TileCCOps": 0.005553245544433594,
+      "TilingProfiler": 0.009899139404296875,
+      "TransformConvOp": 0.0027108192443847656,
+      "TritiumFusion": 0.020798206329345703,
+      "ValueNumbering": 0.003211498260498047,
+      "VectorizeDMA": 0.004341602325439453,
+      "VectorizeMatMult": 0.0021800994873046875,
+      "WeightCoalescing": 0.0030617713928222656,
+      "ZeroSizeTensorElimination": 0.00011968612670898438
+    },
+    "tensorizer": {
+      "DMATilingProfiler::TotalInstructionsAfterTiling": 1396.0,
+      "StaticProfiler::AifUb": 8.992382049560547,
+      "StaticProfiler::ArithmeticIntensityTensorizer": 75.54261016845703,
+      "StaticProfiler::AverageDmaLength": 9594.294921875,
+      "StaticProfiler::AverageFractalPeUtilization": 99.893310546875,
+      "StaticProfiler::AveragePartitionUtilization": 94.61784362792969,
+      "StaticProfiler::AveragePeUtilization": 99.893310546875,
+      "StaticProfiler::DDRTransferBytes": 79837440.0,
+      "StaticProfiler::InternalTransferBytes": 9797632.0,
+      "StaticProfiler::LoadExpanded": 11010.0,
+      "StaticProfiler::LocalizationEfficiency": 840.0734252929688,
+      "StaticProfiler::LocalizationEfficiencyIgnoreNonlocal": 1527.2347412109375,
+      "StaticProfiler::StoreExpanded": 3073.0,
+      "StaticProfiler::TotalDMAExpanded": 14083.0,
+      "StaticProfiler::TotalDynamicInstancesCount": 1442.0,
+      "StaticProfiler::TotalDynamicInstancesWithMmPackedCount": 1442.0,
+      "StaticProfiler::TotalLNCComm": 0.0,
+      "StaticProfiler::TotalLNCCommTransfer": 0.0,
+      "TilingProfiler::AveragePartitionUtilizationAfterTiling": 0.0,
+      "TilingProfiler::AveragePeUtilizationAfterTiling": 0.0,
+      "TilingProfiler::BatchnormInstructionsAfterTiling": 0.0,
+      "TilingProfiler::DmaInstructionsAfterTiling": 0.0,
+      "TilingProfiler::GenericInstructionsAfterTiling": 12.0,
+      "TilingProfiler::MatMultInstructionsAfterTiling": 821.0,
+      "TilingProfiler::NumPfTransposes": 6.0,
+      "TilingProfiler::NumPfTransposesForIo": 0.0,
+      "TilingProfiler::NumPfTransposesForLocal": 5.0,
+      "TilingProfiler::NumPfTransposesForNonlocal": 1.0,
+      "TilingProfiler::PfTransposeInstructions": 72.0,
+      "TilingProfiler::PfTransposeInstructionsForIo": 0.0,
+      "TilingProfiler::PfTransposeInstructionsForLocal": 56.0,
+      "TilingProfiler::PfTransposeInstructionsForNonlocal": 16.0,
+      "TilingProfiler::ReduceInstructionsAfterTiling": 0.0,
+      "TilingProfiler::SimdInstructionsAfterTiling": 101.0,
+      "TilingProfiler::TotalInstructionsAfterTiling": 0.0,
+      "TransformConvOp::Conv1d_depthwise_bf01_oi01_bf01": 0.0,
+      "TransformConvOp::Conv2d_dw_fb01_io01_01bf_rep_nhwc_Pcinh": 0.0,
+      "TransformConvOp::Conv2d_pbp_0f1b_0i1o_01fb_experimental_1": 0.0,
+      "TransformConvOp::Conv2d_pbp_fb01_io01_01bf_experimental_1": 0.0,
+      "TransformConvOp::conv2d_column_packing": 0.0,
+      "TransformConvOp::conv2d_column_packing_1": 0.0,
+      "TransformConvOp::conv2d_column_packing_io10": 0.0,
+      "TransformConvOp::conv2d_depthwise_f01b_o01i_bf01": 0.0
+    }
+  },
+  "sg0001": {
+    "compiletime": {
+      "AGOrderingAnalysisPass": 0.034732818603515625,
+      "AffinePredicateResolution": 0.0015087127685546875,
+      "AliasDependencyElimination": 0.0001227855682373047,
+      "AliasDependencyInduction": 0.008313894271850586,
+      "AliasDependencyReset": 0.044220924377441406,
+      "BFComputeCutting": 0.001974344253540039,
+      "BirCodeGenLoop": 0.03118896484375,
+      "CCOpFusion": 0.018246889114379883,
+      "CanonicalizeDAGForPGTiling": 0.003057718276977539,
+      "CanonicalizeIR": 0.0027036666870117188,
+      "CoalesceCCOp": 0.0046405792236328125,
+      "CommuteConcat": 0.0015790462493896484,
+      "DMALocalityOpt": 0.0015497207641601563,
+      "DMAProfiler": 0.004349708557128906,
+      "DMATilingProfiler": 0.003928422927856445,
+      "DataLocalityOpt": 0.12123703956604004,
+      "DataStreaming": 0.0025773048400878906,
+      "DeConcat": 0.0008485317230224609,
+      "DeadCodeElimination": 0.0012981891632080078,
+      "DeadStoreElimination": 0.034687042236328125,
+      "DelinearIndices": 0.009628772735595703,
+      "Delinearization": 0.003772258758544922,
+      "DoNothing": 7.009506225585938e-05,
+      "DramToDramTranspose": 0.028621673583984375,
+      "DumpGraphAndMetadata": 0.003651142120361328,
+      "EliminateDivs": 0.004262447357177734,
+      "ExpandBatchNorm": 0.002134084701538086,
+      "ExpandISAMacro": 0.0026290416717529297,
+      "FactorizeBlkDims": 0.009716033935546875,
+      "FactorizeThreadAxesInFreeDims": 0.0013210773468017578,
+      "FlattenMacroLoop": 0.002851247787475586,
+      "GenericAccessSimplifier": 0.002216815948486328,
+      "InferInitValue": 0.03134632110595703,
+      "InferIntrinsicOnCC": 0.011671781539916992,
+      "InferNeuronTensor": 0.039717674255371094,
+      "InferNonlocalTensors": 0.030872583389282227,
+      "InferPSumTensor": 0.022834062576293945,
+      "InlineNativeKernels": 0.0021605491638183594,
+      "InsertIOTransposes": 0.017906904220581055,
+      "InsertLocalTransposes": 0.007941961288452148,
+      "InsertOffloadedTransposes": 0.0032515525817871094,
+      "LICM": 0.003479480743408203,
+      "LateLegalizeInst": 0.003596782684326172,
+      "LateLegalizePostSplit": 0.002257108688354492,
+      "LateLowerReshapeOp": 0.0018393993377685547,
+      "LateLowerTensorOp": 0.005475044250488281,
+      "LateNeuronInstComb": 0.017774581909179688,
+      "LayoutPreprocessing": 0.03530263900756836,
+      "LayoutPreprocessingAndAnalysis": 0.11916303634643555,
+      "LayoutRequirementAnalysis": 0.007796525955200195,
+      "LegalizeCCOpLayout": 0.0019328594207763672,
+      "LegalizeOpLevelAlias": 0.001219034194946289,
+      "LegalizePartitionReduce": 0.0009839534759521484,
+      "LegalizeSundaAccess": 0.015137434005737305,
+      "LegalizeSundaMacro": 0.010521173477172852,
+      "LegalizeType": 0.004090547561645508,
+      "LocalLayoutOpt": 0.020325422286987305,
+      "LoopFusion": 0.006730556488037109,
+      "LoopSplitting": 0.00034809112548828125,
+      "LowerBroadcast": 0.001789093017578125,
+      "LowerCCOpBlockAxis": 0.005074977874755859,
+      "LowerComplexBroadcast": 0.0019309520721435547,
+      "LowerIntrinsics": 0.03209352493286133,
+      "LowerTensorOp": 0.012279510498046875,
+      "LowerTranspose": 0.010157585144042969,
+      "MacroGeneration": 0.09246373176574707,
+      "MaskPropagation": 0.003335237503051758,
+      "MemcpyElimination": 0.10414385795593262,
+      "MutateDataType": 0.00220489501953125,
+      "NeuronAliasDependencyInduction": 0.0002532005310058594,
+      "NeuronAliasDependencyReset": 0.03873252868652344,
+      "NeuronInstComb": 0.012767791748046875,
+      "NeuronLICM": 0.006428241729736328,
+      "NeuronLoopFusion": 0.01547694206237793,
+      "NeuronLoopInterchange": 0.0012590885162353516,
+      "NeuronSimplifier": 0.009620428085327148,
+      "NeuronSimplifyPredicates": 0.0022652149200439453,
+      "NeuronValueNumbering": 0.0031261444091796875,
+      "OptimizeAliasedCopyChain": 0.0007045269012451172,
+      "OptimizeNKIKernels": 0.0022683143615722656,
+      "PAGLayoutOpt": 0.11684298515319824,
+      "PComputeCutting": 0.0060575008392333984,
+      "PGLayoutTilingPipeline": 1.5194215774536133,
+      "PGTiling": 0.5792257785797119,
+      "PadElimination": 0.0004138946533203125,
+      "ParAxesAnnotation": 0.08577656745910645,
+      "PartialLoopFusion": 0.010853052139282227,
+      "PartialSimdFusion": 0.010831356048583984,
+      "PerfectLoopNest": 0.0021359920501708984,
+      "RecognizeOpIdiom": 0.004781246185302734,
+      "Recompute": 0.00029349327087402344,
+      "RelaxPredicates": 0.0031125545501708984,
+      "Rematerialization": 0.002535104751586914,
+      "ReshapeWeights": 0.0007915496826171875,
+      "ResolveAccessConflict": 0.004204988479614258,
+      "ResolveComplicatePredicates": 0.0014605522155761719,
+      "RewriteReplicationMatmul": 0.0014035701751708984,
+      "RewriteWeights": 0.0033304691314697266,
+      "SFKVectorizer": 0.11060166358947754,
+      "SimpleAllReduceTiling": 0.0013706684112548828,
+      "Simplifier": 0.004431247711181641,
+      "SimplifyMacroPredicates": 0.005709409713745117,
+      "SimplifyNeuronTensor": 0.005321979522705078,
+      "SimplifySlice": 0.0020780563354492188,
+      "SimplifyTensor": 0.00576329231262207,
+      "SpillPSum": 0.01259160041809082,
+      "SplitAPUnionSets": 0.009907007217407227,
+      "SplitAccGrp": 0.0010552406311035156,
+      "StaticProfiler": 0.0033452510833740234,
+      "StaticTransposeLocalTensor": 0.005699634552001953,
+      "SundaISel": 0.04179859161376953,
+      "TCTransform": 0.0024602413177490234,
+      "TensorInitialization": 0.0022628307342529297,
+      "TensorOpSimplifier": 0.006663322448730469,
+      "TensorOpTransform": 0.03399252891540527,
+      "TileCCOps": 0.0057027339935302734,
+      "TilingProfiler": 0.01235508918762207,
+      "TransformConvOp": 0.0025110244750976563,
+      "TritiumFusion": 0.04343080520629883,
+      "ValueNumbering": 0.0041046142578125,
+      "VectorizeDMA": 0.0033228397369384766,
+      "VectorizeMatMult": 0.0033872127532958984,
+      "WeightCoalescing": 0.0023338794708251953,
+      "ZeroSizeTensorElimination": 0.00011372566223144531
+    },
+    "tensorizer": {
+      "DMATilingProfiler::TotalInstructionsAfterTiling": 5268.0,
+      "StaticProfiler::AifUb": 127.58392333984375,
+      "StaticProfiler::ArithmeticIntensityTensorizer": 129.38287353515625,
+      "StaticProfiler::AverageDmaLength": 6718.79638671875,
+      "StaticProfiler::AverageFractalPeUtilization": 100.0,
+      "StaticProfiler::AveragePartitionUtilization": 99.92172241210938,
+      "StaticProfiler::AveragePeUtilization": 100.0,
+      "StaticProfiler::DDRTransferBytes": 198661120.0,
+      "StaticProfiler::InternalTransferBytes": 10321920.0,
+      "StaticProfiler::LoadExpanded": 27264.0,
+      "StaticProfiler::LocalizationEfficiency": 101.41001892089844,
+      "StaticProfiler::LocalizationEfficiencyIgnoreNonlocal": 103.59725189208984,
+      "StaticProfiler::StoreExpanded": 1153.0,
+      "StaticProfiler::TotalDMAExpanded": 28417.0,
+      "StaticProfiler::TotalDynamicInstancesCount": 5111.0,
+      "StaticProfiler::TotalDynamicInstancesWithMmPackedCount": 5111.0,
+      "StaticProfiler::TotalLNCComm": 0.0,
+      "StaticProfiler::TotalLNCCommTransfer": 0.0,
+      "TilingProfiler::AveragePartitionUtilizationAfterTiling": 0.0,
+      "TilingProfiler::AveragePeUtilizationAfterTiling": 0.0,
+      "TilingProfiler::BatchnormInstructionsAfterTiling": 0.0,
+      "TilingProfiler::DmaInstructionsAfterTiling": 0.0,
+      "TilingProfiler::GenericInstructionsAfterTiling": 8.0,
+      "TilingProfiler::MatMultInstructionsAfterTiling": 4276.0,
+      "TilingProfiler::NumPfTransposes": 8.0,
+      "TilingProfiler::NumPfTransposesForIo": 3.0,
+      "TilingProfiler::NumPfTransposesForLocal": 3.0,
+      "TilingProfiler::NumPfTransposesForNonlocal": 2.0,
+      "TilingProfiler::PfTransposeInstructions": 122.0,
+      "TilingProfiler::PfTransposeInstructionsForIo": 34.0,
+      "TilingProfiler::PfTransposeInstructionsForLocal": 24.0,
+      "TilingProfiler::PfTransposeInstructionsForNonlocal": 64.0,
+      "TilingProfiler::ReduceInstructionsAfterTiling": 0.0,
+      "TilingProfiler::SimdInstructionsAfterTiling": 180.0,
+      "TilingProfiler::TotalInstructionsAfterTiling": 0.0,
+      "TransformConvOp::Conv1d_depthwise_bf01_oi01_bf01": 0.0,
+      "TransformConvOp::Conv2d_dw_fb01_io01_01bf_rep_nhwc_Pcinh": 0.0,
+      "TransformConvOp::Conv2d_pbp_0f1b_0i1o_01fb_experimental_1": 0.0,
+      "TransformConvOp::Conv2d_pbp_fb01_io01_01bf_experimental_1": 0.0,
+      "TransformConvOp::conv2d_column_packing": 0.0,
+      "TransformConvOp::conv2d_column_packing_1": 0.0,
+      "TransformConvOp::conv2d_column_packing_io10": 0.0,
+      "TransformConvOp::conv2d_depthwise_f01b_o01i_bf01": 0.0
+    }
+  },
+  "sg0002": {
+    "compiletime": {
+      "AGOrderingAnalysisPass": 0.018787622451782227,
+      "AffinePredicateResolution": 0.0011818408966064453,
+      "AliasDependencyElimination": 0.00011801719665527344,
+      "AliasDependencyInduction": 0.005483388900756836,
+      "AliasDependencyReset": 0.026019811630249023,
+      "BFComputeCutting": 0.00225830078125,
+      "BirCodeGenLoop": 0.4621126651763916,
+      "CCOpFusion": 0.01928091049194336,
+      "CanonicalizeDAGForPGTiling": 0.004612922668457031,
+      "CanonicalizeIR": 0.0017774105072021484,
+      "CoalesceCCOp": 0.014393329620361328,
+      "CommuteConcat": 0.0020241737365722656,
+      "DMALocalityOpt": 0.0052585601806640625,
+      "DMAProfiler": 0.011700868606567383,
+      "DMATilingProfiler": 0.004782676696777344,
+      "DataLocalityOpt": 0.06629562377929688,
+      "DataStreaming": 0.03746771812438965,
+      "DeConcat": 0.0006563663482666016,
+      "DeadCodeElimination": 0.002358675003051758,
+      "DeadStoreElimination": 0.0055620670318603516,
+      "DelinearIndices": 0.004741668701171875,
+      "Delinearization": 0.0036110877990722656,
+      "DoNothing": 8.368492126464844e-05,
+      "DramToDramTranspose": 0.016016721725463867,
+      "DumpGraphAndMetadata": 0.0853111743927002,
+      "EliminateDivs": 0.0025675296783447266,
+      "ExpandBatchNorm": 0.002092123031616211,
+      "ExpandISAMacro": 0.010552406311035156,
+      "FactorizeBlkDims": 0.0076830387115478516,
+      "FactorizeThreadAxesInFreeDims": 0.002122640609741211,
+      "FlattenMacroLoop": 0.002187013626098633,
+      "GenericAccessSimplifier": 0.0009529590606689453,
+      "InferInitValue": 0.0242159366607666,
+      "InferIntrinsicOnCC": 0.009269952774047852,
+      "InferNeuronTensor": 0.020155906677246094,
+      "InferNonlocalTensors": 0.015646696090698242,
+      "InferPSumTensor": 0.3076965808868408,
+      "InlineNativeKernels": 0.009155511856079102,
+      "InsertIOTransposes": 0.015281438827514648,
+      "InsertLocalTransposes": 0.006501436233520508,
+      "InsertOffloadedTransposes": 0.002702474594116211,
+      "LICM": 0.002913951873779297,
+      "LateLegalizeInst": 0.013724088668823242,
+      "LateLegalizePostSplit": 0.012693405151367188,
+      "LateLowerReshapeOp": 0.0025734901428222656,
+      "LateLowerTensorOp": 0.001531362533569336,
+      "LateNeuronInstComb": 0.008354902267456055,
+      "LayoutPreprocessing": 0.026634931564331055,
+      "LayoutPreprocessingAndAnalysis": 0.5595176219940186,
+      "LayoutRequirementAnalysis": 0.005538463592529297,
+      "LegalizeCCOpLayout": 0.0022728443145751953,
+      "LegalizeOpLevelAlias": 0.001255035400390625,
+      "LegalizePartitionReduce": 0.001256704330444336,
+      "LegalizeSundaAccess": 0.07487797737121582,
+      "LegalizeSundaMacro": 0.010920286178588867,
+      "LegalizeType": 0.012901067733764648,
+      "LocalLayoutOpt": 0.012011289596557617,
+      "LoopFusion": 0.006572723388671875,
+      "LoopSplitting": 0.0003001689910888672,
+      "LowerBroadcast": 0.0016355514526367188,
+      "LowerCCOpBlockAxis": 0.0050678253173828125,
+      "LowerComplexBroadcast": 0.0025262832641601563,
+      "LowerIntrinsics": 0.30371904373168945,
+      "LowerTensorOp": 0.011744022369384766,
+      "LowerTranspose": 0.011518478393554688,
+      "MacroGeneration": 0.026911020278930664,
+      "MaskPropagation": 0.0031325817108154297,
+      "MemcpyElimination": 0.027472257614135742,
+      "MutateDataType": 0.0015196800231933594,
+      "NeuronAliasDependencyInduction": 0.00016927719116210938,
+      "NeuronAliasDependencyReset": 0.0242006778717041,
+      "NeuronInstComb": 0.004147529602050781,
+      "NeuronLICM": 0.036264657974243164,
+      "NeuronLoopFusion": 0.00889277458190918,
+      "NeuronLoopInterchange": 0.002141237258911133,
+      "NeuronSimplifier": 0.00720524787902832,
+      "NeuronSimplifyPredicates": 0.11929655075073242,
+      "NeuronValueNumbering": 0.003022432327270508,
+      "OptimizeAliasedCopyChain": 0.0006387233734130859,
+      "OptimizeNKIKernels": 0.5260024070739746,
+      "PAGLayoutOpt": 0.5680239200592041,
+      "PComputeCutting": 0.0048143863677978516,
+      "PGLayoutTilingPipeline": 1.6304676532745361,
+      "PGTiling": 0.1616363525390625,
+      "PadElimination": 0.0003521442413330078,
+      "ParAxesAnnotation": 0.0544736385345459,
+      "PartialLoopFusion": 0.005907773971557617,
+      "PartialSimdFusion": 0.0038967132568359375,
+      "PerfectLoopNest": 0.0021576881408691406,
+      "RecognizeOpIdiom": 0.0039520263671875,
+      "Recompute": 0.0002884864807128906,
+      "RelaxPredicates": 0.013870716094970703,
+      "Rematerialization": 0.0024657249450683594,
+      "ReshapeWeights": 0.0006930828094482422,
+      "ResolveAccessConflict": 0.0038983821868896484,
+      "ResolveComplicatePredicates": 0.0012950897216796875,
+      "RewriteReplicationMatmul": 0.002060413360595703,
+      "RewriteWeights": 0.0028791427612304688,
+      "SFKVectorizer": 0.28761887550354004,
+      "SimpleAllReduceTiling": 0.008704662322998047,
+      "Simplifier": 0.003449678421020508,
+      "SimplifyMacroPredicates": 0.010317325592041016,
+      "SimplifyNeuronTensor": 1.0378923416137695,
+      "SimplifySlice": 0.0008852481842041016,
+      "SimplifyTensor": 0.005218982696533203,
+      "SpillPSum": 0.009551286697387695,
+      "SplitAPUnionSets": 0.10591006278991699,
+      "SplitAccGrp": 0.0011169910430908203,
+      "StaticProfiler": 0.01290583610534668,
+      "StaticTransposeLocalTensor": 0.003824472427368164,
+      "SundaISel": 0.041872262954711914,
+      "TCTransform": 0.0008666515350341797,
+      "TensorInitialization": 0.013058185577392578,
+      "TensorOpSimplifier": 0.0061550140380859375,
+      "TensorOpTransform": 0.020328521728515625,
+      "TileCCOps": 0.006834983825683594,
+      "TilingProfiler": 0.0072863101959228516,
+      "TransformConvOp": 0.0032320022583007813,
+      "TritiumFusion": 0.03062152862548828,
+      "ValueNumbering": 0.0023603439331054688,
+      "VectorizeDMA": 0.004430294036865234,
+      "VectorizeMatMult": 0.0021605491638183594,
+      "WeightCoalescing": 0.00825953483581543,
+      "ZeroSizeTensorElimination": 0.00011014938354492188
+    },
+    "tensorizer": {
+      "DMATilingProfiler::TotalInstructionsAfterTiling": 42834.0,
+      "StaticProfiler::AifUb": 129.43267822265625,
+      "StaticProfiler::ArithmeticIntensityTensorizer": 128.19729614257813,
+      "StaticProfiler::AverageDmaLength": 4810.17578125,
+      "StaticProfiler::AverageFractalPeUtilization": 99.65389251708984,
+      "StaticProfiler::AveragePartitionUtilization": 97.55139923095703,
+      "StaticProfiler::AveragePeUtilization": 98.60253143310547,
+      "StaticProfiler::DDRTransferBytes": 782946624.0,
+      "StaticProfiler::InternalTransferBytes": 629086720.0,
+      "StaticProfiler::LoadExpanded": 97814.0,
+      "StaticProfiler::LocalizationEfficiency": 99.04553985595703,
+      "StaticProfiler::LocalizationEfficiencyIgnoreNonlocal": 100.20111846923828,
+      "StaticProfiler::StoreExpanded": 1757.0,
+      "StaticProfiler::TotalDMAExpanded": 99571.0,
+      "StaticProfiler::TotalDynamicInstancesCount": 50031.0,
+      "StaticProfiler::TotalDynamicInstancesWithMmPackedCount": 49585.0,
+      "StaticProfiler::TotalLNCComm": 0.0,
+      "StaticProfiler::TotalLNCCommTransfer": 0.0,
+      "TilingProfiler::AveragePartitionUtilizationAfterTiling": 0.0,
+      "TilingProfiler::AveragePeUtilizationAfterTiling": 0.0,
+      "TilingProfiler::BatchnormInstructionsAfterTiling": 0.0,
+      "TilingProfiler::DmaInstructionsAfterTiling": 0.0,
+      "TilingProfiler::GenericInstructionsAfterTiling": 4.0,
+      "TilingProfiler::MatMultInstructionsAfterTiling": 22464.0,
+      "TilingProfiler::NumPfTransposes": 5.0,
+      "TilingProfiler::NumPfTransposesForIo": 1.0,
+      "TilingProfiler::NumPfTransposesForLocal": 1.0,
+      "TilingProfiler::NumPfTransposesForNonlocal": 3.0,
+      "TilingProfiler::PfTransposeInstructions": 19105.0,
+      "TilingProfiler::PfTransposeInstructionsForIo": 19008.0,
+      "TilingProfiler::PfTransposeInstructionsForLocal": 1.0,
+      "TilingProfiler::PfTransposeInstructionsForNonlocal": 96.0,
+      "TilingProfiler::ReduceInstructionsAfterTiling": 4.0,
+      "TilingProfiler::SimdInstructionsAfterTiling": 158.0,
+      "TilingProfiler::TotalInstructionsAfterTiling": 0.0,
+      "TransformConvOp::Conv1d_depthwise_bf01_oi01_bf01": 0.0,
+      "TransformConvOp::Conv2d_dw_fb01_io01_01bf_rep_nhwc_Pcinh": 0.0,
+      "TransformConvOp::Conv2d_pbp_0f1b_0i1o_01fb_experimental_1": 0.0,
+      "TransformConvOp::Conv2d_pbp_fb01_io01_01bf_experimental_1": 0.0,
+      "TransformConvOp::conv2d_column_packing": 0.0,
+      "TransformConvOp::conv2d_column_packing_1": 0.0,
+      "TransformConvOp::conv2d_column_packing_io10": 0.0,
+      "TransformConvOp::conv2d_depthwise_f01b_o01i_bf01": 0.0
+    }
+  },
+  "sg01": {
+    "compiletime": {
+      "CanonicalizeConv": 2.300000051036477e-05,
+      "CanonicalizeForTensorizer": 1.2999999853491317e-05,
+      "Canonicalizer": 0.00028199999360367656,
+      "HoistCompute": 3.000000106112566e-06,
+      "IdentifyCrossPassTensors": 2.4000000848900527e-05,
+      "MemcastMotion": 9.999999747378752e-06,
+      "PenguinizeFunctions": 1.4000000192027073e-05,
+      "PruneFunctions": 1.2000000424450263e-05,
+      "RemoveOptimizationBarriers": 2.2000000171829015e-05,
+      "ScatterMotion": 1.700000029813964e-05,
+      "TensorizerLegalizationPass": 3.400000059627928e-05,
+      "VerifySupportedOps": 1.1000000085914508e-05,
+      "algsimp": 6.500000017695129e-05,
+      "batchnorm_expander": 1.2000000424450263e-05,
+      "boundary-marker-removal": 3.999999989900971e-06,
+      "call-inliner": 9.999999747378752e-06,
+      "canonicalize-boundary-marker": 4.999999873689376e-06,
+      "collective-stream-id-checker": 3.999999989900971e-06,
+      "comparison-expander": 4.999999873689376e-06,
+      "computation-deduplicator": 2.099999983329326e-05,
+      "conditional-to-select": 4.999999873689376e-06,
+      "config-lowering": 2.9000000722589903e-05,
+      "constant_folding": 9.000000318337698e-06,
+      "cse": 1.2000000424450263e-05,
+      "dce": 9.999999974752427e-07,
+      "dynamic-slice-transpose": 3.999999989900971e-06,
+      "eliminate-redundant-compare": 3.999999989900971e-06,
+      "emit-offloaded-dropout": 1.2999999853491317e-05,
+      "flatten-call-graph": 9.000000318337698e-06,
+      "fuse-send-recv": 2.9000000722589903e-05,
+      "hilo::LegalizeAlias": 4.999999873689376e-06,
+      "hilo::NeuronInstCombine": 3.5000000934815034e-05,
+      "hilo::NeuronOpFusion": 1.2999999853491317e-05,
+      "hilo::ReplaceTokenTypeWithU8Pass": 1.5999999959603883e-05,
+      "hilo::ScheduleFusion": 9.999999974752427e-07,
+      "hilo::SixtyFourHack": 9.999999747378752e-06,
+      "hilo::VerifyAliasing": 1.9999999949504854e-06,
+      "hlo-mac-count": 2.5999999706982635e-05,
+      "hlo-verifier": 0.00018899999849963933,
+      "legalize-ccops": 9.999999974752427e-07,
+      "legalize-compare": 3.999999989900971e-06,
+      "lower-argminmax-custom-call": 3.999999989900971e-06,
+      "map-inline": 1.2000000424450263e-05,
+      "metadata-naming": 2.4000000848900527e-05,
+      "mlir::detail::OpToOpPassAdaptor": 2.2000000171829015e-05,
+      "mlir::hlo::MhloToPyPenguin": 0.0008980000275187194,
+      "mlir::mhlo::LowerComplexExtraPass": 7.999999797903001e-05,
+      "mlir::mhlo::LowerComplexPass": 0.00013499999477062374,
+      "native-to-custom-softmax": 7.000000096013537e-06,
+      "native-to-custom-softmax-dx": 1.5999999959603883e-05,
+      "operand_upcaster": 1.8999999156221747e-05,
+      "post-par-pipe-begin": 1.9999999949504854e-06,
+      "post-par-pipe-end": 0.0,
+      "post-partition-simplification": 0.0005530000198632479,
+      "replace-minimum-constant": 6.000000212225132e-06,
+      "reshape-mover": 3.000000106112566e-06,
+      "simplify-concat": 4.8999998398358e-05,
+      "simplify-while-loops": 1.9999999949504854e-06,
+      "transform-variadic-reduce": 9.000000318337698e-06,
+      "tuple-simplifier": 4.999999873689376e-06,
+      "unpack-nested-aws-ntwsr": 3.999999989900971e-06,
+      "unroll-while-loop": 9.999999974752427e-07
+    },
+    "hilo": {
+      "ArithmeticIntensity": 123.27030181884766,
+      "HloMacCount": 12415139840.0,
+      "Traffic": 201429536.0
+    }
+  },
+  "sg02": {
+    "compiletime": {
+      "CanonicalizeConv": 0.0,
+      "CanonicalizeForTensorizer": 1.2999999853491317e-05,
+      "Canonicalizer": 0.00033599999733269215,
+      "HoistCompute": 0.0,
+      "IdentifyCrossPassTensors": 2.2000000171829015e-05,
+      "MemcastMotion": 0.0,
+      "PenguinizeFunctions": 9.000000318337698e-06,
+      "PruneFunctions": 9.999999747378752e-06,
+      "RemoveOptimizationBarriers": 1.8999999156221747e-05,
+      "ScatterMotion": 0.0,
+      "TensorizerLegalizationPass": 6.000000212225132e-06,
+      "VerifySupportedOps": 1.1000000085914508e-05,
+      "algsimp": 5.999999848427251e-05,
+      "batchnorm_expander": 1.1000000085914508e-05,
+      "boundary-marker-removal": 3.999999989900971e-06,
+      "call-inliner": 1.1000000085914508e-05,
+      "canonicalize-boundary-marker": 3.999999989900971e-06,
+      "collective-stream-id-checker": 1.9999999949504854e-06,
+      "comparison-expander": 4.999999873689376e-06,
+      "computation-deduplicator": 2.099999983329326e-05,
+      "conditional-to-select": 6.000000212225132e-06,
+      "config-lowering": 2.5999999706982635e-05,
+      "constant_folding": 9.000000318337698e-06,
+      "cse": 1.1000000085914508e-05,
+      "dce": 9.999999974752427e-07,
+      "dynamic-slice-transpose": 3.999999989900971e-06,
+      "eliminate-redundant-compare": 3.000000106112566e-06,
+      "emit-offloaded-dropout": 1.2000000424450263e-05,
+      "flatten-call-graph": 1.2000000424450263e-05,
+      "fuse-send-recv": 1.8000000636675395e-05,
+      "hilo::LegalizeAlias": 1.9999999949504854e-06,
+      "hilo::NeuronInstCombine": 4.999999873689376e-06,
+      "hilo::NeuronOpFusion": 2.8000000384054147e-05,
+      "hilo::ReplaceTokenTypeWithU8Pass": 7.000000096013537e-06,
+      "hilo::ScheduleFusion": 0.0,
+      "hilo::SixtyFourHack": 3.7000001611886546e-05,
+      "hilo::VerifyAliasing": 9.999999974752427e-07,
+      "hlo-mac-count": 0.00016900000628083944,
+      "hlo-verifier": 0.00015799999528098851,
+      "legalize-ccops": 9.999999974752427e-07,
+      "legalize-compare": 3.000000106112566e-06,
+      "lower-argminmax-custom-call": 3.999999989900971e-06,
+      "map-inline": 1.2999999853491317e-05,
+      "metadata-naming": 1.5999999959603883e-05,
+      "mlir::detail::OpToOpPassAdaptor": 2.700000004551839e-05,
+      "mlir::hlo::MhloToPyPenguin": 0.0008440000237897038,
+      "mlir::mhlo::LowerComplexExtraPass": 6.900000153109431e-05,
+      "mlir::mhlo::LowerComplexPass": 1.5999999959603883e-05,
+      "native-to-custom-softmax": 6.000000212225132e-06,
+      "native-to-custom-softmax-dx": 1.9999999494757503e-05,
+      "operand_upcaster": 1.2000000424450263e-05,
+      "post-par-pipe-begin": 9.999999974752427e-07,
+      "post-par-pipe-end": 0.0,
+      "post-partition-simplification": 0.000526999996509403,
+      "replace-minimum-constant": 9.000000318337698e-06,
+      "reshape-mover": 3.000000106112566e-06,
+      "simplify-concat": 4.400000034365803e-05,
+      "simplify-while-loops": 1.9999999949504854e-06,
+      "transform-variadic-reduce": 4.8999998398358e-05,
+      "tuple-simplifier": 4.999999873689376e-06,
+      "unpack-nested-aws-ntwsr": 3.999999989900971e-06,
+      "unroll-while-loop": 9.999999974752427e-07
+    },
+    "hilo": {
+      "ArithmeticIntensity": 25.691875457763672,
+      "HloMacCount": 9974841344.0,
+      "Traffic": 776497728.0
+    }
+  }
+}

context_encoding_model/_tp0_bk0/graph.neff

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:c0aeace703e08ac36bdcb2027d9a278403cb96ef39f48bddc999b077215e8a36
+size 1557504

context_encoding_model/_tp0_bk0/metaneff.pb

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:0aef68f833b52be82fd0e17410bcfd279e5719338cb746c0619d5139fc4a3d02
+size 1042690

context_encoding_model/_tp0_bk0/model.MODULE_f4171003694760566af4+a9cd68fb.hlo_module.pb

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:d89b9e073981a0b1b7d0bbd0a24f147e9df13c5706d9d6be9971b857124c9496
+size 1119812

context_encoding_model/_tp0_bk0/model.MODULE_f4171003694760566af4+a9cd68fb.neff

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:c0aeace703e08ac36bdcb2027d9a278403cb96ef39f48bddc999b077215e8a36
+size 1557504

context_encoding_model/_tp0_bk0/neuron_config.json

@@ -0,0 +1,220 @@
+{
+  "_attn_implementation_autoset": false,
+  "_name_or_path": "Qwen/Qwen3-8B",
+  "add_cross_attention": false,
+  "architectures": [
+    "Qwen3ForCausalLM"
+  ],
+  "attention_bias": false,
+  "attention_dropout": 0.0,
+  "attribute_map": {},
+  "bad_words_ids": null,
+  "begin_suppress_tokens": null,
+  "bos_token_id": 151643,
+  "chunk_size_feed_forward": 0,
+  "cross_attention_hidden_size": null,
+  "decoder_start_token_id": null,
+  "diversity_penalty": 0.0,
+  "do_sample": false,
+  "early_stopping": false,
+  "encoder_no_repeat_ngram_size": 0,
+  "eos_token_id": 151645,
+  "exponential_decay_length_penalty": null,
+  "finetuning_task": null,
+  "forced_bos_token_id": null,
+  "forced_eos_token_id": null,
+  "fused_spec_config": null,
+  "head_dim": 128,
+  "hidden_act": "silu",
+  "hidden_size": 4096,
+  "id2label": {
+    "0": "LABEL_0",
+    "1": "LABEL_1"
+  },
+  "initializer_range": 0.02,
+  "intermediate_size": 12288,
+  "is_decoder": false,
+  "is_encoder_decoder": false,
+  "label2id": {
+    "LABEL_0": 0,
+    "LABEL_1": 1
+  },
+  "length_penalty": 1.0,
+  "max_length": 20,
+  "max_position_embeddings": 40960,
+  "max_window_layers": 36,
+  "metadata": null,
+  "min_length": 0,
+  "model_type": "qwen3",
+  "neuron_config": {
+    "activation_quantization_type": null,
+    "allow_input_truncation": false,
+    "apply_seq_ids_mask": false,
+    "async_mode": false,
+    "attention_dp_degree": 1,
+    "attention_dtype": null,
+    "attn_block_cte_nki_kernel_enabled": false,
+    "attn_block_tkg_nki_kernel_cache_update": false,
+    "attn_block_tkg_nki_kernel_enabled": false,
+    "attn_cls": {
+      "__module__": "neuronx_distributed_inference.models.qwen3.modeling_qwen3",
+      "__name__": "NeuronQwen3Attention"
+    },
+    "attn_kernel_enabled": null,
+    "attn_tkg_builtin_kernel_enabled": false,
+    "attn_tkg_nki_kernel_enabled": false,
+    "batch_size": 1,
+    "bucket_n_active_tokens": true,
+    "buckets": [
+      128
+    ],
+    "cast_type": "config",
+    "cc_pipeline_tiling_factor": 2,
+    "chunked_prefill_config": null,
+    "context_encoding_buckets": [
+      128
+    ],
+    "cp_degree": 1,
+    "ctx_batch_size": 1,
+    "disable_kv_cache_tiling": false,
+    "draft_model_modules_to_not_convert": null,
+    "enable_bucketing": true,
+    "enable_eagle_draft_input_norm": false,
+    "enable_eagle_speculation": false,
+    "enable_fused_speculation": false,
+    "enable_long_context_mode": false,
+    "enable_output_completion_notifications": false,
+    "enable_spill_reload_dge": false,
+    "enable_token_tree": false,
+    "ep_degree": 1,
+    "expert_mlp_nki_kernel_enabled": null,
+    "flash_decoding_enabled": false,
+    "fused_qkv": false,
+    "fused_rmsnorm_skip_gamma": false,
+    "is_block_kv_layout": null,
+    "is_chunked_prefill": false,
+    "is_continuous_batching": true,
+    "is_eagle_draft": false,
+    "is_medusa": false,
+    "is_prefill_stage": true,
+    "is_prefix_caching": false,
+    "k_cache_transposed": false,
+    "kv_cache_batch_size": 1,
+    "kv_cache_padding_size": 0,
+    "kv_cache_quant": false,
+    "kv_cache_tiling": false,
+    "layer_boundary_markers": false,
+    "lm_head_pad": false,
+    "lm_head_pad_alignment_size": 1,
+    "local_ranks_size": 2,
+    "logical_nc_config": 1,
+    "lora_config": null,
+    "max_batch_size": 1,
+    "max_context_length": 1024,
+    "max_length": 1024,
+    "max_new_tokens": null,
+    "medusa_speculation_length": 0,
+    "medusa_tree": null,
+    "mlp_kernel_enabled": false,
+    "mlp_kernel_fuse_residual_add": false,
+    "modules_to_not_convert": null,
+    "moe_fused_nki_kernel_enabled": null,
+    "n_active_tokens": 1024,
+    "n_positions": 1024,
+    "num_medusa_heads": 0,
+    "on_cpu": false,
+    "on_device_sampling_config": {
+      "deterministic": false,
+      "do_sample": false,
+      "dynamic": true,
+      "global_topk": 256,
+      "on_device_sampling_config": true,
+      "temperature": 1.0,
+      "top_k": 1,
+      "top_k_kernel_enabled": false,
+      "top_p": 1.0
+    },
+    "output_logits": false,
+    "overrides_torch_dtype": true,
+    "pa_block_size": 1024,
+    "pa_num_blocks": 1,
+    "padding_side": "right",
+    "pp_degree": 1,
+    "prefix_buckets": null,
+    "qk_layernorm": false,
+    "qkv_kernel_enabled": false,
+    "qkv_kernel_fuse_residual_add": false,
+    "qkv_kernel_nbsd_layout": false,
+    "quantization_dtype": "int8",
+    "quantization_type": "per_tensor_symmetric",
+    "quantize_clamp_bound": Infinity,
+    "quantized": false,
+    "quantized_checkpoints_path": null,
+    "quantized_mlp_kernel_enabled": false,
+    "rmsnorm_quantize_kernel_enabled": false,
+    "router_topk_nki_kernel_enabled": null,
+    "rpl_reduce_dtype": null,
+    "save_sharded_checkpoint": true,
+    "scratchpad_page_size": null,
+    "seq_len": 1024,
+    "seq_len_threshold_for_cc_tiling": 16384,
+    "sequence_parallel_enabled": false,
+    "shared_mlp_nki_kernel_enabled": null,
+    "skip_sharding": false,
+    "skip_warmup": false,
+    "spec_batch_size": 1,
+    "speculation_length": 0,
+    "start_rank_id": 0,
+    "target": null,
+    "tile_cc": false,
+    "tkg_batch_size": 1,
+    "token_generation_buckets": null,
+    "token_tree_config": null,
+    "torch_dtype": "bfloat16",
+    "tp_degree": 2,
+    "vocab_parallel": false,
+    "weight_gather_seq_len_threshold": 32768,
+    "weights_to_skip_layout_optimization": [],
+    "world_size": 2
+  },
+  "no_repeat_ngram_size": 0,
+  "num_attention_heads": 32,
+  "num_beam_groups": 1,
+  "num_beams": 1,
+  "num_cores_per_group": 1,
+  "num_hidden_layers": 36,
+  "num_key_value_heads": 8,
+  "num_return_sequences": 1,
+  "output_attentions": false,
+  "output_hidden_states": false,
+  "output_scores": false,
+  "pad_token_id": 0,
+  "prefix": null,
+  "problem_type": null,
+  "pruned_heads": {},
+  "remove_invalid_values": false,
+  "repetition_penalty": 1.0,
+  "return_dict": true,
+  "return_dict_in_generate": false,
+  "rms_norm_eps": 1e-06,
+  "rope_scaling": null,
+  "rope_theta": 1000000,
+  "sep_token_id": null,
+  "sliding_window": null,
+  "suppress_tokens": null,
+  "task_specific_params": null,
+  "temperature": 1.0,
+  "tf_legacy_loss": false,
+  "tie_encoder_decoder": false,
+  "tie_word_embeddings": false,
+  "tokenizer_class": null,
+  "top_k": 50,
+  "top_p": 1.0,
+  "torchscript": false,
+  "transformers_version": "4.51.0",
+  "typical_p": 1.0,
+  "use_bfloat16": false,
+  "use_cache": true,
+  "use_sliding_window": false,
+  "vocab_size": 151936
+}

context_encoding_model/_tp0_bk1/command.txt

@@ -0,0 +1 @@
+neuronx-cc compile --framework=XLA model.MODULE_2914133a46cb7b4660ab+d7af8a84.hlo_module.pb --output model.MODULE_2914133a46cb7b4660ab+d7af8a84.neff --target=trn1 --auto-cast=none --model-type=transformer '--tensorizer-options=--enable-ccop-compute-overlap --cc-pipeline-tiling-factor=2 --vectorize-strided-dma ' --lnc=1 -O1 '--internal-hlo2tensorizer-options= --modular-flow-mac-threshold=10  --verify-hlo=true' --logfile=log-neuron-cc.txt --verbose=35

context_encoding_model/_tp0_bk1/compile_flags.MODULE_2914133a46cb7b4660ab+d7af8a84.json

@@ -0,0 +1 @@
+["--target=trn1", "--auto-cast=none", "--model-type=transformer", "--tensorizer-options=--enable-ccop-compute-overlap --cc-pipeline-tiling-factor=2 --vectorize-strided-dma ", "--lnc=1", "-O1", "--internal-hlo2tensorizer-options= --modular-flow-mac-threshold=10  --verify-hlo=true", "--logfile=/home/ubuntu/qwen3/context_encoding_model/_tp0_bk1/log-neuron-cc.txt"]

context_encoding_model/_tp0_bk1/global_metric_store.json

@@ -0,0 +1,1079 @@
+{
+  "Average": {
+    "tensorizer": {
+      "StaticProfiler::AverageFractalPeUtilization": 99.65782165527344,
+      "StaticProfiler::AveragePartitionUtilization": 97.58238220214844,
+      "StaticProfiler::AveragePeUtilization": 98.61824035644531,
+      "StaticProfiler::LocalizationEfficiency": 98.78419494628906,
+      "StaticProfiler::LocalizationEfficiencyIgnoreNonlocal": 100.47209167480469,
+      "TilingProfiler::AveragePartitionUtilizationAfterTiling": 0.0,
+      "TilingProfiler::AveragePeUtilizationAfterTiling": 0.0
+    }
+  },
+  "Count": {
+    "tensorizer": {
+      "StaticProfiler::AverageFractalPeUtilization": 1.0,
+      "StaticProfiler::AveragePartitionUtilization": 1.0,
+      "StaticProfiler::AveragePeUtilization": 1.0,
+      "StaticProfiler::LocalizationEfficiency": 1.0,
+      "StaticProfiler::LocalizationEfficiencyIgnoreNonlocal": 1.0,
+      "TilingProfiler::AveragePartitionUtilizationAfterTiling": 1.0,
+      "TilingProfiler::AveragePeUtilizationAfterTiling": 1.0
+    }
+  },
+  "Sum": {
+    "compiletime": {
+      "AGOrderingAnalysisPass": 0.019578933715820313,
+      "AffinePredicateResolution": 0.0019481182098388672,
+      "AliasDependencyElimination": 0.0001239776611328125,
+      "AliasDependencyInduction": 0.00577092170715332,
+      "AliasDependencyReset": 0.027690649032592773,
+      "BFComputeCutting": 0.0023322105407714844,
+      "BirCodeGenLoop": 0.4628438949584961,
+      "CCOpFusion": 0.022275209426879883,
+      "CanonicalizeConv": 3.300000025774352e-05,
+      "CanonicalizeDAGForPGTiling": 0.005593061447143555,
+      "CanonicalizeForTensorizer": 4.400000034365803e-05,
+      "CanonicalizeIR": 0.001634359359741211,
+      "Canonicalizer": 0.0008999999845400453,
+      "CoalesceCCOp": 0.015577077865600586,
+      "CommuteConcat": 0.0008616447448730469,
+      "DMALocalityOpt": 0.007327079772949219,
+      "DMAProfiler": 0.012569665908813477,
+      "DMATilingProfiler": 0.0037431716918945313,
+      "DataLocalityOpt": 0.06741714477539063,
+      "DataStreaming": 0.03615880012512207,
+      "DeConcat": 0.0005049705505371094,
+      "DeadCodeElimination": 0.0009002685546875,
+      "DeadStoreElimination": 0.0056514739990234375,
+      "DelinearIndices": 0.004773139953613281,
+      "Delinearization": 0.0026137828826904297,
+      "DoNothing": 0.0001933574676513672,
+      "DramToDramTranspose": 0.019293546676635742,
+      "DumpGraphAndMetadata": 0.10360383987426758,
+      "EliminateDivs": 0.003831148147583008,
+      "ExpandBatchNorm": 0.0019576549530029297,
+      "ExpandISAMacro": 0.012068033218383789,
+      "FactorizeBlkDims": 0.008942604064941406,
+      "FactorizeThreadAxesInFreeDims": 0.001847982406616211,
+      "FlattenMacroLoop": 0.003529787063598633,
+      "GenericAccessSimplifier": 0.0008223056793212891,
+      "HoistCompute": 7.999999979801942e-06,
+      "IdentifyCrossPassTensors": 4.8000001697801054e-05,
+      "InferInitValue": 0.025947093963623047,
+      "InferIntrinsicOnCC": 0.00908350944519043,
+      "InferNeuronTensor": 0.02371978759765625,
+      "InferNonlocalTensors": 0.014753341674804688,
+      "InferPSumTensor": 0.309035062789917,
+      "InlineNativeKernels": 0.008690595626831055,
+      "InsertIOTransposes": 0.01906275749206543,
+      "InsertLocalTransposes": 0.004312276840209961,
+      "InsertOffloadedTransposes": 0.002802133560180664,
+      "LICM": 0.003081083297729492,
+      "LateLegalizeInst": 0.014100313186645508,
+      "LateLegalizePostSplit": 0.012533903121948242,
+      "LateLowerReshapeOp": 0.001035451889038086,
+      "LateLowerTensorOp": 0.002605438232421875,
+      "LateNeuronInstComb": 0.009373188018798828,
+      "LayoutPreprocessing": 0.03434133529663086,
+      "LayoutPreprocessingAndAnalysis": 0.07319903373718262,
+      "LayoutRequirementAnalysis": 0.005194187164306641,
+      "LegalizeCCOpLayout": 0.0025322437286376953,
+      "LegalizeOpLevelAlias": 0.0020308494567871094,
+      "LegalizePartitionReduce": 0.0010001659393310547,
+      "LegalizeSundaAccess": 0.0786747932434082,
+      "LegalizeSundaMacro": 0.011176109313964844,
+      "LegalizeType": 0.014636754989624023,
+      "LocalLayoutOpt": 0.014019250869750977,
+      "LoopFusion": 0.005472898483276367,
+      "LoopSplitting": 0.00038623809814453125,
+      "LowerBroadcast": 0.0027265548706054688,
+      "LowerCCOpBlockAxis": 0.0058476924896240234,
+      "LowerComplexBroadcast": 0.00213623046875,
+      "LowerIntrinsics": 0.3070671558380127,
+      "LowerTensorOp": 0.010679960250854492,
+      "LowerTranspose": 0.012553691864013672,
+      "MacroGeneration": 0.029733657836914063,
+      "MaskPropagation": 0.0028328895568847656,
+      "MemcastMotion": 1.8999999156221747e-05,
+      "MemcpyElimination": 0.026583433151245117,
+      "MutateDataType": 0.0020093917846679688,
+      "NeuronAliasDependencyInduction": 0.00018548965454101563,
+      "NeuronAliasDependencyReset": 0.02524423599243164,
+      "NeuronInstComb": 0.004286766052246094,
+      "NeuronLICM": 0.03554058074951172,
+      "NeuronLoopFusion": 0.007987260818481445,
+      "NeuronLoopInterchange": 0.0023233890533447266,
+      "NeuronSimplifier": 0.0075054168701171875,
+      "NeuronSimplifyPredicates": 0.12207841873168945,
+      "NeuronValueNumbering": 0.0038213729858398438,
+      "OptimizeAliasedCopyChain": 0.0005936622619628906,
+      "OptimizeNKIKernels": 0.44962644577026367,
+      "PAGLayoutOpt": 0.0999138355255127,
+      "PComputeCutting": 0.005170106887817383,
+      "PGLayoutTilingPipeline": 0.7408750057220459,
+      "PGTiling": 0.29245758056640625,
+      "PadElimination": 0.000308990478515625,
+      "ParAxesAnnotation": 0.05283546447753906,
+      "PartialLoopFusion": 0.0043125152587890625,
+      "PartialSimdFusion": 0.004901885986328125,
+      "PenguinizeFunctions": 4.3000000005122274e-05,
+      "PerfectLoopNest": 0.001722574234008789,
+      "PruneFunctions": 4.199999966658652e-05,
+      "RecognizeOpIdiom": 0.004076480865478516,
+      "Recompute": 0.0002620220184326172,
+      "RelaxPredicates": 0.013286828994750977,
+      "Rematerialization": 0.0021238327026367188,
+      "RemoveOptimizationBarriers": 4.3000000005122274e-05,
+      "ReshapeWeights": 0.0006799697875976563,
+      "ResolveAccessConflict": 0.0040090084075927734,
+      "ResolveComplicatePredicates": 0.001981496810913086,
+      "RewriteReplicationMatmul": 0.0021796226501464844,
+      "RewriteWeights": 0.0022602081298828125,
+      "SFKVectorizer": 0.274188756942749,
+      "ScatterMotion": 5.7999997807201e-05,
+      "SimpleAllReduceTiling": 0.009164094924926758,
+      "Simplifier": 0.0046122074127197266,
+      "SimplifyMacroPredicates": 0.010458230972290039,
+      "SimplifyNeuronTensor": 1.0516629219055176,
+      "SimplifySlice": 0.0009145736694335938,
+      "SimplifyTensor": 0.00577855110168457,
+      "SpillPSum": 0.012692689895629883,
+      "SplitAPUnionSets": 0.10518908500671387,
+      "SplitAccGrp": 0.001172780990600586,
+      "StaticProfiler": 0.0124053955078125,
+      "StaticTransposeLocalTensor": 0.0038576126098632813,
+      "SundaISel": 0.04396390914916992,
+      "TCTransform": 0.0018804073333740234,
+      "TensorInitialization": 0.012793779373168945,
+      "TensorOpSimplifier": 0.0045316219329833984,
+      "TensorOpTransform": 0.021115541458129883,
+      "TensorizerLegalizationPass": 6.999999459367245e-05,
+      "TileCCOps": 0.0056231021881103516,
+      "TilingProfiler": 0.00790858268737793,
+      "TransformConvOp": 0.0030431747436523438,
+      "TritiumFusion": 0.03186154365539551,
+      "ValueNumbering": 0.0038623809814453125,
+      "VectorizeDMA": 0.0021522045135498047,
+      "VectorizeMatMult": 0.003453969955444336,
+      "VerifySupportedOps": 3.300000025774352e-05,
+      "WeightCoalescing": 0.009244203567504883,
+      "ZeroSizeTensorElimination": 0.00011420249938964844,
+      "algsimp": 0.0026100000832229853,
+      "batchnorm_expander": 3.9999998989515007e-05,
+      "boundary-marker-removal": 1.2000000424450263e-05,
+      "call-inliner": 0.00046499999007210135,
+      "canonicalize-boundary-marker": 1.8000000636675395e-05,
+      "collective-stream-id-checker": 9.200000204145908e-05,
+      "comparison-expander": 0.0005959999980404973,
+      "computation-deduplicator": 6.900000153109431e-05,
+      "conditional-to-select": 1.700000029813964e-05,
+      "config-lowering": 7.79999973019585e-05,
+      "constant-statistics": 0.0005530000198632479,
+      "constant_folding": 0.0003320000250823796,
+      "cse": 3.7000001611886546e-05,
+      "dce": 7.800000457791612e-05,
+      "dot_decomposer": 0.0014440000522881746,
+      "dynamic-slice-transpose": 1.2000000424450263e-05,
+      "eliminate-redundant-compare": 0.00028100001509301364,
+      "emit-offloaded-dropout": 4.099999932805076e-05,
+      "flatten-call-graph": 0.0009379999246448278,
+      "fuse-send-recv": 7.200000254670158e-05,
+      "hilo::LegalizeAlias": 1.2999999853491317e-05,
+      "hilo::NeuronInstCombine": 0.00010099999781232327,
+      "hilo::NeuronOpFusion": 2.5000001187436283e-05,
+      "hilo::ReplaceTokenTypeWithU8Pass": 4.3000000005122274e-05,
+      "hilo::ScheduleFusion": 1.9999999949504854e-06,
+      "hilo::SixtyFourHack": 8.900000102585182e-05,
+      "hilo::VerifyAliasing": 4.999999873689376e-06,
+      "hlo-mac-count": 0.0012799999676644802,
+      "hlo-verifier": 0.007751000113785267,
+      "instruction-histogram": 0.0006590000120922923,
+      "io-con-pipe-begin": 6.000000212225132e-06,
+      "io-con-pipe-end": 9.999999974752427e-07,
+      "io-layout-normalization": 0.0014029999729245901,
+      "io-statistics": 6.199999916134402e-05,
+      "legalize-ccops": 3.999999989900971e-06,
+      "legalize-compare": 1.1000000085914508e-05,
+      "lower-argminmax-custom-call": 1.2000000424450263e-05,
+      "map-inline": 0.0008909999742172658,
+      "metadata-naming": 5.7999997807201e-05,
+      "mlir::detail::OpToOpPassAdaptor": 0.00016799999866634607,
+      "mlir::hlo::MhloToPyPenguin": 0.0028260000981390476,
+      "mlir::mhlo::LowerComplexExtraPass": 0.00026000000070780516,
+      "mlir::mhlo::LowerComplexPass": 0.0002699999895412475,
+      "native-to-custom-softmax": 0.0007219999679364264,
+      "native-to-custom-softmax-dx": 0.0005740000051446259,
+      "operand_upcaster": 6.399999983841553e-05,
+      "opt-barrier-removal": 0.0005649999948218465,
+      "post-par-pipe-begin": 9.600000339560211e-05,
+      "post-par-pipe-end": 0.0,
+      "post-partition-simplification": 0.0016929999692365527,
+      "pre-par-pipe-begin": 9.999999974752427e-07,
+      "pre-par-pipe-end": 0.0,
+      "pre-partition-simplification": 0.1934960037469864,
+      "replace-minimum-constant": 0.00044299999717622995,
+      "reshape-mover": 0.00010800000018207356,
+      "simplify-concat": 0.00014099999680183828,
+      "simplify-while-loops": 9.600000339560211e-05,
+      "transform-variadic-reduce": 7.900000491645187e-05,
+      "tuple-simplifier": 0.0002980000281240791,
+      "unpack-nested-aws-ntwsr": 0.0004720000142697245,
+      "unroll-while-loop": 1.8999999156221747e-05,
+      "zero_sized_hlo_elimination": 0.0008989999769255519
+    },
+    "hilo": {
+      "ConstantSize": 599333.0,
+      "HloInputCount": 475.0,
+      "HloMacCount": 50240159744.0,
+      "HloOutputCount": 73.0,
+      "IfmapSize": 8266543104.0,
+      "OfmapSize": 75497472.0,
+      "OutputsReadFromCount": 0.0,
+      "PassthroughTensorsCount": 0.0,
+      "RedundantOutputCount": 0.0,
+      "Traffic": 1663506816.0
+    },
+    "tensorizer": {
+      "DMATilingProfiler::TotalInstructionsAfterTiling": 43318.0,
+      "StaticProfiler::AifUb": 154.8094024658203,
+      "StaticProfiler::ArithmeticIntensityTensorizer": 152.92723083496094,
+      "StaticProfiler::AverageDmaLength": 4809.89794921875,
+      "StaticProfiler::DDRTransferBytes": 787141440.0,
+      "StaticProfiler::InternalTransferBytes": 634853888.0,
+      "StaticProfiler::LoadExpanded": 98070.0,
+      "StaticProfiler::StoreExpanded": 2397.0,
+      "StaticProfiler::TotalDMAExpanded": 100467.0,
+      "StaticProfiler::TotalDynamicInstancesCount": 50670.0,
+      "StaticProfiler::TotalDynamicInstancesWithMmPackedCount": 50224.0,
+      "StaticProfiler::TotalLNCComm": 0.0,
+      "StaticProfiler::TotalLNCCommTransfer": 0.0,
+      "TilingProfiler::BatchnormInstructionsAfterTiling": 0.0,
+      "TilingProfiler::DmaInstructionsAfterTiling": 0.0,
+      "TilingProfiler::GenericInstructionsAfterTiling": 4.0,
+      "TilingProfiler::MatMultInstructionsAfterTiling": 22848.0,
+      "TilingProfiler::NumPfTransposes": 5.0,
+      "TilingProfiler::NumPfTransposesForIo": 1.0,
+      "TilingProfiler::NumPfTransposesForLocal": 1.0,
+      "TilingProfiler::NumPfTransposesForNonlocal": 3.0,
+      "TilingProfiler::PfTransposeInstructions": 19201.0,
+      "TilingProfiler::PfTransposeInstructionsForIo": 19008.0,
+      "TilingProfiler::PfTransposeInstructionsForLocal": 1.0,
+      "TilingProfiler::PfTransposeInstructionsForNonlocal": 192.0,
+      "TilingProfiler::ReduceInstructionsAfterTiling": 4.0,
+      "TilingProfiler::SimdInstructionsAfterTiling": 158.0,
+      "TilingProfiler::TotalInstructionsAfterTiling": 0.0,
+      "TransformConvOp::Conv1d_depthwise_bf01_oi01_bf01": 0.0,
+      "TransformConvOp::Conv2d_dw_fb01_io01_01bf_rep_nhwc_Pcinh": 0.0,
+      "TransformConvOp::Conv2d_pbp_0f1b_0i1o_01fb_experimental_1": 0.0,
+      "TransformConvOp::Conv2d_pbp_fb01_io01_01bf_experimental_1": 0.0,
+      "TransformConvOp::conv2d_column_packing": 0.0,
+      "TransformConvOp::conv2d_column_packing_1": 0.0,
+      "TransformConvOp::conv2d_column_packing_io10": 0.0,
+      "TransformConvOp::conv2d_depthwise_f01b_o01i_bf01": 0.0
+    }
+  },
+  "all": {
+    "compiletime": {
+      "algsimp": 0.0024220000486820936,
+      "call-inliner": 0.0004349999944679439,
+      "collective-stream-id-checker": 8.199999865610152e-05,
+      "comparison-expander": 0.0005810000002384186,
+      "constant-statistics": 0.0005530000198632479,
+      "constant_folding": 0.0003060000017285347,
+      "dce": 7.500000356230885e-05,
+      "dot_decomposer": 0.0014440000522881746,
+      "eliminate-redundant-compare": 0.0002690000110305846,
+      "flatten-call-graph": 0.0009069999796338379,
+      "hlo-mac-count": 0.0010560000082477927,
+      "hlo-verifier": 0.007164000067859888,
+      "instruction-histogram": 0.0006590000120922923,
+      "io-con-pipe-begin": 6.000000212225132e-06,
+      "io-con-pipe-end": 9.999999974752427e-07,
+      "io-layout-normalization": 0.0014029999729245901,
+      "io-statistics": 6.199999916134402e-05,
+      "map-inline": 0.0008549999911338091,
+      "native-to-custom-softmax": 0.0007029999978840351,
+      "native-to-custom-softmax-dx": 0.000522000016644597,
+      "opt-barrier-removal": 0.0005649999948218465,
+      "pre-par-pipe-begin": 9.999999974752427e-07,
+      "pre-par-pipe-end": 0.0,
+      "pre-partition-simplification": 0.1934960037469864,
+      "replace-minimum-constant": 0.00042100000428035855,
+      "reshape-mover": 9.7999996796716e-05,
+      "simplify-while-loops": 9.000000136438757e-05,
+      "tuple-simplifier": 0.00028300000121816993,
+      "unpack-nested-aws-ntwsr": 0.0004600000102072954,
+      "unroll-while-loop": 1.8999999156221747e-05,
+      "zero_sized_hlo_elimination": 0.0008989999769255519
+    }
+  },
+  "cumsum": {
+    "compiletime": {
+      "CoalesceCCOp": 0.0002090930938720703,
+      "DMALocalityOpt": 0.00018835067749023438,
+      "DMAProfiler": 0.0008924007415771484,
+      "DataStreaming": 0.0002593994140625,
+      "DoNothing": 0.00011873245239257813,
+      "ExpandISAMacro": 0.0005505084991455078,
+      "FactorizeBlkDims": 0.0004696846008300781,
+      "InferPSumTensor": 0.0004990100860595703,
+      "LateLegalizeInst": 0.0004222393035888672,
+      "LateNeuronInstComb": 0.0005340576171875,
+      "LegalizeSundaAccess": 0.0017271041870117188,
+      "LegalizeType": 0.0002815723419189453,
+      "LowerBroadcast": 0.0002243518829345703,
+      "LowerIntrinsics": 0.0002181529998779297,
+      "LowerTranspose": 0.00024199485778808594,
+      "NeuronInstComb": 0.0004971027374267578,
+      "NeuronLICM": 0.0004258155822753906,
+      "NeuronSimplifyPredicates": 0.002941608428955078,
+      "NeuronValueNumbering": 0.0004222393035888672,
+      "SFKVectorizer": 0.002941131591796875,
+      "SimpleAllReduceTiling": 0.00019812583923339844,
+      "SimplifyNeuronTensor": 0.00045800209045410156,
+      "SpillPSum": 0.0005657672882080078,
+      "WeightCoalescing": 0.00020837783813476563
+    }
+  },
+  "sg00": {
+    "compiletime": {
+      "CanonicalizeConv": 3.099999958067201e-05,
+      "CanonicalizeForTensorizer": 1.5999999959603883e-05,
+      "Canonicalizer": 0.00032900000223889947,
+      "HoistCompute": 3.000000106112566e-06,
+      "IdentifyCrossPassTensors": 2.8000000384054147e-05,
+      "MemcastMotion": 1.1000000085914508e-05,
+      "PenguinizeFunctions": 1.5999999959603883e-05,
+      "PruneFunctions": 1.2999999853491317e-05,
+      "RemoveOptimizationBarriers": 2.300000051036477e-05,
+      "ScatterMotion": 1.9999999494757503e-05,
+      "TensorizerLegalizationPass": 4.3000000005122274e-05,
+      "VerifySupportedOps": 1.2000000424450263e-05,
+      "algsimp": 6.600000051548705e-05,
+      "batchnorm_expander": 1.4000000192027073e-05,
+      "boundary-marker-removal": 3.999999989900971e-06,
+      "call-inliner": 9.000000318337698e-06,
+      "canonicalize-boundary-marker": 6.000000212225132e-06,
+      "collective-stream-id-checker": 3.000000106112566e-06,
+      "comparison-expander": 4.999999873689376e-06,
+      "computation-deduplicator": 1.9999999494757503e-05,
+      "conditional-to-select": 4.999999873689376e-06,
+      "config-lowering": 3.099999958067201e-05,
+      "constant_folding": 9.000000318337698e-06,
+      "cse": 1.2999999853491317e-05,
+      "dce": 9.999999974752427e-07,
+      "dynamic-slice-transpose": 3.999999989900971e-06,
+      "eliminate-redundant-compare": 3.999999989900971e-06,
+      "emit-offloaded-dropout": 1.2999999853491317e-05,
+      "flatten-call-graph": 9.999999747378752e-06,
+      "fuse-send-recv": 2.499999936844688e-05,
+      "hilo::LegalizeAlias": 6.000000212225132e-06,
+      "hilo::NeuronInstCombine": 4.3000000005122274e-05,
+      "hilo::NeuronOpFusion": 9.000000318337698e-06,
+      "hilo::ReplaceTokenTypeWithU8Pass": 1.5999999959603883e-05,
+      "hilo::ScheduleFusion": 9.999999974752427e-07,
+      "hilo::SixtyFourHack": 1.4999999621068127e-05,
+      "hilo::VerifyAliasing": 1.9999999949504854e-06,
+      "hlo-mac-count": 2.5999999706982635e-05,
+      "hlo-verifier": 0.0001939999929163605,
+      "legalize-ccops": 1.9999999949504854e-06,
+      "legalize-compare": 3.999999989900971e-06,
+      "lower-argminmax-custom-call": 3.999999989900971e-06,
+      "map-inline": 1.2000000424450263e-05,
+      "metadata-naming": 1.8000000636675395e-05,
+      "mlir::detail::OpToOpPassAdaptor": 1.9999999494757503e-05,
+      "mlir::hlo::MhloToPyPenguin": 0.0009980000322684646,
+      "mlir::mhlo::LowerComplexExtraPass": 8.800000068731606e-05,
+      "mlir::mhlo::LowerComplexPass": 0.00015999999595806003,
+      "native-to-custom-softmax": 7.000000096013537e-06,
+      "native-to-custom-softmax-dx": 1.2999999853491317e-05,
+      "operand_upcaster": 1.9999999494757503e-05,
+      "post-par-pipe-begin": 8.900000102585182e-05,
+      "post-par-pipe-end": 0.0,
+      "post-partition-simplification": 0.000582000007852912,
+      "replace-minimum-constant": 7.000000096013537e-06,
+      "reshape-mover": 3.999999989900971e-06,
+      "simplify-concat": 4.8000001697801054e-05,
+      "simplify-while-loops": 1.9999999949504854e-06,
+      "transform-variadic-reduce": 9.000000318337698e-06,
+      "tuple-simplifier": 4.999999873689376e-06,
+      "unpack-nested-aws-ntwsr": 3.999999989900971e-06,
+      "unroll-while-loop": 0.0
+    },
+    "hilo": {
+      "ArithmeticIntensity": 16.6773738861084,
+      "ConstantSize": 599333.0,
+      "HloInputCount": 475.0,
+      "HloMacCount": 5637144576.0,
+      "HloOutputCount": 73.0,
+      "IfmapSize": 8266543104.0,
+      "OfmapSize": 75497472.0,
+      "OutputsReadFromCount": 0.0,
+      "PassthroughTensorsCount": 0.0,
+      "RedundantOutputCount": 0.0,
+      "Traffic": 676023104.0
+    }
+  },
+  "sg0000": {
+    "compiletime": {
+      "AGOrderingAnalysisPass": 0.08161520957946777,
+      "AffinePredicateResolution": 0.001527547836303711,
+      "AliasDependencyElimination": 0.00012493133544921875,
+      "AliasDependencyInduction": 0.008615970611572266,
+      "AliasDependencyReset": 0.03425288200378418,
+      "BFComputeCutting": 0.003037691116333008,
+      "BirCodeGenLoop": 0.05175900459289551,
+      "CCOpFusion": 0.024791479110717773,
+      "CanonicalizeDAGForPGTiling": 0.003105640411376953,
+      "CanonicalizeIR": 0.0020570755004882813,
+      "CoalesceCCOp": 0.005420684814453125,
+      "CommuteConcat": 0.0015554428100585938,
+      "DMALocalityOpt": 0.0025992393493652344,
+      "DMAProfiler": 0.004426240921020508,
+      "DMATilingProfiler": 0.00414586067199707,
+      "DataLocalityOpt": 0.11810016632080078,
+      "DataStreaming": 0.0053942203521728516,
+      "DeConcat": 0.0011267662048339844,
+      "DeadCodeElimination": 0.0016050338745117188,
+      "DeadStoreElimination": 0.030996084213256836,
+      "DelinearIndices": 0.007958412170410156,
+      "Delinearization": 0.003355741500854492,
+      "DoNothing": 7.987022399902344e-05,
+      "DramToDramTranspose": 0.03346753120422363,
+      "DumpGraphAndMetadata": 0.005443096160888672,
+      "EliminateDivs": 0.004342555999755859,
+      "ExpandBatchNorm": 0.0018055438995361328,
+      "ExpandISAMacro": 0.003648519515991211,
+      "FactorizeBlkDims": 0.019720077514648438,
+      "FactorizeThreadAxesInFreeDims": 0.0019965171813964844,
+      "FlattenMacroLoop": 0.003274679183959961,
+      "GenericAccessSimplifier": 0.0009877681732177734,
+      "InferInitValue": 0.032111167907714844,
+      "InferIntrinsicOnCC": 0.014227867126464844,
+      "InferNeuronTensor": 0.04684329032897949,
+      "InferNonlocalTensors": 0.10579586029052734,
+      "InferPSumTensor": 0.04808926582336426,
+      "InlineNativeKernels": 0.0025835037231445313,
+      "InsertIOTransposes": 0.012038707733154297,
+      "InsertLocalTransposes": 0.007574796676635742,
+      "InsertOffloadedTransposes": 0.003882884979248047,
+      "LICM": 0.003116607666015625,
+      "LateLegalizeInst": 0.006630420684814453,
+      "LateLegalizePostSplit": 0.0030584335327148438,
+      "LateLowerReshapeOp": 0.002176046371459961,
+      "LateLowerTensorOp": 0.005063295364379883,
+      "LateNeuronInstComb": 0.024392366409301758,
+      "LayoutPreprocessing": 0.03173065185546875,
+      "LayoutPreprocessingAndAnalysis": 0.07484269142150879,
+      "LayoutRequirementAnalysis": 0.007186174392700195,
+      "LegalizeCCOpLayout": 0.003088235855102539,
+      "LegalizeOpLevelAlias": 0.0011813640594482422,
+      "LegalizePartitionReduce": 0.0013763904571533203,
+      "LegalizeSundaAccess": 0.04270172119140625,
+      "LegalizeSundaMacro": 0.009444236755371094,
+      "LegalizeType": 0.004534721374511719,
+      "LocalLayoutOpt": 0.01777815818786621,
+      "LoopFusion": 0.0060007572174072266,
+      "LoopSplitting": 0.000377655029296875,
+      "LowerBroadcast": 0.0016138553619384766,
+      "LowerCCOpBlockAxis": 0.004978179931640625,
+      "LowerComplexBroadcast": 0.0023903846740722656,
+      "LowerIntrinsics": 0.034012556076049805,
+      "LowerTensorOp": 0.01333928108215332,
+      "LowerTranspose": 0.011911869049072266,
+      "MacroGeneration": 0.07152104377746582,
+      "MaskPropagation": 0.004988193511962891,
+      "MemcpyElimination": 0.11162376403808594,
+      "MutateDataType": 0.0014476776123046875,
+      "NeuronAliasDependencyInduction": 0.0002269744873046875,
+      "NeuronAliasDependencyReset": 0.15035724639892578,
+      "NeuronInstComb": 0.015686750411987305,
+      "NeuronLICM": 0.011453866958618164,
+      "NeuronLoopFusion": 0.018696069717407227,
+      "NeuronLoopInterchange": 0.0018415451049804688,
+      "NeuronSimplifier": 0.011624336242675781,
+      "NeuronSimplifyPredicates": 0.005795955657958984,
+      "NeuronValueNumbering": 0.0040967464447021484,
+      "OptimizeAliasedCopyChain": 0.0014064311981201172,
+      "OptimizeNKIKernels": 0.0021300315856933594,
+      "PAGLayoutOpt": 0.33215951919555664,
+      "PComputeCutting": 0.008408308029174805,
+      "PGLayoutTilingPipeline": 1.3294909000396729,
+      "PGTiling": 0.3412203788757324,
+      "PadElimination": 0.0018661022186279297,
+      "ParAxesAnnotation": 0.29718852043151855,
+      "PartialLoopFusion": 0.024113893508911133,
+      "PartialSimdFusion": 0.029590368270874023,
+      "PerfectLoopNest": 0.0021219253540039063,
+      "RecognizeOpIdiom": 0.004444122314453125,
+      "Recompute": 0.00028204917907714844,
+      "RelaxPredicates": 0.004793405532836914,
+      "Rematerialization": 0.004267692565917969,
+      "ReshapeWeights": 0.0014717578887939453,
+      "ResolveAccessConflict": 0.0038602352142333984,
+      "ResolveComplicatePredicates": 0.001505136489868164,
+      "RewriteReplicationMatmul": 0.0020885467529296875,
+      "RewriteWeights": 0.003512144088745117,
+      "SFKVectorizer": 0.3296499252319336,
+      "SimpleAllReduceTiling": 0.002294301986694336,
+      "Simplifier": 0.004443168640136719,
+      "SimplifyMacroPredicates": 0.013223648071289063,
+      "SimplifyNeuronTensor": 0.011357307434082031,
+      "SimplifySlice": 0.0010068416595458984,
+      "SimplifyTensor": 0.006380319595336914,
+      "SpillPSum": 0.018645763397216797,
+      "SplitAPUnionSets": 0.031983375549316406,
+      "SplitAccGrp": 0.0017464160919189453,
+      "StaticProfiler": 0.004789590835571289,
+      "StaticTransposeLocalTensor": 0.0048563480377197266,
+      "SundaISel": 0.046004533767700195,
+      "TCTransform": 0.0017864704132080078,
+      "TensorInitialization": 0.015267372131347656,
+      "TensorOpSimplifier": 0.006502866744995117,
+      "TensorOpTransform": 0.029101848602294922,
+      "TileCCOps": 0.0055658817291259766,
+      "TilingProfiler": 0.014283895492553711,
+      "TransformConvOp": 0.0028002262115478516,
+      "TritiumFusion": 0.037850379943847656,
+      "ValueNumbering": 0.002534627914428711,
+      "VectorizeDMA": 0.0056002140045166016,
+      "VectorizeMatMult": 0.004069805145263672,
+      "WeightCoalescing": 0.0033059120178222656,
+      "ZeroSizeTensorElimination": 0.00012040138244628906
+    },
+    "tensorizer": {
+      "DMATilingProfiler::TotalInstructionsAfterTiling": 1945.0,
+      "StaticProfiler::AifUb": 18.54642677307129,
+      "StaticProfiler::ArithmeticIntensityTensorizer": 234.4757080078125,
+      "StaticProfiler::AverageDmaLength": 3607.790283203125,
+      "StaticProfiler::AverageFractalPeUtilization": 99.84349822998047,
+      "StaticProfiler::AveragePartitionUtilization": 96.70350646972656,
+      "StaticProfiler::AveragePeUtilization": 99.51932525634766,
+      "StaticProfiler::DDRTransferBytes": 53226752.0,
+      "StaticProfiler::InternalTransferBytes": 27462656.0,
+      "StaticProfiler::LoadExpanded": 10244.0,
+      "StaticProfiler::LocalizationEfficiency": 1264.2635498046875,
+      "StaticProfiler::LocalizationEfficiencyIgnoreNonlocal": 1466.4949951171875,
+      "StaticProfiler::StoreExpanded": 3713.0,
+      "StaticProfiler::TotalDMAExpanded": 13957.0,
+      "StaticProfiler::TotalDynamicInstancesCount": 2107.0,
+      "StaticProfiler::TotalDynamicInstancesWithMmPackedCount": 2103.0,
+      "StaticProfiler::TotalLNCComm": 0.0,
+      "StaticProfiler::TotalLNCCommTransfer": 0.0,
+      "TilingProfiler::AveragePartitionUtilizationAfterTiling": 0.0,
+      "TilingProfiler::AveragePeUtilizationAfterTiling": 0.0,
+      "TilingProfiler::BatchnormInstructionsAfterTiling": 0.0,
+      "TilingProfiler::DmaInstructionsAfterTiling": 0.0,
+      "TilingProfiler::GenericInstructionsAfterTiling": 24.0,
+      "TilingProfiler::MatMultInstructionsAfterTiling": 1010.0,
+      "TilingProfiler::NumPfTransposes": 6.0,
+      "TilingProfiler::NumPfTransposesForIo": 0.0,
+      "TilingProfiler::NumPfTransposesForLocal": 5.0,
+      "TilingProfiler::NumPfTransposesForNonlocal": 1.0,
+      "TilingProfiler::PfTransposeInstructions": 176.0,
+      "TilingProfiler::PfTransposeInstructionsForIo": 0.0,
+      "TilingProfiler::PfTransposeInstructionsForLocal": 144.0,
+      "TilingProfiler::PfTransposeInstructionsForNonlocal": 32.0,
+      "TilingProfiler::ReduceInstructionsAfterTiling": 0.0,
+      "TilingProfiler::SimdInstructionsAfterTiling": 177.0,
+      "TilingProfiler::TotalInstructionsAfterTiling": 0.0,
+      "TransformConvOp::Conv1d_depthwise_bf01_oi01_bf01": 0.0,
+      "TransformConvOp::Conv2d_dw_fb01_io01_01bf_rep_nhwc_Pcinh": 0.0,
+      "TransformConvOp::Conv2d_pbp_0f1b_0i1o_01fb_experimental_1": 0.0,
+      "TransformConvOp::Conv2d_pbp_fb01_io01_01bf_experimental_1": 0.0,
+      "TransformConvOp::conv2d_column_packing": 0.0,
+      "TransformConvOp::conv2d_column_packing_1": 0.0,
+      "TransformConvOp::conv2d_column_packing_io10": 0.0,
+      "TransformConvOp::conv2d_depthwise_f01b_o01i_bf01": 0.0
+    }
+  },
+  "sg0001": {
+    "compiletime": {
+      "AGOrderingAnalysisPass": 0.035902976989746094,
+      "AffinePredicateResolution": 0.0021402835845947266,
+      "AliasDependencyElimination": 0.0001494884490966797,
+      "AliasDependencyInduction": 0.00843667984008789,
+      "AliasDependencyReset": 0.07715225219726563,
+      "BFComputeCutting": 0.002821207046508789,
+      "BirCodeGenLoop": 0.03489971160888672,
+      "CCOpFusion": 0.03669166564941406,
+      "CanonicalizeDAGForPGTiling": 0.0034394264221191406,
+      "CanonicalizeIR": 0.001888275146484375,
+      "CoalesceCCOp": 0.0048944950103759766,
+      "CommuteConcat": 0.001985788345336914,
+      "DMALocalityOpt": 0.0010595321655273438,
+      "DMAProfiler": 0.0038537979125976563,
+      "DMATilingProfiler": 0.0052776336669921875,
+      "DataLocalityOpt": 0.13663840293884277,
+      "DataStreaming": 0.004033327102661133,
+      "DeConcat": 0.0017592906951904297,
+      "DeadCodeElimination": 0.0027074813842773438,
+      "DeadStoreElimination": 0.03486442565917969,
+      "DelinearIndices": 0.010581493377685547,
+      "Delinearization": 0.004877567291259766,
+      "DoNothing": 6.914138793945313e-05,
+      "DramToDramTranspose": 0.03982400894165039,
+      "DumpGraphAndMetadata": 0.004088640213012695,
+      "EliminateDivs": 0.0045583248138427734,
+      "ExpandBatchNorm": 0.0018122196197509766,
+      "ExpandISAMacro": 0.0023725032806396484,
+      "FactorizeBlkDims": 0.013248920440673828,
+      "FactorizeThreadAxesInFreeDims": 0.0023849010467529297,
+      "FlattenMacroLoop": 0.0036728382110595703,
+      "GenericAccessSimplifier": 0.0026085376739501953,
+      "InferInitValue": 0.038416147232055664,
+      "InferIntrinsicOnCC": 0.010096549987792969,
+      "InferNeuronTensor": 0.05150651931762695,
+      "InferNonlocalTensors": 0.031507015228271484,
+      "InferPSumTensor": 0.03166079521179199,
+      "InlineNativeKernels": 0.0021262168884277344,
+      "InsertIOTransposes": 0.022419452667236328,
+      "InsertLocalTransposes": 0.0071408748626708984,
+      "InsertOffloadedTransposes": 0.0034465789794921875,
+      "LICM": 0.004317283630371094,
+      "LateLegalizeInst": 0.004563570022583008,
+      "LateLegalizePostSplit": 0.0027570724487304688,
+      "LateLowerReshapeOp": 0.0013232231140136719,
+      "LateLowerTensorOp": 0.004618406295776367,
+      "LateNeuronInstComb": 0.020873546600341797,
+      "LayoutPreprocessing": 0.037287235260009766,
+      "LayoutPreprocessingAndAnalysis": 0.10860347747802734,
+      "LayoutRequirementAnalysis": 0.007799863815307617,
+      "LegalizeCCOpLayout": 0.001935720443725586,
+      "LegalizeOpLevelAlias": 0.0012698173522949219,
+      "LegalizePartitionReduce": 0.002346515655517578,
+      "LegalizeSundaAccess": 0.016484975814819336,
+      "LegalizeSundaMacro": 0.011503934860229492,
+      "LegalizeType": 0.0047261714935302734,
+      "LocalLayoutOpt": 0.02424001693725586,
+      "LoopFusion": 0.007829427719116211,
+      "LoopSplitting": 0.00044846534729003906,
+      "LowerBroadcast": 0.0014789104461669922,
+      "LowerCCOpBlockAxis": 0.0059947967529296875,
+      "LowerComplexBroadcast": 0.0023598670959472656,
+      "LowerIntrinsics": 0.035590410232543945,
+      "LowerTensorOp": 0.012118339538574219,
+      "LowerTranspose": 0.011335611343383789,
+      "MacroGeneration": 0.11938071250915527,
+      "MaskPropagation": 0.003367900848388672,
+      "MemcpyElimination": 0.10591435432434082,
+      "MutateDataType": 0.002183198928833008,
+      "NeuronAliasDependencyInduction": 0.0002372264862060547,
+      "NeuronAliasDependencyReset": 0.02314162254333496,
+      "NeuronInstComb": 0.01471090316772461,
+      "NeuronLICM": 0.007970094680786133,
+      "NeuronLoopFusion": 0.022555112838745117,
+      "NeuronLoopInterchange": 0.0015497207641601563,
+      "NeuronSimplifier": 0.012836694717407227,
+      "NeuronSimplifyPredicates": 0.001605987548828125,
+      "NeuronValueNumbering": 0.0046231746673583984,
+      "OptimizeAliasedCopyChain": 0.00162506103515625,
+      "OptimizeNKIKernels": 0.0015685558319091797,
+      "PAGLayoutOpt": 0.14427471160888672,
+      "PComputeCutting": 0.00727081298828125,
+      "PGLayoutTilingPipeline": 1.2423913478851318,
+      "PGTiling": 0.5181164741516113,
+      "PadElimination": 0.00038051605224609375,
+      "ParAxesAnnotation": 0.09470343589782715,
+      "PartialLoopFusion": 0.018784761428833008,
+      "PartialSimdFusion": 0.027338027954101563,
+      "PerfectLoopNest": 0.0021829605102539063,
+      "RecognizeOpIdiom": 0.0048656463623046875,
+      "Recompute": 0.0002601146697998047,
+      "RelaxPredicates": 0.0033593177795410156,
+      "Rematerialization": 0.0023822784423828125,
+      "ReshapeWeights": 0.0014538764953613281,
+      "ResolveAccessConflict": 0.0047032833099365234,
+      "ResolveComplicatePredicates": 0.0019354820251464844,
+      "RewriteReplicationMatmul": 0.002605438232421875,
+      "RewriteWeights": 0.004354715347290039,
+      "SFKVectorizer": 0.16805624961853027,
+      "SimpleAllReduceTiling": 0.0025529861450195313,
+      "Simplifier": 0.00439763069152832,
+      "SimplifyMacroPredicates": 0.007683992385864258,
+      "SimplifyNeuronTensor": 0.0066149234771728516,
+      "SimplifySlice": 0.0023670196533203125,
+      "SimplifyTensor": 0.0063228607177734375,
+      "SpillPSum": 0.01709151268005371,
+      "SplitAPUnionSets": 0.018975019454956055,
+      "SplitAccGrp": 0.002074003219604492,
+      "StaticProfiler": 0.0037796497344970703,
+      "StaticTransposeLocalTensor": 0.005953311920166016,
+      "SundaISel": 0.0426335334777832,
+      "TCTransform": 0.0011513233184814453,
+      "TensorInitialization": 0.002532958984375,
+      "TensorOpSimplifier": 0.006600379943847656,
+      "TensorOpTransform": 0.034122467041015625,
+      "TileCCOps": 0.0059397220611572266,
+      "TilingProfiler": 0.013670921325683594,
+      "TransformConvOp": 0.002622365951538086,
+      "TritiumFusion": 0.05379676818847656,
+      "ValueNumbering": 0.0030698776245117188,
+      "VectorizeDMA": 0.0016117095947265625,
+      "VectorizeMatMult": 0.005866289138793945,
+      "WeightCoalescing": 0.0026290416717529297,
+      "ZeroSizeTensorElimination": 0.00011897087097167969
+    },
+    "tensorizer": {
+      "DMATilingProfiler::TotalInstructionsAfterTiling": 6049.0,
+      "StaticProfiler::AifUb": 251.7889862060547,
+      "StaticProfiler::ArithmeticIntensityTensorizer": 253.54466247558594,
+      "StaticProfiler::AverageDmaLength": 6385.9599609375,
+      "StaticProfiler::AverageFractalPeUtilization": 100.0,
+      "StaticProfiler::AveragePartitionUtilization": 99.86996459960938,
+      "StaticProfiler::AveragePeUtilization": 100.0,
+      "StaticProfiler::DDRTransferBytes": 204350464.0,
+      "StaticProfiler::InternalTransferBytes": 21430272.0,
+      "StaticProfiler::LoadExpanded": 27520.0,
+      "StaticProfiler::LocalizationEfficiency": 100.69728088378906,
+      "StaticProfiler::LocalizationEfficiencyIgnoreNonlocal": 105.00786590576172,
+      "StaticProfiler::StoreExpanded": 2305.0,
+      "StaticProfiler::TotalDMAExpanded": 29825.0,
+      "StaticProfiler::TotalDynamicInstancesCount": 6153.0,
+      "StaticProfiler::TotalDynamicInstancesWithMmPackedCount": 6153.0,
+      "StaticProfiler::TotalLNCComm": 0.0,
+      "StaticProfiler::TotalLNCCommTransfer": 0.0,
+      "TilingProfiler::AveragePartitionUtilizationAfterTiling": 0.0,
+      "TilingProfiler::AveragePeUtilizationAfterTiling": 0.0,
+      "TilingProfiler::BatchnormInstructionsAfterTiling": 0.0,
+      "TilingProfiler::DmaInstructionsAfterTiling": 0.0,
+      "TilingProfiler::GenericInstructionsAfterTiling": 16.0,
+      "TilingProfiler::MatMultInstructionsAfterTiling": 4848.0,
+      "TilingProfiler::NumPfTransposes": 8.0,
+      "TilingProfiler::NumPfTransposesForIo": 3.0,
+      "TilingProfiler::NumPfTransposesForLocal": 3.0,
+      "TilingProfiler::NumPfTransposesForNonlocal": 2.0,
+      "TilingProfiler::PfTransposeInstructions": 276.0,
+      "TilingProfiler::PfTransposeInstructionsForIo": 68.0,
+      "TilingProfiler::PfTransposeInstructionsForLocal": 80.0,
+      "TilingProfiler::PfTransposeInstructionsForNonlocal": 128.0,
+      "TilingProfiler::ReduceInstructionsAfterTiling": 0.0,
+      "TilingProfiler::SimdInstructionsAfterTiling": 216.0,
+      "TilingProfiler::TotalInstructionsAfterTiling": 0.0,
+      "TransformConvOp::Conv1d_depthwise_bf01_oi01_bf01": 0.0,
+      "TransformConvOp::Conv2d_dw_fb01_io01_01bf_rep_nhwc_Pcinh": 0.0,
+      "TransformConvOp::Conv2d_pbp_0f1b_0i1o_01fb_experimental_1": 0.0,
+      "TransformConvOp::Conv2d_pbp_fb01_io01_01bf_experimental_1": 0.0,
+      "TransformConvOp::conv2d_column_packing": 0.0,
+      "TransformConvOp::conv2d_column_packing_1": 0.0,
+      "TransformConvOp::conv2d_column_packing_io10": 0.0,
+      "TransformConvOp::conv2d_depthwise_f01b_o01i_bf01": 0.0
+    }
+  },
+  "sg0002": {
+    "compiletime": {
+      "AGOrderingAnalysisPass": 0.019578933715820313,
+      "AffinePredicateResolution": 0.0019481182098388672,
+      "AliasDependencyElimination": 0.0001239776611328125,
+      "AliasDependencyInduction": 0.00577092170715332,
+      "AliasDependencyReset": 0.027690649032592773,
+      "BFComputeCutting": 0.0023322105407714844,
+      "BirCodeGenLoop": 0.4628438949584961,
+      "CCOpFusion": 0.022275209426879883,
+      "CanonicalizeDAGForPGTiling": 0.005593061447143555,
+      "CanonicalizeIR": 0.001634359359741211,
+      "CoalesceCCOp": 0.015367984771728516,
+      "CommuteConcat": 0.0008616447448730469,
+      "DMALocalityOpt": 0.007138729095458984,
+      "DMAProfiler": 0.011677265167236328,
+      "DMATilingProfiler": 0.0037431716918945313,
+      "DataLocalityOpt": 0.06741714477539063,
+      "DataStreaming": 0.03589940071105957,
+      "DeConcat": 0.0005049705505371094,
+      "DeadCodeElimination": 0.0009002685546875,
+      "DeadStoreElimination": 0.0056514739990234375,
+      "DelinearIndices": 0.004773139953613281,
+      "Delinearization": 0.0026137828826904297,
+      "DoNothing": 7.462501525878906e-05,
+      "DramToDramTranspose": 0.019293546676635742,
+      "DumpGraphAndMetadata": 0.10360383987426758,
+      "EliminateDivs": 0.003831148147583008,
+      "ExpandBatchNorm": 0.0019576549530029297,
+      "ExpandISAMacro": 0.011517524719238281,
+      "FactorizeBlkDims": 0.008472919464111328,
+      "FactorizeThreadAxesInFreeDims": 0.001847982406616211,
+      "FlattenMacroLoop": 0.003529787063598633,
+      "GenericAccessSimplifier": 0.0008223056793212891,
+      "InferInitValue": 0.025947093963623047,
+      "InferIntrinsicOnCC": 0.00908350944519043,
+      "InferNeuronTensor": 0.02371978759765625,
+      "InferNonlocalTensors": 0.014753341674804688,
+      "InferPSumTensor": 0.3085360527038574,
+      "InlineNativeKernels": 0.008690595626831055,
+      "InsertIOTransposes": 0.01906275749206543,
+      "InsertLocalTransposes": 0.004312276840209961,
+      "InsertOffloadedTransposes": 0.002802133560180664,
+      "LICM": 0.003081083297729492,
+      "LateLegalizeInst": 0.01367807388305664,
+      "LateLegalizePostSplit": 0.012533903121948242,
+      "LateLowerReshapeOp": 0.001035451889038086,
+      "LateLowerTensorOp": 0.002605438232421875,
+      "LateNeuronInstComb": 0.008839130401611328,
+      "LayoutPreprocessing": 0.03434133529663086,
+      "LayoutPreprocessingAndAnalysis": 0.07319903373718262,
+      "LayoutRequirementAnalysis": 0.005194187164306641,
+      "LegalizeCCOpLayout": 0.0025322437286376953,
+      "LegalizeOpLevelAlias": 0.0020308494567871094,
+      "LegalizePartitionReduce": 0.0010001659393310547,
+      "LegalizeSundaAccess": 0.07694768905639648,
+      "LegalizeSundaMacro": 0.011176109313964844,
+      "LegalizeType": 0.014355182647705078,
+      "LocalLayoutOpt": 0.014019250869750977,
+      "LoopFusion": 0.005472898483276367,
+      "LoopSplitting": 0.00038623809814453125,
+      "LowerBroadcast": 0.0025022029876708984,
+      "LowerCCOpBlockAxis": 0.0058476924896240234,
+      "LowerComplexBroadcast": 0.00213623046875,
+      "LowerIntrinsics": 0.30684900283813477,
+      "LowerTensorOp": 0.010679960250854492,
+      "LowerTranspose": 0.012311697006225586,
+      "MacroGeneration": 0.029733657836914063,
+      "MaskPropagation": 0.0028328895568847656,
+      "MemcpyElimination": 0.026583433151245117,
+      "MutateDataType": 0.0020093917846679688,
+      "NeuronAliasDependencyInduction": 0.00018548965454101563,
+      "NeuronAliasDependencyReset": 0.02524423599243164,
+      "NeuronInstComb": 0.003789663314819336,
+      "NeuronLICM": 0.03511476516723633,
+      "NeuronLoopFusion": 0.007987260818481445,
+      "NeuronLoopInterchange": 0.0023233890533447266,
+      "NeuronSimplifier": 0.0075054168701171875,
+      "NeuronSimplifyPredicates": 0.11913681030273438,
+      "NeuronValueNumbering": 0.0033991336822509766,
+      "OptimizeAliasedCopyChain": 0.0005936622619628906,
+      "OptimizeNKIKernels": 0.44962644577026367,
+      "PAGLayoutOpt": 0.0999138355255127,
+      "PComputeCutting": 0.005170106887817383,
+      "PGLayoutTilingPipeline": 0.7408750057220459,
+      "PGTiling": 0.29245758056640625,
+      "PadElimination": 0.000308990478515625,
+      "ParAxesAnnotation": 0.05283546447753906,
+      "PartialLoopFusion": 0.0043125152587890625,
+      "PartialSimdFusion": 0.004901885986328125,
+      "PerfectLoopNest": 0.001722574234008789,
+      "RecognizeOpIdiom": 0.004076480865478516,
+      "Recompute": 0.0002620220184326172,
+      "RelaxPredicates": 0.013286828994750977,
+      "Rematerialization": 0.0021238327026367188,
+      "ReshapeWeights": 0.0006799697875976563,
+      "ResolveAccessConflict": 0.0040090084075927734,
+      "ResolveComplicatePredicates": 0.001981496810913086,
+      "RewriteReplicationMatmul": 0.0021796226501464844,
+      "RewriteWeights": 0.0022602081298828125,
+      "SFKVectorizer": 0.27124762535095215,
+      "SimpleAllReduceTiling": 0.00896596908569336,
+      "Simplifier": 0.0046122074127197266,
+      "SimplifyMacroPredicates": 0.010458230972290039,
+      "SimplifyNeuronTensor": 1.0512049198150635,
+      "SimplifySlice": 0.0009145736694335938,
+      "SimplifyTensor": 0.00577855110168457,
+      "SpillPSum": 0.012126922607421875,
+      "SplitAPUnionSets": 0.10518908500671387,
+      "SplitAccGrp": 0.001172780990600586,
+      "StaticProfiler": 0.0124053955078125,
+      "StaticTransposeLocalTensor": 0.0038576126098632813,
+      "SundaISel": 0.04396390914916992,
+      "TCTransform": 0.0018804073333740234,
+      "TensorInitialization": 0.012793779373168945,
+      "TensorOpSimplifier": 0.0045316219329833984,
+      "TensorOpTransform": 0.021115541458129883,
+      "TileCCOps": 0.0056231021881103516,
+      "TilingProfiler": 0.00790858268737793,
+      "TransformConvOp": 0.0030431747436523438,
+      "TritiumFusion": 0.03186154365539551,
+      "ValueNumbering": 0.0038623809814453125,
+      "VectorizeDMA": 0.0021522045135498047,
+      "VectorizeMatMult": 0.003453969955444336,
+      "WeightCoalescing": 0.009035825729370117,
+      "ZeroSizeTensorElimination": 0.00011420249938964844
+    },
+    "tensorizer": {
+      "DMATilingProfiler::TotalInstructionsAfterTiling": 43318.0,
+      "StaticProfiler::AifUb": 154.8094024658203,
+      "StaticProfiler::ArithmeticIntensityTensorizer": 152.92723083496094,
+      "StaticProfiler::AverageDmaLength": 4809.89794921875,
+      "StaticProfiler::AverageFractalPeUtilization": 99.65782165527344,
+      "StaticProfiler::AveragePartitionUtilization": 97.58238220214844,
+      "StaticProfiler::AveragePeUtilization": 98.61824035644531,
+      "StaticProfiler::DDRTransferBytes": 787141440.0,
+      "StaticProfiler::InternalTransferBytes": 634853888.0,
+      "StaticProfiler::LoadExpanded": 98070.0,
+      "StaticProfiler::LocalizationEfficiency": 98.78419494628906,
+      "StaticProfiler::LocalizationEfficiencyIgnoreNonlocal": 100.47209167480469,
+      "StaticProfiler::StoreExpanded": 2397.0,
+      "StaticProfiler::TotalDMAExpanded": 100467.0,
+      "StaticProfiler::TotalDynamicInstancesCount": 50670.0,
+      "StaticProfiler::TotalDynamicInstancesWithMmPackedCount": 50224.0,
+      "StaticProfiler::TotalLNCComm": 0.0,
+      "StaticProfiler::TotalLNCCommTransfer": 0.0,
+      "TilingProfiler::AveragePartitionUtilizationAfterTiling": 0.0,
+      "TilingProfiler::AveragePeUtilizationAfterTiling": 0.0,
+      "TilingProfiler::BatchnormInstructionsAfterTiling": 0.0,
+      "TilingProfiler::DmaInstructionsAfterTiling": 0.0,
+      "TilingProfiler::GenericInstructionsAfterTiling": 4.0,
+      "TilingProfiler::MatMultInstructionsAfterTiling": 22848.0,
+      "TilingProfiler::NumPfTransposes": 5.0,
+      "TilingProfiler::NumPfTransposesForIo": 1.0,
+      "TilingProfiler::NumPfTransposesForLocal": 1.0,
+      "TilingProfiler::NumPfTransposesForNonlocal": 3.0,
+      "TilingProfiler::PfTransposeInstructions": 19201.0,
+      "TilingProfiler::PfTransposeInstructionsForIo": 19008.0,
+      "TilingProfiler::PfTransposeInstructionsForLocal": 1.0,
+      "TilingProfiler::PfTransposeInstructionsForNonlocal": 192.0,
+      "TilingProfiler::ReduceInstructionsAfterTiling": 4.0,
+      "TilingProfiler::SimdInstructionsAfterTiling": 158.0,
+      "TilingProfiler::TotalInstructionsAfterTiling": 0.0,
+      "TransformConvOp::Conv1d_depthwise_bf01_oi01_bf01": 0.0,
+      "TransformConvOp::Conv2d_dw_fb01_io01_01bf_rep_nhwc_Pcinh": 0.0,
+      "TransformConvOp::Conv2d_pbp_0f1b_0i1o_01fb_experimental_1": 0.0,
+      "TransformConvOp::Conv2d_pbp_fb01_io01_01bf_experimental_1": 0.0,
+      "TransformConvOp::conv2d_column_packing": 0.0,
+      "TransformConvOp::conv2d_column_packing_1": 0.0,
+      "TransformConvOp::conv2d_column_packing_io10": 0.0,
+      "TransformConvOp::conv2d_depthwise_f01b_o01i_bf01": 0.0
+    }
+  },
+  "sg01": {
+    "compiletime": {
+      "CanonicalizeConv": 1.9999999949504854e-06,
+      "CanonicalizeForTensorizer": 1.4999999621068127e-05,
+      "Canonicalizer": 0.0002589999930933118,
+      "HoistCompute": 4.999999873689376e-06,
+      "IdentifyCrossPassTensors": 7.999999979801942e-06,
+      "MemcastMotion": 7.999999979801942e-06,
+      "PenguinizeFunctions": 1.5999999959603883e-05,
+      "PruneFunctions": 2.099999983329326e-05,
+      "RemoveOptimizationBarriers": 7.999999979801942e-06,
+      "ScatterMotion": 3.7999998312443495e-05,
+      "TensorizerLegalizationPass": 1.9999999494757503e-05,
+      "VerifySupportedOps": 9.999999747378752e-06,
+      "algsimp": 6.199999916134402e-05,
+      "batchnorm_expander": 1.2999999853491317e-05,
+      "boundary-marker-removal": 3.999999989900971e-06,
+      "call-inliner": 9.999999747378752e-06,
+      "canonicalize-boundary-marker": 6.000000212225132e-06,
+      "collective-stream-id-checker": 3.999999989900971e-06,
+      "comparison-expander": 4.999999873689376e-06,
+      "computation-deduplicator": 2.4000000848900527e-05,
+      "conditional-to-select": 4.999999873689376e-06,
+      "config-lowering": 2.099999983329326e-05,
+      "constant_folding": 7.999999979801942e-06,
+      "cse": 1.2000000424450263e-05,
+      "dce": 9.999999974752427e-07,
+      "dynamic-slice-transpose": 3.999999989900971e-06,
+      "eliminate-redundant-compare": 3.999999989900971e-06,
+      "emit-offloaded-dropout": 1.4000000192027073e-05,
+      "flatten-call-graph": 9.000000318337698e-06,
+      "fuse-send-recv": 2.9000000722589903e-05,
+      "hilo::LegalizeAlias": 4.999999873689376e-06,
+      "hilo::NeuronInstCombine": 4.5000000682193786e-05,
+      "hilo::NeuronOpFusion": 1.5999999959603883e-05,
+      "hilo::ReplaceTokenTypeWithU8Pass": 1.1000000085914508e-05,
+      "hilo::ScheduleFusion": 9.999999974752427e-07,
+      "hilo::SixtyFourHack": 1.4999999621068127e-05,
+      "hilo::VerifyAliasing": 1.9999999949504854e-06,
+      "hlo-mac-count": 2.5999999706982635e-05,
+      "hlo-verifier": 0.00020500000391621143,
+      "legalize-ccops": 9.999999974752427e-07,
+      "legalize-compare": 3.999999989900971e-06,
+      "lower-argminmax-custom-call": 3.999999989900971e-06,
+      "map-inline": 1.2000000424450263e-05,
+      "metadata-naming": 2.499999936844688e-05,
+      "mlir::detail::OpToOpPassAdaptor": 0.00012799999967683107,
+      "mlir::hlo::MhloToPyPenguin": 0.0009619999909773469,
+      "mlir::mhlo::LowerComplexExtraPass": 8.099999831756577e-05,
+      "mlir::mhlo::LowerComplexPass": 3.999999989900971e-06,
+      "native-to-custom-softmax": 6.000000212225132e-06,
+      "native-to-custom-softmax-dx": 1.5999999959603883e-05,
+      "operand_upcaster": 2.099999983329326e-05,
+      "post-par-pipe-begin": 4.999999873689376e-06,
+      "post-par-pipe-end": 0.0,
+      "post-partition-simplification": 0.0005779999773949385,
+      "replace-minimum-constant": 6.000000212225132e-06,
+      "reshape-mover": 3.000000106112566e-06,
+      "simplify-concat": 4.8999998398358e-05,
+      "simplify-while-loops": 1.9999999949504854e-06,
+      "transform-variadic-reduce": 9.000000318337698e-06,
+      "tuple-simplifier": 4.999999873689376e-06,
+      "unpack-nested-aws-ntwsr": 3.999999989900971e-06,
+      "unroll-while-loop": 0.0
+    },
+    "hilo": {
+      "ArithmeticIntensity": 240.22828674316406,
+      "HloMacCount": 24964497408.0,
+      "Traffic": 207839776.0
+    }
+  },
+  "sg02": {
+    "compiletime": {
+      "CanonicalizeConv": 0.0,
+      "CanonicalizeForTensorizer": 1.2999999853491317e-05,
+      "Canonicalizer": 0.000311999989207834,
+      "HoistCompute": 0.0,
+      "IdentifyCrossPassTensors": 1.2000000424450263e-05,
+      "MemcastMotion": 0.0,
+      "PenguinizeFunctions": 1.1000000085914508e-05,
+      "PruneFunctions": 7.999999979801942e-06,
+      "RemoveOptimizationBarriers": 1.2000000424450263e-05,
+      "ScatterMotion": 0.0,
+      "TensorizerLegalizationPass": 7.000000096013537e-06,
+      "VerifySupportedOps": 1.1000000085914508e-05,
+      "algsimp": 5.999999848427251e-05,
+      "batchnorm_expander": 1.2999999853491317e-05,
+      "boundary-marker-removal": 3.999999989900971e-06,
+      "call-inliner": 1.1000000085914508e-05,
+      "canonicalize-boundary-marker": 6.000000212225132e-06,
+      "collective-stream-id-checker": 3.000000106112566e-06,
+      "comparison-expander": 4.999999873689376e-06,
+      "computation-deduplicator": 2.499999936844688e-05,
+      "conditional-to-select": 7.000000096013537e-06,
+      "config-lowering": 2.5999999706982635e-05,
+      "constant_folding": 9.000000318337698e-06,
+      "cse": 1.2000000424450263e-05,
+      "dce": 9.999999974752427e-07,
+      "dynamic-slice-transpose": 3.999999989900971e-06,
+      "eliminate-redundant-compare": 3.999999989900971e-06,
+      "emit-offloaded-dropout": 1.4000000192027073e-05,
+      "flatten-call-graph": 1.2000000424450263e-05,
+      "fuse-send-recv": 1.8000000636675395e-05,
+      "hilo::LegalizeAlias": 1.9999999949504854e-06,
+      "hilo::NeuronInstCombine": 1.2999999853491317e-05,
+      "hilo::NeuronOpFusion": 0.0,
+      "hilo::ReplaceTokenTypeWithU8Pass": 1.5999999959603883e-05,
+      "hilo::ScheduleFusion": 0.0,
+      "hilo::SixtyFourHack": 5.900000178371556e-05,
+      "hilo::VerifyAliasing": 9.999999974752427e-07,
+      "hlo-mac-count": 0.0001720000000204891,
+      "hlo-verifier": 0.0001880000054370612,
+      "legalize-ccops": 9.999999974752427e-07,
+      "legalize-compare": 3.000000106112566e-06,
+      "lower-argminmax-custom-call": 3.999999989900971e-06,
+      "map-inline": 1.2000000424450263e-05,
+      "metadata-naming": 1.4999999621068127e-05,
+      "mlir::detail::OpToOpPassAdaptor": 1.9999999494757503e-05,
+      "mlir::hlo::MhloToPyPenguin": 0.0008660000166855752,
+      "mlir::mhlo::LowerComplexExtraPass": 9.100000170292333e-05,
+      "mlir::mhlo::LowerComplexPass": 0.00010599999950500205,
+      "native-to-custom-softmax": 6.000000212225132e-06,
+      "native-to-custom-softmax-dx": 2.300000051036477e-05,
+      "operand_upcaster": 2.300000051036477e-05,
+      "post-par-pipe-begin": 1.9999999949504854e-06,
+      "post-par-pipe-end": 0.0,
+      "post-partition-simplification": 0.0005329999839887023,
+      "replace-minimum-constant": 9.000000318337698e-06,
+      "reshape-mover": 3.000000106112566e-06,
+      "simplify-concat": 4.400000034365803e-05,
+      "simplify-while-loops": 1.9999999949504854e-06,
+      "transform-variadic-reduce": 6.0999998822808266e-05,
+      "tuple-simplifier": 4.999999873689376e-06,
+      "unpack-nested-aws-ntwsr": 3.999999989900971e-06,
+      "unroll-while-loop": 0.0
+    },
+    "hilo": {
+      "ArithmeticIntensity": 50.378170013427734,
+      "HloMacCount": 19638517760.0,
+      "Traffic": 779643968.0
+    }
+  }
+}

context_encoding_model/_tp0_bk1/graph.neff

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:60a8e4c285a690a146d149c675038f0498f62f761e4e3893706941d7ca8af583
+size 1659904

context_encoding_model/_tp0_bk1/metaneff.pb

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:24459c80d98d706b0a4aca22eda28ff6c09f03a08393e76b58ee0ca668d1b851
+size 1152551

context_encoding_model/_tp0_bk1/model.MODULE_2914133a46cb7b4660ab+d7af8a84.hlo_module.pb

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:b03debb723d63387ea26771f63729d616ac71a0dbfcb78d21d2194ff723fcbc1
+size 1229637

context_encoding_model/_tp0_bk1/model.MODULE_2914133a46cb7b4660ab+d7af8a84.neff

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:60a8e4c285a690a146d149c675038f0498f62f761e4e3893706941d7ca8af583
+size 1659904

context_encoding_model/_tp0_bk1/neuron_config.json

@@ -0,0 +1,220 @@
+{
+  "_attn_implementation_autoset": false,
+  "_name_or_path": "Qwen/Qwen3-8B",
+  "add_cross_attention": false,
+  "architectures": [
+    "Qwen3ForCausalLM"
+  ],
+  "attention_bias": false,
+  "attention_dropout": 0.0,
+  "attribute_map": {},
+  "bad_words_ids": null,
+  "begin_suppress_tokens": null,
+  "bos_token_id": 151643,
+  "chunk_size_feed_forward": 0,
+  "cross_attention_hidden_size": null,
+  "decoder_start_token_id": null,
+  "diversity_penalty": 0.0,
+  "do_sample": false,
+  "early_stopping": false,
+  "encoder_no_repeat_ngram_size": 0,
+  "eos_token_id": 151645,
+  "exponential_decay_length_penalty": null,
+  "finetuning_task": null,
+  "forced_bos_token_id": null,
+  "forced_eos_token_id": null,
+  "fused_spec_config": null,
+  "head_dim": 128,
+  "hidden_act": "silu",
+  "hidden_size": 4096,
+  "id2label": {
+    "0": "LABEL_0",
+    "1": "LABEL_1"
+  },
+  "initializer_range": 0.02,
+  "intermediate_size": 12288,
+  "is_decoder": false,
+  "is_encoder_decoder": false,
+  "label2id": {
+    "LABEL_0": 0,
+    "LABEL_1": 1
+  },
+  "length_penalty": 1.0,
+  "max_length": 20,
+  "max_position_embeddings": 40960,
+  "max_window_layers": 36,
+  "metadata": null,
+  "min_length": 0,
+  "model_type": "qwen3",
+  "neuron_config": {
+    "activation_quantization_type": null,
+    "allow_input_truncation": false,
+    "apply_seq_ids_mask": false,
+    "async_mode": false,
+    "attention_dp_degree": 1,
+    "attention_dtype": null,
+    "attn_block_cte_nki_kernel_enabled": false,
+    "attn_block_tkg_nki_kernel_cache_update": false,
+    "attn_block_tkg_nki_kernel_enabled": false,
+    "attn_cls": {
+      "__module__": "neuronx_distributed_inference.models.qwen3.modeling_qwen3",
+      "__name__": "NeuronQwen3Attention"
+    },
+    "attn_kernel_enabled": null,
+    "attn_tkg_builtin_kernel_enabled": false,
+    "attn_tkg_nki_kernel_enabled": false,
+    "batch_size": 1,
+    "bucket_n_active_tokens": true,
+    "buckets": [
+      256
+    ],
+    "cast_type": "config",
+    "cc_pipeline_tiling_factor": 2,
+    "chunked_prefill_config": null,
+    "context_encoding_buckets": [
+      256
+    ],
+    "cp_degree": 1,
+    "ctx_batch_size": 1,
+    "disable_kv_cache_tiling": false,
+    "draft_model_modules_to_not_convert": null,
+    "enable_bucketing": true,
+    "enable_eagle_draft_input_norm": false,
+    "enable_eagle_speculation": false,
+    "enable_fused_speculation": false,
+    "enable_long_context_mode": false,
+    "enable_output_completion_notifications": false,
+    "enable_spill_reload_dge": false,
+    "enable_token_tree": false,
+    "ep_degree": 1,
+    "expert_mlp_nki_kernel_enabled": null,
+    "flash_decoding_enabled": false,
+    "fused_qkv": false,
+    "fused_rmsnorm_skip_gamma": false,
+    "is_block_kv_layout": null,
+    "is_chunked_prefill": false,
+    "is_continuous_batching": true,
+    "is_eagle_draft": false,
+    "is_medusa": false,
+    "is_prefill_stage": true,
+    "is_prefix_caching": false,
+    "k_cache_transposed": false,
+    "kv_cache_batch_size": 1,
+    "kv_cache_padding_size": 0,
+    "kv_cache_quant": false,
+    "kv_cache_tiling": false,
+    "layer_boundary_markers": false,
+    "lm_head_pad": false,
+    "lm_head_pad_alignment_size": 1,
+    "local_ranks_size": 2,
+    "logical_nc_config": 1,
+    "lora_config": null,
+    "max_batch_size": 1,
+    "max_context_length": 1024,
+    "max_length": 1024,
+    "max_new_tokens": null,
+    "medusa_speculation_length": 0,
+    "medusa_tree": null,
+    "mlp_kernel_enabled": false,
+    "mlp_kernel_fuse_residual_add": false,
+    "modules_to_not_convert": null,
+    "moe_fused_nki_kernel_enabled": null,
+    "n_active_tokens": 1024,
+    "n_positions": 1024,
+    "num_medusa_heads": 0,
+    "on_cpu": false,
+    "on_device_sampling_config": {
+      "deterministic": false,
+      "do_sample": false,
+      "dynamic": true,
+      "global_topk": 256,
+      "on_device_sampling_config": true,
+      "temperature": 1.0,
+      "top_k": 1,
+      "top_k_kernel_enabled": false,
+      "top_p": 1.0
+    },
+    "output_logits": false,
+    "overrides_torch_dtype": true,
+    "pa_block_size": 1024,
+    "pa_num_blocks": 1,
+    "padding_side": "right",
+    "pp_degree": 1,
+    "prefix_buckets": null,
+    "qk_layernorm": false,
+    "qkv_kernel_enabled": false,
+    "qkv_kernel_fuse_residual_add": false,
+    "qkv_kernel_nbsd_layout": false,
+    "quantization_dtype": "int8",
+    "quantization_type": "per_tensor_symmetric",
+    "quantize_clamp_bound": Infinity,
+    "quantized": false,
+    "quantized_checkpoints_path": null,
+    "quantized_mlp_kernel_enabled": false,
+    "rmsnorm_quantize_kernel_enabled": false,
+    "router_topk_nki_kernel_enabled": null,
+    "rpl_reduce_dtype": null,
+    "save_sharded_checkpoint": true,
+    "scratchpad_page_size": null,
+    "seq_len": 1024,
+    "seq_len_threshold_for_cc_tiling": 16384,
+    "sequence_parallel_enabled": false,
+    "shared_mlp_nki_kernel_enabled": null,
+    "skip_sharding": false,
+    "skip_warmup": false,
+    "spec_batch_size": 1,
+    "speculation_length": 0,
+    "start_rank_id": 0,
+    "target": null,
+    "tile_cc": false,
+    "tkg_batch_size": 1,
+    "token_generation_buckets": null,
+    "token_tree_config": null,
+    "torch_dtype": "bfloat16",
+    "tp_degree": 2,
+    "vocab_parallel": false,
+    "weight_gather_seq_len_threshold": 32768,
+    "weights_to_skip_layout_optimization": [],
+    "world_size": 2
+  },
+  "no_repeat_ngram_size": 0,
+  "num_attention_heads": 32,
+  "num_beam_groups": 1,
+  "num_beams": 1,
+  "num_cores_per_group": 1,
+  "num_hidden_layers": 36,
+  "num_key_value_heads": 8,
+  "num_return_sequences": 1,
+  "output_attentions": false,
+  "output_hidden_states": false,
+  "output_scores": false,
+  "pad_token_id": 0,
+  "prefix": null,
+  "problem_type": null,
+  "pruned_heads": {},
+  "remove_invalid_values": false,
+  "repetition_penalty": 1.0,
+  "return_dict": true,
+  "return_dict_in_generate": false,
+  "rms_norm_eps": 1e-06,
+  "rope_scaling": null,
+  "rope_theta": 1000000,
+  "sep_token_id": null,
+  "sliding_window": null,
+  "suppress_tokens": null,
+  "task_specific_params": null,
+  "temperature": 1.0,
+  "tf_legacy_loss": false,
+  "tie_encoder_decoder": false,
+  "tie_word_embeddings": false,
+  "tokenizer_class": null,
+  "top_k": 50,
+  "top_p": 1.0,
+  "torchscript": false,
+  "transformers_version": "4.51.0",
+  "typical_p": 1.0,
+  "use_bfloat16": false,
+  "use_cache": true,
+  "use_sliding_window": false,
+  "vocab_size": 151936
+}

context_encoding_model/_tp0_bk2/command.txt

@@ -0,0 +1 @@
+neuronx-cc compile --framework=XLA model.MODULE_00594b8bc68e927f3dbe+1ad60ced.hlo_module.pb --output model.MODULE_00594b8bc68e927f3dbe+1ad60ced.neff --target=trn1 --auto-cast=none --model-type=transformer '--tensorizer-options=--enable-ccop-compute-overlap --cc-pipeline-tiling-factor=2 --vectorize-strided-dma ' --lnc=1 -O1 '--internal-hlo2tensorizer-options= --modular-flow-mac-threshold=10  --verify-hlo=true' --logfile=log-neuron-cc.txt --verbose=35

context_encoding_model/_tp0_bk2/compile_flags.MODULE_00594b8bc68e927f3dbe+1ad60ced.json

@@ -0,0 +1 @@
+["--target=trn1", "--auto-cast=none", "--model-type=transformer", "--tensorizer-options=--enable-ccop-compute-overlap --cc-pipeline-tiling-factor=2 --vectorize-strided-dma ", "--lnc=1", "-O1", "--internal-hlo2tensorizer-options= --modular-flow-mac-threshold=10  --verify-hlo=true", "--logfile=/home/ubuntu/qwen3/context_encoding_model/_tp0_bk2/log-neuron-cc.txt"]

context_encoding_model/_tp0_bk2/global_metric_store.json

@@ -0,0 +1,1079 @@
+{
+  "Average": {
+    "tensorizer": {
+      "StaticProfiler::AverageFractalPeUtilization": 99.66542053222656,
+      "StaticProfiler::AveragePartitionUtilization": 97.7269515991211,
+      "StaticProfiler::AveragePeUtilization": 98.64861297607422,
+      "StaticProfiler::LocalizationEfficiency": 98.26979064941406,
+      "StaticProfiler::LocalizationEfficiencyIgnoreNonlocal": 101.01405334472656,
+      "TilingProfiler::AveragePartitionUtilizationAfterTiling": 0.0,
+      "TilingProfiler::AveragePeUtilizationAfterTiling": 0.0
+    }
+  },
+  "Count": {
+    "tensorizer": {
+      "StaticProfiler::AverageFractalPeUtilization": 1.0,
+      "StaticProfiler::AveragePartitionUtilization": 1.0,
+      "StaticProfiler::AveragePeUtilization": 1.0,
+      "StaticProfiler::LocalizationEfficiency": 1.0,
+      "StaticProfiler::LocalizationEfficiencyIgnoreNonlocal": 1.0,
+      "TilingProfiler::AveragePartitionUtilizationAfterTiling": 1.0,
+      "TilingProfiler::AveragePeUtilizationAfterTiling": 1.0
+    }
+  },
+  "Sum": {
+    "compiletime": {
+      "AGOrderingAnalysisPass": 0.018257856369018555,
+      "AffinePredicateResolution": 0.0011677742004394531,
+      "AliasDependencyElimination": 0.0001201629638671875,
+      "AliasDependencyInduction": 0.0052988529205322266,
+      "AliasDependencyReset": 0.029210567474365234,
+      "BFComputeCutting": 0.0032625198364257813,
+      "BirCodeGenLoop": 0.4527714252471924,
+      "CCOpFusion": 0.02410125732421875,
+      "CanonicalizeConv": 0.00029399999766610563,
+      "CanonicalizeDAGForPGTiling": 0.004324913024902344,
+      "CanonicalizeForTensorizer": 4.8000001697801054e-05,
+      "CanonicalizeIR": 0.0019502639770507813,
+      "Canonicalizer": 0.0010809999657794833,
+      "CoalesceCCOp": 0.014672040939331055,
+      "CommuteConcat": 0.0008339881896972656,
+      "DMALocalityOpt": 0.005767107009887695,
+      "DMAProfiler": 0.012850046157836914,
+      "DMATilingProfiler": 0.004332065582275391,
+      "DataLocalityOpt": 0.07260942459106445,
+      "DataStreaming": 0.03969836235046387,
+      "DeConcat": 0.0005326271057128906,
+      "DeadCodeElimination": 0.0009255409240722656,
+      "DeadStoreElimination": 0.0055675506591796875,
+      "DelinearIndices": 0.004735231399536133,
+      "Delinearization": 0.0030374526977539063,
+      "DoNothing": 0.00018930435180664063,
+      "DramToDramTranspose": 0.018135547637939453,
+      "DumpGraphAndMetadata": 0.09476375579833984,
+      "EliminateDivs": 0.002595663070678711,
+      "ExpandBatchNorm": 0.002063274383544922,
+      "ExpandISAMacro": 0.011973381042480469,
+      "FactorizeBlkDims": 0.009292840957641602,
+      "FactorizeThreadAxesInFreeDims": 0.0010046958923339844,
+      "FlattenMacroLoop": 0.002232074737548828,
+      "GenericAccessSimplifier": 0.0018167495727539063,
+      "HoistCompute": 7.999999979801942e-06,
+      "IdentifyCrossPassTensors": 7.79999973019585e-05,
+      "InferInitValue": 0.024865150451660156,
+      "InferIntrinsicOnCC": 0.009101152420043945,
+      "InferNeuronTensor": 0.023293495178222656,
+      "InferNonlocalTensors": 0.01632833480834961,
+      "InferPSumTensor": 0.27726316452026367,
+      "InlineNativeKernels": 0.0081634521484375,
+      "InsertIOTransposes": 0.019203901290893555,
+      "InsertLocalTransposes": 0.0042340755462646484,
+      "InsertOffloadedTransposes": 0.002811431884765625,
+      "LICM": 0.0029730796813964844,
+      "LateLegalizeInst": 0.014307022094726563,
+      "LateLegalizePostSplit": 0.012536048889160156,
+      "LateLowerReshapeOp": 0.0018641948699951172,
+      "LateLowerTensorOp": 0.0014081001281738281,
+      "LateNeuronInstComb": 0.00915217399597168,
+      "LayoutPreprocessing": 0.02658390998840332,
+      "LayoutPreprocessingAndAnalysis": 0.10707235336303711,
+      "LayoutRequirementAnalysis": 0.005135536193847656,
+      "LegalizeCCOpLayout": 0.002307415008544922,
+      "LegalizeOpLevelAlias": 0.0012297630310058594,
+      "LegalizePartitionReduce": 0.0010194778442382813,
+      "LegalizeSundaAccess": 0.07808256149291992,
+      "LegalizeSundaMacro": 0.010968446731567383,
+      "LegalizeType": 0.012074947357177734,
+      "LocalLayoutOpt": 0.013799905776977539,
+      "LoopFusion": 0.0052182674407958984,
+      "LoopSplitting": 0.0003161430358886719,
+      "LowerBroadcast": 0.0015821456909179688,
+      "LowerCCOpBlockAxis": 0.0040547847747802734,
+      "LowerComplexBroadcast": 0.002165079116821289,
+      "LowerIntrinsics": 0.31156492233276367,
+      "LowerTensorOp": 0.010558843612670898,
+      "LowerTranspose": 0.012494325637817383,
+      "MacroGeneration": 0.029862642288208008,
+      "MaskPropagation": 0.002757549285888672,
+      "MemcastMotion": 3.400000059627928e-05,
+      "MemcpyElimination": 0.025969266891479492,
+      "MutateDataType": 0.002087831497192383,
+      "NeuronAliasDependencyInduction": 0.00016880035400390625,
+      "NeuronAliasDependencyReset": 0.020352602005004883,
+      "NeuronInstComb": 0.004656076431274414,
+      "NeuronLICM": 0.03560137748718262,
+      "NeuronLoopFusion": 0.007991313934326172,
+      "NeuronLoopInterchange": 0.002409219741821289,
+      "NeuronSimplifier": 0.007069587707519531,
+      "NeuronSimplifyPredicates": 0.12419009208679199,
+      "NeuronValueNumbering": 0.0032753944396972656,
+      "OptimizeAliasedCopyChain": 0.0005936622619628906,
+      "OptimizeNKIKernels": 0.5374257564544678,
+      "PAGLayoutOpt": 0.08115577697753906,
+      "PComputeCutting": 0.004801273345947266,
+      "PGLayoutTilingPipeline": 0.5454635620117188,
+      "PGTiling": 0.14933419227600098,
+      "PadElimination": 0.00034046173095703125,
+      "ParAxesAnnotation": 0.053552865982055664,
+      "PartialLoopFusion": 0.0067539215087890625,
+      "PartialSimdFusion": 0.00693058967590332,
+      "PenguinizeFunctions": 4.5000000682193786e-05,
+      "PerfectLoopNest": 0.0035321712493896484,
+      "PruneFunctions": 5.199999941396527e-05,
+      "RecognizeOpIdiom": 0.003947257995605469,
+      "Recompute": 0.00024962425231933594,
+      "RelaxPredicates": 0.013285398483276367,
+      "Rematerialization": 0.002062082290649414,
+      "RemoveOptimizationBarriers": 8.70000003487803e-05,
+      "ReshapeWeights": 0.002131223678588867,
+      "ResolveAccessConflict": 0.0038597583770751953,
+      "ResolveComplicatePredicates": 0.002032756805419922,
+      "RewriteReplicationMatmul": 0.001924753189086914,
+      "RewriteWeights": 0.002452373504638672,
+      "SFKVectorizer": 0.2718319892883301,
+      "ScatterMotion": 3.7999998312443495e-05,
+      "SimpleAllReduceTiling": 0.008960247039794922,
+      "Simplifier": 0.004038810729980469,
+      "SimplifyMacroPredicates": 0.010622739791870117,
+      "SimplifyNeuronTensor": 1.0594146251678467,
+      "SimplifySlice": 0.0009577274322509766,
+      "SimplifyTensor": 0.005341768264770508,
+      "SpillPSum": 0.012076139450073242,
+      "SplitAPUnionSets": 0.10771751403808594,
+      "SplitAccGrp": 0.002201557159423828,
+      "StaticProfiler": 0.012447118759155273,
+      "StaticTransposeLocalTensor": 0.0038712024688720703,
+      "SundaISel": 0.04214668273925781,
+      "TCTransform": 0.0008432865142822266,
+      "TensorInitialization": 0.012825727462768555,
+      "TensorOpSimplifier": 0.004651308059692383,
+      "TensorOpTransform": 0.019537687301635742,
+      "TensorizerLegalizationPass": 5.7999997807201e-05,
+      "TileCCOps": 0.006766319274902344,
+      "TilingProfiler": 0.006911277770996094,
+      "TransformConvOp": 0.0030303001403808594,
+      "TritiumFusion": 0.04502224922180176,
+      "ValueNumbering": 0.001996755599975586,
+      "VectorizeDMA": 0.0019402503967285156,
+      "VectorizeMatMult": 0.0027413368225097656,
+      "VerifySupportedOps": 3.7000001611886546e-05,
+      "WeightCoalescing": 0.008520841598510742,
+      "ZeroSizeTensorElimination": 0.00013709068298339844,
+      "algsimp": 0.0026940000243484974,
+      "batchnorm_expander": 4.400000034365803e-05,
+      "boundary-marker-removal": 1.5999999959603883e-05,
+      "call-inliner": 0.00046999999904073775,
+      "canonicalize-boundary-marker": 1.8999999156221747e-05,
+      "collective-stream-id-checker": 7.300000288523734e-05,
+      "comparison-expander": 0.0005740000051446259,
+      "computation-deduplicator": 7.999999797903001e-05,
+      "conditional-to-select": 1.8000000636675395e-05,
+      "config-lowering": 0.0003279999946244061,
+      "constant-statistics": 0.0005329999839887023,
+      "constant_folding": 0.0003260000084992498,
+      "cse": 4.5000000682193786e-05,
+      "dce": 8.399999933317304e-05,
+      "dot_decomposer": 0.0013409999664872885,
+      "dynamic-slice-transpose": 1.3999999282532372e-05,
+      "eliminate-redundant-compare": 0.0002959999837912619,
+      "emit-offloaded-dropout": 6.399999983841553e-05,
+      "flatten-call-graph": 0.0009319999953731894,
+      "fuse-send-recv": 6.999999459367245e-05,
+      "hilo::LegalizeAlias": 1.3999999282532372e-05,
+      "hilo::NeuronInstCombine": 0.0001660000125411898,
+      "hilo::NeuronOpFusion": 2.5000001187436283e-05,
+      "hilo::ReplaceTokenTypeWithU8Pass": 5.2999999752501026e-05,
+      "hilo::ScheduleFusion": 7.000000096013537e-06,
+      "hilo::SixtyFourHack": 7.299999560927972e-05,
+      "hilo::VerifyAliasing": 6.000000212225132e-06,
+      "hlo-mac-count": 0.0013429999817162752,
+      "hlo-verifier": 0.007542999926954508,
+      "instruction-histogram": 0.0006709999870508909,
+      "io-con-pipe-begin": 4.999999873689376e-06,
+      "io-con-pipe-end": 9.999999974752427e-07,
+      "io-layout-normalization": 0.001310999970883131,
+      "io-statistics": 8.499999967170879e-05,
+      "legalize-ccops": 3.999999989900971e-06,
+      "legalize-compare": 1.2999999853491317e-05,
+      "lower-argminmax-custom-call": 1.300000076298602e-05,
+      "map-inline": 0.0008850000449456275,
+      "metadata-naming": 5.999999848427251e-05,
+      "mlir::detail::OpToOpPassAdaptor": 0.00014399999054148793,
+      "mlir::hlo::MhloToPyPenguin": 0.004429999738931656,
+      "mlir::mhlo::LowerComplexExtraPass": 0.00027299998328089714,
+      "mlir::mhlo::LowerComplexPass": 0.0004909999552182853,
+      "native-to-custom-softmax": 0.0007070000283420086,
+      "native-to-custom-softmax-dx": 0.0005990000208839774,
+      "operand_upcaster": 4.900000203633681e-05,
+      "opt-barrier-removal": 0.0005510000046342611,
+      "post-par-pipe-begin": 8.999999408842996e-06,
+      "post-par-pipe-end": 0.0,
+      "post-partition-simplification": 0.0018570000538602471,
+      "pre-par-pipe-begin": 9.999999974752427e-07,
+      "pre-par-pipe-end": 0.0,
+      "pre-partition-simplification": 0.12893199920654297,
+      "replace-minimum-constant": 0.0004569999873638153,
+      "reshape-mover": 0.00012599999899975955,
+      "simplify-concat": 0.00015899998834356666,
+      "simplify-while-loops": 0.00010400000610388815,
+      "transform-variadic-reduce": 7.000000186963007e-05,
+      "tuple-simplifier": 0.0003150000120513141,
+      "unpack-nested-aws-ntwsr": 0.0004349999944679439,
+      "unroll-while-loop": 2.099999983329326e-05,
+      "zero_sized_hlo_elimination": 0.0008670000243000686
+    },
+    "hilo": {
+      "ConstantSize": 1189157.0,
+      "HloInputCount": 475.0,
+      "HloMacCount": 101242896384.0,
+      "HloOutputCount": 73.0,
+      "IfmapSize": 8266545152.0,
+      "OfmapSize": 75497472.0,
+      "OutputsReadFromCount": 0.0,
+      "PassthroughTensorsCount": 0.0,
+      "RedundantOutputCount": 0.0,
+      "Traffic": 1692493184.0
+    },
+    "tensorizer": {
+      "DMATilingProfiler::TotalInstructionsAfterTiling": 44382.0,
+      "StaticProfiler::AifUb": 205.154296875,
+      "StaticProfiler::ArithmeticIntensityTensorizer": 201.6046905517578,
+      "StaticProfiler::AverageDmaLength": 1901.806396484375,
+      "StaticProfiler::DDRTransferBytes": 795531072.0,
+      "StaticProfiler::InternalTransferBytes": 646388224.0,
+      "StaticProfiler::LoadExpanded": 376342.0,
+      "StaticProfiler::StoreExpanded": 4189.0,
+      "StaticProfiler::TotalDMAExpanded": 380531.0,
+      "StaticProfiler::TotalDynamicInstancesCount": 53882.0,
+      "StaticProfiler::TotalDynamicInstancesWithMmPackedCount": 53436.0,
+      "StaticProfiler::TotalLNCComm": 0.0,
+      "StaticProfiler::TotalLNCCommTransfer": 0.0,
+      "TilingProfiler::BatchnormInstructionsAfterTiling": 0.0,
+      "TilingProfiler::DmaInstructionsAfterTiling": 0.0,
+      "TilingProfiler::GenericInstructionsAfterTiling": 4.0,
+      "TilingProfiler::MatMultInstructionsAfterTiling": 23616.0,
+      "TilingProfiler::NumPfTransposes": 5.0,
+      "TilingProfiler::NumPfTransposesForIo": 1.0,
+      "TilingProfiler::NumPfTransposesForLocal": 1.0,
+      "TilingProfiler::NumPfTransposesForNonlocal": 3.0,
+      "TilingProfiler::PfTransposeInstructions": 19393.0,
+      "TilingProfiler::PfTransposeInstructionsForIo": 19008.0,
+      "TilingProfiler::PfTransposeInstructionsForLocal": 1.0,
+      "TilingProfiler::PfTransposeInstructionsForNonlocal": 384.0,
+      "TilingProfiler::ReduceInstructionsAfterTiling": 4.0,
+      "TilingProfiler::SimdInstructionsAfterTiling": 158.0,
+      "TilingProfiler::TotalInstructionsAfterTiling": 0.0,
+      "TransformConvOp::Conv1d_depthwise_bf01_oi01_bf01": 0.0,
+      "TransformConvOp::Conv2d_dw_fb01_io01_01bf_rep_nhwc_Pcinh": 0.0,
+      "TransformConvOp::Conv2d_pbp_0f1b_0i1o_01fb_experimental_1": 0.0,
+      "TransformConvOp::Conv2d_pbp_fb01_io01_01bf_experimental_1": 0.0,
+      "TransformConvOp::conv2d_column_packing": 0.0,
+      "TransformConvOp::conv2d_column_packing_1": 0.0,
+      "TransformConvOp::conv2d_column_packing_io10": 0.0,
+      "TransformConvOp::conv2d_depthwise_f01b_o01i_bf01": 0.0
+    }
+  },
+  "all": {
+    "compiletime": {
+      "algsimp": 0.002466999925673008,
+      "call-inliner": 0.0004360000020824373,
+      "collective-stream-id-checker": 6.299999949987978e-05,
+      "comparison-expander": 0.0005569999921135604,
+      "constant-statistics": 0.0005329999839887023,
+      "constant_folding": 0.0002969999914057553,
+      "dce": 7.999999797903001e-05,
+      "dot_decomposer": 0.0013409999664872885,
+      "eliminate-redundant-compare": 0.00028199999360367656,
+      "flatten-call-graph": 0.0008999999845400453,
+      "hlo-mac-count": 0.0010720000136643648,
+      "hlo-verifier": 0.0069679999724030495,
+      "instruction-histogram": 0.0006709999870508909,
+      "io-con-pipe-begin": 4.999999873689376e-06,
+      "io-con-pipe-end": 9.999999974752427e-07,
+      "io-layout-normalization": 0.001310999970883131,
+      "io-statistics": 8.499999967170879e-05,
+      "map-inline": 0.0008440000237897038,
+      "native-to-custom-softmax": 0.0006750000175088644,
+      "native-to-custom-softmax-dx": 0.0005000000237487257,
+      "opt-barrier-removal": 0.0005510000046342611,
+      "pre-par-pipe-begin": 9.999999974752427e-07,
+      "pre-par-pipe-end": 0.0,
+      "pre-partition-simplification": 0.12893199920654297,
+      "replace-minimum-constant": 0.0004309999931138009,
+      "reshape-mover": 0.00011500000255182385,
+      "simplify-while-loops": 9.600000339560211e-05,
+      "tuple-simplifier": 0.0002969999914057553,
+      "unpack-nested-aws-ntwsr": 0.00042100000428035855,
+      "unroll-while-loop": 1.9999999494757503e-05,
+      "zero_sized_hlo_elimination": 0.0008670000243000686
+    }
+  },
+  "cumsum": {
+    "compiletime": {
+      "CoalesceCCOp": 0.00020885467529296875,
+      "DMALocalityOpt": 0.00016832351684570313,
+      "DMAProfiler": 0.0007588863372802734,
+      "DataStreaming": 0.00029587745666503906,
+      "DoNothing": 0.00011897087097167969,
+      "ExpandISAMacro": 0.0005011558532714844,
+      "FactorizeBlkDims": 0.00043463706970214844,
+      "InferPSumTensor": 0.00044608116149902344,
+      "LateLegalizeInst": 0.0004031658172607422,
+      "LateNeuronInstComb": 0.0005033016204833984,
+      "LegalizeSundaAccess": 0.0021431446075439453,
+      "LegalizeType": 0.00024056434631347656,
+      "LowerBroadcast": 0.00022101402282714844,
+      "LowerIntrinsics": 0.00023508071899414063,
+      "LowerTranspose": 0.0002219676971435547,
+      "NeuronInstComb": 0.0005297660827636719,
+      "NeuronLICM": 0.00041484832763671875,
+      "NeuronSimplifyPredicates": 0.0028023719787597656,
+      "NeuronValueNumbering": 0.00043582916259765625,
+      "SFKVectorizer": 0.002759695053100586,
+      "SimpleAllReduceTiling": 0.00020432472229003906,
+      "SimplifyNeuronTensor": 0.0004029273986816406,
+      "SpillPSum": 0.0005388259887695313,
+      "WeightCoalescing": 0.0002307891845703125
+    }
+  },
+  "sg00": {
+    "compiletime": {
+      "CanonicalizeConv": 2.300000051036477e-05,
+      "CanonicalizeForTensorizer": 2.300000051036477e-05,
+      "Canonicalizer": 0.0005249999812804163,
+      "HoistCompute": 3.000000106112566e-06,
+      "IdentifyCrossPassTensors": 3.099999958067201e-05,
+      "MemcastMotion": 9.999999747378752e-06,
+      "PenguinizeFunctions": 2.2000000171829015e-05,
+      "PruneFunctions": 1.2999999853491317e-05,
+      "RemoveOptimizationBarriers": 4.400000034365803e-05,
+      "ScatterMotion": 6.000000212225132e-06,
+      "TensorizerLegalizationPass": 3.600000127335079e-05,
+      "VerifySupportedOps": 1.700000029813964e-05,
+      "algsimp": 0.0001049999991664663,
+      "batchnorm_expander": 1.8999999156221747e-05,
+      "boundary-marker-removal": 7.000000096013537e-06,
+      "call-inliner": 1.4000000192027073e-05,
+      "canonicalize-boundary-marker": 7.999999979801942e-06,
+      "collective-stream-id-checker": 3.999999989900971e-06,
+      "comparison-expander": 7.000000096013537e-06,
+      "computation-deduplicator": 2.099999983329326e-05,
+      "conditional-to-select": 7.000000096013537e-06,
+      "config-lowering": 0.00027600000612437725,
+      "constant_folding": 1.2000000424450263e-05,
+      "cse": 2.2000000171829015e-05,
+      "dce": 1.9999999949504854e-06,
+      "dynamic-slice-transpose": 6.000000212225132e-06,
+      "eliminate-redundant-compare": 6.000000212225132e-06,
+      "emit-offloaded-dropout": 3.7999998312443495e-05,
+      "flatten-call-graph": 1.2999999853491317e-05,
+      "fuse-send-recv": 3.099999958067201e-05,
+      "hilo::LegalizeAlias": 7.000000096013537e-06,
+      "hilo::NeuronInstCombine": 6.299999949987978e-05,
+      "hilo::NeuronOpFusion": 6.000000212225132e-06,
+      "hilo::ReplaceTokenTypeWithU8Pass": 2.300000051036477e-05,
+      "hilo::ScheduleFusion": 1.9999999949504854e-06,
+      "hilo::SixtyFourHack": 2.099999983329326e-05,
+      "hilo::VerifyAliasing": 3.000000106112566e-06,
+      "hlo-mac-count": 7.300000288523734e-05,
+      "hlo-verifier": 0.00023600000713486224,
+      "legalize-ccops": 1.9999999949504854e-06,
+      "legalize-compare": 6.000000212225132e-06,
+      "lower-argminmax-custom-call": 6.000000212225132e-06,
+      "map-inline": 1.700000029813964e-05,
+      "metadata-naming": 2.499999936844688e-05,
+      "mlir::detail::OpToOpPassAdaptor": 2.2000000171829015e-05,
+      "mlir::hlo::MhloToPyPenguin": 0.002633000025525689,
+      "mlir::mhlo::LowerComplexExtraPass": 0.0001049999991664663,
+      "mlir::mhlo::LowerComplexPass": 0.00017299999308306724,
+      "native-to-custom-softmax": 2.099999983329326e-05,
+      "native-to-custom-softmax-dx": 6.600000051548705e-05,
+      "operand_upcaster": 2.2000000171829015e-05,
+      "post-par-pipe-begin": 4.999999873689376e-06,
+      "post-par-pipe-end": 0.0,
+      "post-partition-simplification": 0.0008430000161752105,
+      "replace-minimum-constant": 1.1000000085914508e-05,
+      "reshape-mover": 4.999999873689376e-06,
+      "simplify-concat": 6.70000008540228e-05,
+      "simplify-while-loops": 3.999999989900971e-06,
+      "transform-variadic-reduce": 1.2999999853491317e-05,
+      "tuple-simplifier": 7.999999979801942e-06,
+      "unpack-nested-aws-ntwsr": 6.000000212225132e-06,
+      "unroll-while-loop": 9.999999974752427e-07
+    },
+    "hilo": {
+      "ArithmeticIntensity": 34.445003509521484,
+      "ConstantSize": 1189157.0,
+      "HloInputCount": 475.0,
+      "HloMacCount": 11811160064.0,
+      "HloOutputCount": 73.0,
+      "IfmapSize": 8266545152.0,
+      "OfmapSize": 75497472.0,
+      "OutputsReadFromCount": 0.0,
+      "PassthroughTensorsCount": 0.0,
+      "RedundantOutputCount": 0.0,
+      "Traffic": 685798208.0
+    }
+  },
+  "sg0000": {
+    "compiletime": {
+      "AGOrderingAnalysisPass": 0.07801461219787598,
+      "AffinePredicateResolution": 0.0017647743225097656,
+      "AliasDependencyElimination": 0.0001277923583984375,
+      "AliasDependencyInduction": 0.00855708122253418,
+      "AliasDependencyReset": 0.08457040786743164,
+      "BFComputeCutting": 0.003294229507446289,
+      "BirCodeGenLoop": 0.05274701118469238,
+      "CCOpFusion": 0.030017614364624023,
+      "CanonicalizeDAGForPGTiling": 0.003341197967529297,
+      "CanonicalizeIR": 0.0022792816162109375,
+      "CoalesceCCOp": 0.0053555965423583984,
+      "CommuteConcat": 0.0023560523986816406,
+      "DMALocalityOpt": 0.0013885498046875,
+      "DMAProfiler": 0.00625157356262207,
+      "DMATilingProfiler": 0.003763914108276367,
+      "DataLocalityOpt": 0.09786868095397949,
+      "DataStreaming": 0.004992246627807617,
+      "DeConcat": 0.002264261245727539,
+      "DeadCodeElimination": 0.002042531967163086,
+      "DeadStoreElimination": 0.030755043029785156,
+      "DelinearIndices": 0.009100914001464844,
+      "Delinearization": 0.004424571990966797,
+      "DoNothing": 6.914138793945313e-05,
+      "DramToDramTranspose": 0.03130936622619629,
+      "DumpGraphAndMetadata": 0.005283832550048828,
+      "EliminateDivs": 0.0042150020599365234,
+      "ExpandBatchNorm": 0.0019366741180419922,
+      "ExpandISAMacro": 0.002724170684814453,
+      "FactorizeBlkDims": 0.011873722076416016,
+      "FactorizeThreadAxesInFreeDims": 0.002283811569213867,
+      "FlattenMacroLoop": 0.0031974315643310547,
+      "GenericAccessSimplifier": 0.002216339111328125,
+      "InferInitValue": 0.030458927154541016,
+      "InferIntrinsicOnCC": 0.011402368545532227,
+      "InferNeuronTensor": 0.04513859748840332,
+      "InferNonlocalTensors": 0.10613727569580078,
+      "InferPSumTensor": 0.037427663803100586,
+      "InlineNativeKernels": 0.00368499755859375,
+      "InsertIOTransposes": 0.012629508972167969,
+      "InsertLocalTransposes": 0.007400989532470703,
+      "InsertOffloadedTransposes": 0.0025758743286132813,
+      "LICM": 0.0031554698944091797,
+      "LateLegalizeInst": 0.005858182907104492,
+      "LateLegalizePostSplit": 0.0029172897338867188,
+      "LateLowerReshapeOp": 0.0018696784973144531,
+      "LateLowerTensorOp": 0.004997968673706055,
+      "LateNeuronInstComb": 0.019808530807495117,
+      "LayoutPreprocessing": 0.04119300842285156,
+      "LayoutPreprocessingAndAnalysis": 0.10642147064208984,
+      "LayoutRequirementAnalysis": 0.0070705413818359375,
+      "LegalizeCCOpLayout": 0.004191398620605469,
+      "LegalizeOpLevelAlias": 0.0015521049499511719,
+      "LegalizePartitionReduce": 0.002257108688354492,
+      "LegalizeSundaAccess": 0.03900027275085449,
+      "LegalizeSundaMacro": 0.010483741760253906,
+      "LegalizeType": 0.0038602352142333984,
+      "LocalLayoutOpt": 0.01764845848083496,
+      "LoopFusion": 0.006066322326660156,
+      "LoopSplitting": 0.0015685558319091797,
+      "LowerBroadcast": 0.0020384788513183594,
+      "LowerCCOpBlockAxis": 0.005359172821044922,
+      "LowerComplexBroadcast": 0.0019440650939941406,
+      "LowerIntrinsics": 0.030491113662719727,
+      "LowerTensorOp": 0.012917041778564453,
+      "LowerTranspose": 0.010635852813720703,
+      "MacroGeneration": 0.06435012817382813,
+      "MaskPropagation": 0.0051097869873046875,
+      "MemcpyElimination": 0.11022067070007324,
+      "MutateDataType": 0.0014224052429199219,
+      "NeuronAliasDependencyInduction": 0.00023031234741210938,
+      "NeuronAliasDependencyReset": 0.021604061126708984,
+      "NeuronInstComb": 0.013072729110717773,
+      "NeuronLICM": 0.01006174087524414,
+      "NeuronLoopFusion": 0.017573833465576172,
+      "NeuronLoopInterchange": 0.0020608901977539063,
+      "NeuronSimplifier": 0.010074615478515625,
+      "NeuronSimplifyPredicates": 0.0060672760009765625,
+      "NeuronValueNumbering": 0.0041046142578125,
+      "OptimizeAliasedCopyChain": 0.0014190673828125,
+      "OptimizeNKIKernels": 0.0021109580993652344,
+      "PAGLayoutOpt": 0.3779466152191162,
+      "PComputeCutting": 0.008729696273803711,
+      "PGLayoutTilingPipeline": 1.5334703922271729,
+      "PGTiling": 0.47260475158691406,
+      "PadElimination": 0.0015625953674316406,
+      "ParAxesAnnotation": 0.2937772274017334,
+      "PartialLoopFusion": 0.016366004943847656,
+      "PartialSimdFusion": 0.01980447769165039,
+      "PerfectLoopNest": 0.0021877288818359375,
+      "RecognizeOpIdiom": 0.004831075668334961,
+      "Recompute": 0.00025010108947753906,
+      "RelaxPredicates": 0.0039484500885009766,
+      "Rematerialization": 0.004274129867553711,
+      "ReshapeWeights": 0.000804901123046875,
+      "ResolveAccessConflict": 0.0038733482360839844,
+      "ResolveComplicatePredicates": 0.0016858577728271484,
+      "RewriteReplicationMatmul": 0.0014014244079589844,
+      "RewriteWeights": 0.00405120849609375,
+      "SFKVectorizer": 0.20196890830993652,
+      "SimpleAllReduceTiling": 0.002203702926635742,
+      "Simplifier": 0.004297018051147461,
+      "SimplifyMacroPredicates": 0.01361393928527832,
+      "SimplifyNeuronTensor": 0.009984970092773438,
+      "SimplifySlice": 0.0010356903076171875,
+      "SimplifyTensor": 0.006205558776855469,
+      "SpillPSum": 0.016466140747070313,
+      "SplitAPUnionSets": 0.029446840286254883,
+      "SplitAccGrp": 0.0020453929901123047,
+      "StaticProfiler": 0.004591464996337891,
+      "StaticTransposeLocalTensor": 0.005173683166503906,
+      "SundaISel": 0.04554462432861328,
+      "TCTransform": 0.002426624298095703,
+      "TensorInitialization": 0.009510517120361328,
+      "TensorOpSimplifier": 0.0067560672760009766,
+      "TensorOpTransform": 0.028885841369628906,
+      "TileCCOps": 0.005466938018798828,
+      "TilingProfiler": 0.013426065444946289,
+      "TransformConvOp": 0.002458810806274414,
+      "TritiumFusion": 0.0620732307434082,
+      "ValueNumbering": 0.002520322799682617,
+      "VectorizeDMA": 0.005783796310424805,
+      "VectorizeMatMult": 0.005175352096557617,
+      "WeightCoalescing": 0.0029850006103515625,
+      "ZeroSizeTensorElimination": 0.00011801719665527344
+    },
+    "tensorizer": {
+      "DMATilingProfiler::TotalInstructionsAfterTiling": 2597.0,
+      "StaticProfiler::AifUb": 40.028141021728516,
+      "StaticProfiler::ArithmeticIntensityTensorizer": 420.0349426269531,
+      "StaticProfiler::AverageDmaLength": 1921.007568359375,
+      "StaticProfiler::AverageFractalPeUtilization": 99.95317840576172,
+      "StaticProfiler::AveragePartitionUtilization": 99.87249755859375,
+      "StaticProfiler::AveragePeUtilization": 99.80845642089844,
+      "StaticProfiler::DDRTransferBytes": 64558336.0,
+      "StaticProfiler::InternalTransferBytes": 52297728.0,
+      "StaticProfiler::LoadExpanded": 23298.0,
+      "StaticProfiler::LocalizationEfficiency": 1049.3489990234375,
+      "StaticProfiler::LocalizationEfficiencyIgnoreNonlocal": 1358.191162109375,
+      "StaticProfiler::StoreExpanded": 5505.0,
+      "StaticProfiler::TotalDMAExpanded": 28803.0,
+      "StaticProfiler::TotalDynamicInstancesCount": 3692.0,
+      "StaticProfiler::TotalDynamicInstancesWithMmPackedCount": 3689.0,
+      "StaticProfiler::TotalLNCComm": 0.0,
+      "StaticProfiler::TotalLNCCommTransfer": 0.0,
+      "TilingProfiler::AveragePartitionUtilizationAfterTiling": 0.0,
+      "TilingProfiler::AveragePeUtilizationAfterTiling": 0.0,
+      "TilingProfiler::BatchnormInstructionsAfterTiling": 0.0,
+      "TilingProfiler::DmaInstructionsAfterTiling": 0.0,
+      "TilingProfiler::GenericInstructionsAfterTiling": 48.0,
+      "TilingProfiler::MatMultInstructionsAfterTiling": 1412.0,
+      "TilingProfiler::NumPfTransposes": 7.0,
+      "TilingProfiler::NumPfTransposesForIo": 1.0,
+      "TilingProfiler::NumPfTransposesForLocal": 5.0,
+      "TilingProfiler::NumPfTransposesForNonlocal": 1.0,
+      "TilingProfiler::PfTransposeInstructions": 608.0,
+      "TilingProfiler::PfTransposeInstructionsForIo": 128.0,
+      "TilingProfiler::PfTransposeInstructionsForLocal": 416.0,
+      "TilingProfiler::PfTransposeInstructionsForNonlocal": 64.0,
+      "TilingProfiler::ReduceInstructionsAfterTiling": 0.0,
+      "TilingProfiler::SimdInstructionsAfterTiling": 257.0,
+      "TilingProfiler::TotalInstructionsAfterTiling": 0.0,
+      "TransformConvOp::Conv1d_depthwise_bf01_oi01_bf01": 0.0,
+      "TransformConvOp::Conv2d_dw_fb01_io01_01bf_rep_nhwc_Pcinh": 0.0,
+      "TransformConvOp::Conv2d_pbp_0f1b_0i1o_01fb_experimental_1": 0.0,
+      "TransformConvOp::Conv2d_pbp_fb01_io01_01bf_experimental_1": 0.0,
+      "TransformConvOp::conv2d_column_packing": 0.0,
+      "TransformConvOp::conv2d_column_packing_1": 0.0,
+      "TransformConvOp::conv2d_column_packing_io10": 0.0,
+      "TransformConvOp::conv2d_depthwise_f01b_o01i_bf01": 0.0
+    }
+  },
+  "sg0001": {
+    "compiletime": {
+      "AGOrderingAnalysisPass": 0.03313565254211426,
+      "AffinePredicateResolution": 0.0015239715576171875,
+      "AliasDependencyElimination": 0.00011467933654785156,
+      "AliasDependencyInduction": 0.009088993072509766,
+      "AliasDependencyReset": 1.062025547027588,
+      "BFComputeCutting": 0.0024559497833251953,
+      "BirCodeGenLoop": 0.03748297691345215,
+      "CCOpFusion": 0.04092240333557129,
+      "CanonicalizeDAGForPGTiling": 0.004329681396484375,
+      "CanonicalizeIR": 0.002464771270751953,
+      "CoalesceCCOp": 0.004778146743774414,
+      "CommuteConcat": 0.0011680126190185547,
+      "DMALocalityOpt": 0.0016834735870361328,
+      "DMAProfiler": 0.0039997100830078125,
+      "DMATilingProfiler": 0.004555702209472656,
+      "DataLocalityOpt": 0.13762187957763672,
+      "DataStreaming": 0.0044286251068115234,
+      "DeConcat": 0.0015981197357177734,
+      "DeadCodeElimination": 0.0020780563354492188,
+      "DeadStoreElimination": 0.03435230255126953,
+      "DelinearIndices": 0.00969839096069336,
+      "Delinearization": 0.0038826465606689453,
+      "DoNothing": 9.846687316894531e-05,
+      "DramToDramTranspose": 0.03438973426818848,
+      "DumpGraphAndMetadata": 0.00426793098449707,
+      "EliminateDivs": 0.004217386245727539,
+      "ExpandBatchNorm": 0.0019202232360839844,
+      "ExpandISAMacro": 0.0024042129516601563,
+      "FactorizeBlkDims": 0.01425933837890625,
+      "FactorizeThreadAxesInFreeDims": 0.0026972293853759766,
+      "FlattenMacroLoop": 0.002768993377685547,
+      "GenericAccessSimplifier": 0.001058816909790039,
+      "InferInitValue": 0.03559255599975586,
+      "InferIntrinsicOnCC": 0.009636163711547852,
+      "InferNeuronTensor": 0.04922318458557129,
+      "InferNonlocalTensors": 0.030732393264770508,
+      "InferPSumTensor": 0.03249359130859375,
+      "InlineNativeKernels": 0.0014734268188476563,
+      "InsertIOTransposes": 0.021765470504760742,
+      "InsertLocalTransposes": 0.006593465805053711,
+      "InsertOffloadedTransposes": 0.0034906864166259766,
+      "LICM": 0.003262758255004883,
+      "LateLegalizeInst": 0.00400543212890625,
+      "LateLegalizePostSplit": 0.00289154052734375,
+      "LateLowerReshapeOp": 0.002287149429321289,
+      "LateLowerTensorOp": 0.0046651363372802734,
+      "LateNeuronInstComb": 0.019269704818725586,
+      "LayoutPreprocessing": 0.03711414337158203,
+      "LayoutPreprocessingAndAnalysis": 0.2516040802001953,
+      "LayoutRequirementAnalysis": 0.007753133773803711,
+      "LegalizeCCOpLayout": 0.003732919692993164,
+      "LegalizeOpLevelAlias": 0.0016019344329833984,
+      "LegalizePartitionReduce": 0.0020945072174072266,
+      "LegalizeSundaAccess": 0.016069650650024414,
+      "LegalizeSundaMacro": 0.010806083679199219,
+      "LegalizeType": 0.004706859588623047,
+      "LocalLayoutOpt": 0.02442765235900879,
+      "LoopFusion": 0.0067822933197021484,
+      "LoopSplitting": 0.00033974647521972656,
+      "LowerBroadcast": 0.0019419193267822266,
+      "LowerCCOpBlockAxis": 0.005570650100708008,
+      "LowerComplexBroadcast": 0.0020999908447265625,
+      "LowerIntrinsics": 0.03607368469238281,
+      "LowerTensorOp": 0.011876583099365234,
+      "LowerTranspose": 0.011530637741088867,
+      "MacroGeneration": 0.10653066635131836,
+      "MaskPropagation": 0.003092050552368164,
+      "MemcpyElimination": 0.10495471954345703,
+      "MutateDataType": 0.0014193058013916016,
+      "NeuronAliasDependencyInduction": 0.0002295970916748047,
+      "NeuronAliasDependencyReset": 0.021070480346679688,
+      "NeuronInstComb": 0.012903451919555664,
+      "NeuronLICM": 0.00844264030456543,
+      "NeuronLoopFusion": 0.020880460739135742,
+      "NeuronLoopInterchange": 0.0021686553955078125,
+      "NeuronSimplifier": 0.011090755462646484,
+      "NeuronSimplifyPredicates": 0.0016274452209472656,
+      "NeuronValueNumbering": 0.004062652587890625,
+      "OptimizeAliasedCopyChain": 0.0014641284942626953,
+      "OptimizeNKIKernels": 0.0023856163024902344,
+      "PAGLayoutOpt": 0.17638587951660156,
+      "PComputeCutting": 0.00709986686706543,
+      "PGLayoutTilingPipeline": 1.142796516418457,
+      "PGTiling": 0.39766955375671387,
+      "PadElimination": 0.0015380382537841797,
+      "ParAxesAnnotation": 0.09186458587646484,
+      "PartialLoopFusion": 0.015995025634765625,
+      "PartialSimdFusion": 0.026766300201416016,
+      "PerfectLoopNest": 0.002192258834838867,
+      "RecognizeOpIdiom": 0.004943370819091797,
+      "Recompute": 0.00025773048400878906,
+      "RelaxPredicates": 0.003591299057006836,
+      "Rematerialization": 0.0025196075439453125,
+      "ReshapeWeights": 0.0007069110870361328,
+      "ResolveAccessConflict": 0.00481104850769043,
+      "ResolveComplicatePredicates": 0.002285003662109375,
+      "RewriteReplicationMatmul": 0.0021715164184570313,
+      "RewriteWeights": 0.003401041030883789,
+      "SFKVectorizer": 0.14661574363708496,
+      "SimpleAllReduceTiling": 0.0016207695007324219,
+      "Simplifier": 0.00443577766418457,
+      "SimplifyMacroPredicates": 0.006165742874145508,
+      "SimplifyNeuronTensor": 0.006829500198364258,
+      "SimplifySlice": 0.0013000965118408203,
+      "SimplifyTensor": 0.0061337947845458984,
+      "SpillPSum": 0.018761634826660156,
+      "SplitAPUnionSets": 0.017923593521118164,
+      "SplitAccGrp": 0.002531290054321289,
+      "StaticProfiler": 0.003990888595581055,
+      "StaticTransposeLocalTensor": 0.004915952682495117,
+      "SundaISel": 0.04209589958190918,
+      "TCTransform": 0.0012347698211669922,
+      "TensorInitialization": 0.002599954605102539,
+      "TensorOpSimplifier": 0.006845712661743164,
+      "TensorOpTransform": 0.03345227241516113,
+      "TileCCOps": 0.005617856979370117,
+      "TilingProfiler": 0.015013933181762695,
+      "TransformConvOp": 0.002393960952758789,
+      "TritiumFusion": 0.09340715408325195,
+      "ValueNumbering": 0.0031540393829345703,
+      "VectorizeDMA": 0.0015842914581298828,
+      "VectorizeMatMult": 0.0071103572845458984,
+      "WeightCoalescing": 0.0026235580444335938,
+      "ZeroSizeTensorElimination": 0.0001163482666015625
+    },
+    "tensorizer": {
+      "DMATilingProfiler::TotalInstructionsAfterTiling": 7847.0,
+      "StaticProfiler::AifUb": 490.6532287597656,
+      "StaticProfiler::ArithmeticIntensityTensorizer": 487.63507080078125,
+      "StaticProfiler::AverageDmaLength": 869.1515502929688,
+      "StaticProfiler::AverageFractalPeUtilization": 100.0,
+      "StaticProfiler::AveragePartitionUtilization": 99.83790588378906,
+      "StaticProfiler::AveragePeUtilization": 100.0,
+      "StaticProfiler::DDRTransferBytes": 215827456.0,
+      "StaticProfiler::InternalTransferBytes": 43515904.0,
+      "StaticProfiler::LoadExpanded": 238976.0,
+      "StaticProfiler::LocalizationEfficiency": 99.38487243652344,
+      "StaticProfiler::LocalizationEfficiencyIgnoreNonlocal": 107.76165771484375,
+      "StaticProfiler::StoreExpanded": 5121.0,
+      "StaticProfiler::TotalDMAExpanded": 244097.0,
+      "StaticProfiler::TotalDynamicInstancesCount": 9872.0,
+      "StaticProfiler::TotalDynamicInstancesWithMmPackedCount": 9872.0,
+      "StaticProfiler::TotalLNCComm": 0.0,
+      "StaticProfiler::TotalLNCCommTransfer": 0.0,
+      "TilingProfiler::AveragePartitionUtilizationAfterTiling": 0.0,
+      "TilingProfiler::AveragePeUtilizationAfterTiling": 0.0,
+      "TilingProfiler::BatchnormInstructionsAfterTiling": 0.0,
+      "TilingProfiler::DmaInstructionsAfterTiling": 0.0,
+      "TilingProfiler::GenericInstructionsAfterTiling": 32.0,
+      "TilingProfiler::MatMultInstructionsAfterTiling": 6016.0,
+      "TilingProfiler::NumPfTransposes": 8.0,
+      "TilingProfiler::NumPfTransposesForIo": 3.0,
+      "TilingProfiler::NumPfTransposesForLocal": 3.0,
+      "TilingProfiler::NumPfTransposesForNonlocal": 2.0,
+      "TilingProfiler::PfTransposeInstructions": 680.0,
+      "TilingProfiler::PfTransposeInstructionsForIo": 136.0,
+      "TilingProfiler::PfTransposeInstructionsForLocal": 288.0,
+      "TilingProfiler::PfTransposeInstructionsForNonlocal": 256.0,
+      "TilingProfiler::ReduceInstructionsAfterTiling": 0.0,
+      "TilingProfiler::SimdInstructionsAfterTiling": 288.0,
+      "TilingProfiler::TotalInstructionsAfterTiling": 0.0,
+      "TransformConvOp::Conv1d_depthwise_bf01_oi01_bf01": 0.0,
+      "TransformConvOp::Conv2d_dw_fb01_io01_01bf_rep_nhwc_Pcinh": 0.0,
+      "TransformConvOp::Conv2d_pbp_0f1b_0i1o_01fb_experimental_1": 0.0,
+      "TransformConvOp::Conv2d_pbp_fb01_io01_01bf_experimental_1": 0.0,
+      "TransformConvOp::conv2d_column_packing": 0.0,
+      "TransformConvOp::conv2d_column_packing_1": 0.0,
+      "TransformConvOp::conv2d_column_packing_io10": 0.0,
+      "TransformConvOp::conv2d_depthwise_f01b_o01i_bf01": 0.0
+    }
+  },
+  "sg0002": {
+    "compiletime": {
+      "AGOrderingAnalysisPass": 0.018257856369018555,
+      "AffinePredicateResolution": 0.0011677742004394531,
+      "AliasDependencyElimination": 0.0001201629638671875,
+      "AliasDependencyInduction": 0.0052988529205322266,
+      "AliasDependencyReset": 0.029210567474365234,
+      "BFComputeCutting": 0.0032625198364257813,
+      "BirCodeGenLoop": 0.4527714252471924,
+      "CCOpFusion": 0.02410125732421875,
+      "CanonicalizeDAGForPGTiling": 0.004324913024902344,
+      "CanonicalizeIR": 0.0019502639770507813,
+      "CoalesceCCOp": 0.014463186264038086,
+      "CommuteConcat": 0.0008339881896972656,
+      "DMALocalityOpt": 0.005598783493041992,
+      "DMAProfiler": 0.01209115982055664,
+      "DMATilingProfiler": 0.004332065582275391,
+      "DataLocalityOpt": 0.07260942459106445,
+      "DataStreaming": 0.03940248489379883,
+      "DeConcat": 0.0005326271057128906,
+      "DeadCodeElimination": 0.0009255409240722656,
+      "DeadStoreElimination": 0.0055675506591796875,
+      "DelinearIndices": 0.004735231399536133,
+      "Delinearization": 0.0030374526977539063,
+      "DoNothing": 7.033348083496094e-05,
+      "DramToDramTranspose": 0.018135547637939453,
+      "DumpGraphAndMetadata": 0.09476375579833984,
+      "EliminateDivs": 0.002595663070678711,
+      "ExpandBatchNorm": 0.002063274383544922,
+      "ExpandISAMacro": 0.011472225189208984,
+      "FactorizeBlkDims": 0.008858203887939453,
+      "FactorizeThreadAxesInFreeDims": 0.0010046958923339844,
+      "FlattenMacroLoop": 0.002232074737548828,
+      "GenericAccessSimplifier": 0.0018167495727539063,
+      "InferInitValue": 0.024865150451660156,
+      "InferIntrinsicOnCC": 0.009101152420043945,
+      "InferNeuronTensor": 0.023293495178222656,
+      "InferNonlocalTensors": 0.01632833480834961,
+      "InferPSumTensor": 0.27681708335876465,
+      "InlineNativeKernels": 0.0081634521484375,
+      "InsertIOTransposes": 0.019203901290893555,
+      "InsertLocalTransposes": 0.0042340755462646484,
+      "InsertOffloadedTransposes": 0.002811431884765625,
+      "LICM": 0.0029730796813964844,
+      "LateLegalizeInst": 0.01390385627746582,
+      "LateLegalizePostSplit": 0.012536048889160156,
+      "LateLowerReshapeOp": 0.0018641948699951172,
+      "LateLowerTensorOp": 0.0014081001281738281,
+      "LateNeuronInstComb": 0.008648872375488281,
+      "LayoutPreprocessing": 0.02658390998840332,
+      "LayoutPreprocessingAndAnalysis": 0.10707235336303711,
+      "LayoutRequirementAnalysis": 0.005135536193847656,
+      "LegalizeCCOpLayout": 0.002307415008544922,
+      "LegalizeOpLevelAlias": 0.0012297630310058594,
+      "LegalizePartitionReduce": 0.0010194778442382813,
+      "LegalizeSundaAccess": 0.07593941688537598,
+      "LegalizeSundaMacro": 0.010968446731567383,
+      "LegalizeType": 0.011834383010864258,
+      "LocalLayoutOpt": 0.013799905776977539,
+      "LoopFusion": 0.0052182674407958984,
+      "LoopSplitting": 0.0003161430358886719,
+      "LowerBroadcast": 0.0013611316680908203,
+      "LowerCCOpBlockAxis": 0.0040547847747802734,
+      "LowerComplexBroadcast": 0.002165079116821289,
+      "LowerIntrinsics": 0.31132984161376953,
+      "LowerTensorOp": 0.010558843612670898,
+      "LowerTranspose": 0.012272357940673828,
+      "MacroGeneration": 0.029862642288208008,
+      "MaskPropagation": 0.002757549285888672,
+      "MemcpyElimination": 0.025969266891479492,
+      "MutateDataType": 0.002087831497192383,
+      "NeuronAliasDependencyInduction": 0.00016880035400390625,
+      "NeuronAliasDependencyReset": 0.020352602005004883,
+      "NeuronInstComb": 0.004126310348510742,
+      "NeuronLICM": 0.0351865291595459,
+      "NeuronLoopFusion": 0.007991313934326172,
+      "NeuronLoopInterchange": 0.002409219741821289,
+      "NeuronSimplifier": 0.007069587707519531,
+      "NeuronSimplifyPredicates": 0.12138772010803223,
+      "NeuronValueNumbering": 0.0028395652770996094,
+      "OptimizeAliasedCopyChain": 0.0005936622619628906,
+      "OptimizeNKIKernels": 0.5374257564544678,
+      "PAGLayoutOpt": 0.08115577697753906,
+      "PComputeCutting": 0.004801273345947266,
+      "PGLayoutTilingPipeline": 0.5454635620117188,
+      "PGTiling": 0.14933419227600098,
+      "PadElimination": 0.00034046173095703125,
+      "ParAxesAnnotation": 0.053552865982055664,
+      "PartialLoopFusion": 0.0067539215087890625,
+      "PartialSimdFusion": 0.00693058967590332,
+      "PerfectLoopNest": 0.0035321712493896484,
+      "RecognizeOpIdiom": 0.003947257995605469,
+      "Recompute": 0.00024962425231933594,
+      "RelaxPredicates": 0.013285398483276367,
+      "Rematerialization": 0.002062082290649414,
+      "ReshapeWeights": 0.002131223678588867,
+      "ResolveAccessConflict": 0.0038597583770751953,
+      "ResolveComplicatePredicates": 0.002032756805419922,
+      "RewriteReplicationMatmul": 0.001924753189086914,
+      "RewriteWeights": 0.002452373504638672,
+      "SFKVectorizer": 0.2690722942352295,
+      "SimpleAllReduceTiling": 0.008755922317504883,
+      "Simplifier": 0.004038810729980469,
+      "SimplifyMacroPredicates": 0.010622739791870117,
+      "SimplifyNeuronTensor": 1.059011697769165,
+      "SimplifySlice": 0.0009577274322509766,
+      "SimplifyTensor": 0.005341768264770508,
+      "SpillPSum": 0.011537313461303711,
+      "SplitAPUnionSets": 0.10771751403808594,
+      "SplitAccGrp": 0.002201557159423828,
+      "StaticProfiler": 0.012447118759155273,
+      "StaticTransposeLocalTensor": 0.0038712024688720703,
+      "SundaISel": 0.04214668273925781,
+      "TCTransform": 0.0008432865142822266,
+      "TensorInitialization": 0.012825727462768555,
+      "TensorOpSimplifier": 0.004651308059692383,
+      "TensorOpTransform": 0.019537687301635742,
+      "TileCCOps": 0.006766319274902344,
+      "TilingProfiler": 0.006911277770996094,
+      "TransformConvOp": 0.0030303001403808594,
+      "TritiumFusion": 0.04502224922180176,
+      "ValueNumbering": 0.001996755599975586,
+      "VectorizeDMA": 0.0019402503967285156,
+      "VectorizeMatMult": 0.0027413368225097656,
+      "WeightCoalescing": 0.00829005241394043,
+      "ZeroSizeTensorElimination": 0.00013709068298339844
+    },
+    "tensorizer": {
+      "DMATilingProfiler::TotalInstructionsAfterTiling": 44382.0,
+      "StaticProfiler::AifUb": 205.154296875,
+      "StaticProfiler::ArithmeticIntensityTensorizer": 201.6046905517578,
+      "StaticProfiler::AverageDmaLength": 1901.806396484375,
+      "StaticProfiler::AverageFractalPeUtilization": 99.66542053222656,
+      "StaticProfiler::AveragePartitionUtilization": 97.7269515991211,
+      "StaticProfiler::AveragePeUtilization": 98.64861297607422,
+      "StaticProfiler::DDRTransferBytes": 795531072.0,
+      "StaticProfiler::InternalTransferBytes": 646388224.0,
+      "StaticProfiler::LoadExpanded": 376342.0,
+      "StaticProfiler::LocalizationEfficiency": 98.26979064941406,
+      "StaticProfiler::LocalizationEfficiencyIgnoreNonlocal": 101.01405334472656,
+      "StaticProfiler::StoreExpanded": 4189.0,
+      "StaticProfiler::TotalDMAExpanded": 380531.0,
+      "StaticProfiler::TotalDynamicInstancesCount": 53882.0,
+      "StaticProfiler::TotalDynamicInstancesWithMmPackedCount": 53436.0,
+      "StaticProfiler::TotalLNCComm": 0.0,
+      "StaticProfiler::TotalLNCCommTransfer": 0.0,
+      "TilingProfiler::AveragePartitionUtilizationAfterTiling": 0.0,
+      "TilingProfiler::AveragePeUtilizationAfterTiling": 0.0,
+      "TilingProfiler::BatchnormInstructionsAfterTiling": 0.0,
+      "TilingProfiler::DmaInstructionsAfterTiling": 0.0,
+      "TilingProfiler::GenericInstructionsAfterTiling": 4.0,
+      "TilingProfiler::MatMultInstructionsAfterTiling": 23616.0,
+      "TilingProfiler::NumPfTransposes": 5.0,
+      "TilingProfiler::NumPfTransposesForIo": 1.0,
+      "TilingProfiler::NumPfTransposesForLocal": 1.0,
+      "TilingProfiler::NumPfTransposesForNonlocal": 3.0,
+      "TilingProfiler::PfTransposeInstructions": 19393.0,
+      "TilingProfiler::PfTransposeInstructionsForIo": 19008.0,
+      "TilingProfiler::PfTransposeInstructionsForLocal": 1.0,
+      "TilingProfiler::PfTransposeInstructionsForNonlocal": 384.0,
+      "TilingProfiler::ReduceInstructionsAfterTiling": 4.0,
+      "TilingProfiler::SimdInstructionsAfterTiling": 158.0,
+      "TilingProfiler::TotalInstructionsAfterTiling": 0.0,
+      "TransformConvOp::Conv1d_depthwise_bf01_oi01_bf01": 0.0,
+      "TransformConvOp::Conv2d_dw_fb01_io01_01bf_rep_nhwc_Pcinh": 0.0,
+      "TransformConvOp::Conv2d_pbp_0f1b_0i1o_01fb_experimental_1": 0.0,
+      "TransformConvOp::Conv2d_pbp_fb01_io01_01bf_experimental_1": 0.0,
+      "TransformConvOp::conv2d_column_packing": 0.0,
+      "TransformConvOp::conv2d_column_packing_1": 0.0,
+      "TransformConvOp::conv2d_column_packing_io10": 0.0,
+      "TransformConvOp::conv2d_depthwise_f01b_o01i_bf01": 0.0
+    }
+  },
+  "sg01": {
+    "compiletime": {
+      "CanonicalizeConv": 1.2000000424450263e-05,
+      "CanonicalizeForTensorizer": 1.2999999853491317e-05,
+      "Canonicalizer": 0.0002500000118743628,
+      "HoistCompute": 3.000000106112566e-06,
+      "IdentifyCrossPassTensors": 2.300000051036477e-05,
+      "MemcastMotion": 1.1000000085914508e-05,
+      "PenguinizeFunctions": 1.4000000192027073e-05,
+      "PruneFunctions": 3.099999958067201e-05,
+      "RemoveOptimizationBarriers": 2.2000000171829015e-05,
+      "ScatterMotion": 2.9999999242136255e-05,
+      "TensorizerLegalizationPass": 1.700000029813964e-05,
+      "VerifySupportedOps": 9.000000318337698e-06,
+      "algsimp": 6.299999949987978e-05,
+      "batchnorm_expander": 1.2999999853491317e-05,
+      "boundary-marker-removal": 4.999999873689376e-06,
+      "call-inliner": 9.000000318337698e-06,
+      "canonicalize-boundary-marker": 6.000000212225132e-06,
+      "collective-stream-id-checker": 3.000000106112566e-06,
+      "comparison-expander": 4.999999873689376e-06,
+      "computation-deduplicator": 1.8000000636675395e-05,
+      "conditional-to-select": 4.999999873689376e-06,
+      "config-lowering": 2.5999999706982635e-05,
+      "constant_folding": 7.999999979801942e-06,
+      "cse": 1.2000000424450263e-05,
+      "dce": 9.999999974752427e-07,
+      "dynamic-slice-transpose": 3.999999989900971e-06,
+      "eliminate-redundant-compare": 3.999999989900971e-06,
+      "emit-offloaded-dropout": 1.2999999853491317e-05,
+      "flatten-call-graph": 7.999999979801942e-06,
+      "fuse-send-recv": 2.099999983329326e-05,
+      "hilo::LegalizeAlias": 4.999999873689376e-06,
+      "hilo::NeuronInstCombine": 4.5000000682193786e-05,
+      "hilo::NeuronOpFusion": 1.700000029813964e-05,
+      "hilo::ReplaceTokenTypeWithU8Pass": 2.099999983329326e-05,
+      "hilo::ScheduleFusion": 9.999999974752427e-07,
+      "hilo::SixtyFourHack": 1.2999999853491317e-05,
+      "hilo::VerifyAliasing": 1.9999999949504854e-06,
+      "hlo-mac-count": 2.9999999242136255e-05,
+      "hlo-verifier": 0.00018000000272877514,
+      "legalize-ccops": 9.999999974752427e-07,
+      "legalize-compare": 3.999999989900971e-06,
+      "lower-argminmax-custom-call": 3.999999989900971e-06,
+      "map-inline": 1.2000000424450263e-05,
+      "metadata-naming": 1.8000000636675395e-05,
+      "mlir::detail::OpToOpPassAdaptor": 9.999999747378752e-05,
+      "mlir::hlo::MhloToPyPenguin": 0.0009420000133104622,
+      "mlir::mhlo::LowerComplexExtraPass": 7.999999797903001e-05,
+      "mlir::mhlo::LowerComplexPass": 0.00015799999528098851,
+      "native-to-custom-softmax": 6.000000212225132e-06,
+      "native-to-custom-softmax-dx": 1.2999999853491317e-05,
+      "operand_upcaster": 1.4999999621068127e-05,
+      "post-par-pipe-begin": 1.9999999949504854e-06,
+      "post-par-pipe-end": 0.0,
+      "post-partition-simplification": 0.0005130000063218176,
+      "replace-minimum-constant": 6.000000212225132e-06,
+      "reshape-mover": 3.000000106112566e-06,
+      "simplify-concat": 4.8999998398358e-05,
+      "simplify-while-loops": 1.9999999949504854e-06,
+      "transform-variadic-reduce": 9.000000318337698e-06,
+      "tuple-simplifier": 4.999999873689376e-06,
+      "unpack-nested-aws-ntwsr": 3.999999989900971e-06,
+      "unroll-while-loop": 0.0
+    },
+    "hilo": {
+      "ArithmeticIntensity": 457.20416259765625,
+      "HloMacCount": 50465865728.0,
+      "Traffic": 220758560.0
+    }
+  },
+  "sg02": {
+    "compiletime": {
+      "CanonicalizeConv": 0.0002589999930933118,
+      "CanonicalizeForTensorizer": 1.2000000424450263e-05,
+      "Canonicalizer": 0.0003060000017285347,
+      "HoistCompute": 1.9999999949504854e-06,
+      "IdentifyCrossPassTensors": 2.4000000848900527e-05,
+      "MemcastMotion": 1.2999999853491317e-05,
+      "PenguinizeFunctions": 9.000000318337698e-06,
+      "PruneFunctions": 7.999999979801942e-06,
+      "RemoveOptimizationBarriers": 2.099999983329326e-05,
+      "ScatterMotion": 1.9999999949504854e-06,
+      "TensorizerLegalizationPass": 4.999999873689376e-06,
+      "VerifySupportedOps": 1.1000000085914508e-05,
+      "algsimp": 5.900000178371556e-05,
+      "batchnorm_expander": 1.2000000424450263e-05,
+      "boundary-marker-removal": 3.999999989900971e-06,
+      "call-inliner": 1.1000000085914508e-05,
+      "canonicalize-boundary-marker": 4.999999873689376e-06,
+      "collective-stream-id-checker": 3.000000106112566e-06,
+      "comparison-expander": 4.999999873689376e-06,
+      "computation-deduplicator": 4.099999932805076e-05,
+      "conditional-to-select": 6.000000212225132e-06,
+      "config-lowering": 2.5999999706982635e-05,
+      "constant_folding": 9.000000318337698e-06,
+      "cse": 1.1000000085914508e-05,
+      "dce": 9.999999974752427e-07,
+      "dynamic-slice-transpose": 3.999999989900971e-06,
+      "eliminate-redundant-compare": 3.999999989900971e-06,
+      "emit-offloaded-dropout": 1.2999999853491317e-05,
+      "flatten-call-graph": 1.1000000085914508e-05,
+      "fuse-send-recv": 1.8000000636675395e-05,
+      "hilo::LegalizeAlias": 1.9999999949504854e-06,
+      "hilo::NeuronInstCombine": 5.8000001445179805e-05,
+      "hilo::NeuronOpFusion": 1.9999999949504854e-06,
+      "hilo::ReplaceTokenTypeWithU8Pass": 9.000000318337698e-06,
+      "hilo::ScheduleFusion": 3.999999989900971e-06,
+      "hilo::SixtyFourHack": 3.899999865097925e-05,
+      "hilo::VerifyAliasing": 9.999999974752427e-07,
+      "hlo-mac-count": 0.00016799999866634607,
+      "hlo-verifier": 0.00015900000289548188,
+      "legalize-ccops": 9.999999974752427e-07,
+      "legalize-compare": 3.000000106112566e-06,
+      "lower-argminmax-custom-call": 3.000000106112566e-06,
+      "map-inline": 1.2000000424450263e-05,
+      "metadata-naming": 1.700000029813964e-05,
+      "mlir::detail::OpToOpPassAdaptor": 2.2000000171829015e-05,
+      "mlir::hlo::MhloToPyPenguin": 0.0008549999911338091,
+      "mlir::mhlo::LowerComplexExtraPass": 8.800000068731606e-05,
+      "mlir::mhlo::LowerComplexPass": 0.00015999999595806003,
+      "native-to-custom-softmax": 4.999999873689376e-06,
+      "native-to-custom-softmax-dx": 1.9999999494757503e-05,
+      "operand_upcaster": 1.2000000424450263e-05,
+      "post-par-pipe-begin": 1.9999999949504854e-06,
+      "post-par-pipe-end": 0.0,
+      "post-partition-simplification": 0.0005009999731555581,
+      "replace-minimum-constant": 9.000000318337698e-06,
+      "reshape-mover": 3.000000106112566e-06,
+      "simplify-concat": 4.3000000005122274e-05,
+      "simplify-while-loops": 1.9999999949504854e-06,
+      "transform-variadic-reduce": 4.8000001697801054e-05,
+      "tuple-simplifier": 4.999999873689376e-06,
+      "unpack-nested-aws-ntwsr": 3.999999989900971e-06,
+      "unroll-while-loop": 0.0
+    },
+    "hilo": {
+      "ArithmeticIntensity": 99.1578140258789,
+      "HloMacCount": 38965870592.0,
+      "Traffic": 785936448.0
+    }
+  }
+}

context_encoding_model/_tp0_bk2/graph.neff

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:d7e216fd8f0f2acfef59524e7cdb4ead506b2c17c584ce45dd222cd4dc4e3f4f
+size 1987584

context_encoding_model/_tp0_bk2/metaneff.pb

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:99c279a1a32451ce56757879c7a74b6ff23378ae19871f2aee2c2746ceda57f3
+size 1373735

context_encoding_model/_tp0_bk2/model.MODULE_00594b8bc68e927f3dbe+1ad60ced.hlo_module.pb

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:970c5138d61d773fc00bacb9090fbc05a05573925b8d91068006c211596d3f78
+size 1450821

context_encoding_model/_tp0_bk2/model.MODULE_00594b8bc68e927f3dbe+1ad60ced.neff

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:d7e216fd8f0f2acfef59524e7cdb4ead506b2c17c584ce45dd222cd4dc4e3f4f
+size 1987584

context_encoding_model/_tp0_bk2/neuron_config.json

@@ -0,0 +1,220 @@
+{
+  "_attn_implementation_autoset": false,
+  "_name_or_path": "Qwen/Qwen3-8B",
+  "add_cross_attention": false,
+  "architectures": [
+    "Qwen3ForCausalLM"
+  ],
+  "attention_bias": false,
+  "attention_dropout": 0.0,
+  "attribute_map": {},
+  "bad_words_ids": null,
+  "begin_suppress_tokens": null,
+  "bos_token_id": 151643,
+  "chunk_size_feed_forward": 0,
+  "cross_attention_hidden_size": null,
+  "decoder_start_token_id": null,
+  "diversity_penalty": 0.0,
+  "do_sample": false,
+  "early_stopping": false,
+  "encoder_no_repeat_ngram_size": 0,
+  "eos_token_id": 151645,
+  "exponential_decay_length_penalty": null,
+  "finetuning_task": null,
+  "forced_bos_token_id": null,
+  "forced_eos_token_id": null,
+  "fused_spec_config": null,
+  "head_dim": 128,
+  "hidden_act": "silu",
+  "hidden_size": 4096,
+  "id2label": {
+    "0": "LABEL_0",
+    "1": "LABEL_1"
+  },
+  "initializer_range": 0.02,
+  "intermediate_size": 12288,
+  "is_decoder": false,
+  "is_encoder_decoder": false,
+  "label2id": {
+    "LABEL_0": 0,
+    "LABEL_1": 1
+  },
+  "length_penalty": 1.0,
+  "max_length": 20,
+  "max_position_embeddings": 40960,
+  "max_window_layers": 36,
+  "metadata": null,
+  "min_length": 0,
+  "model_type": "qwen3",
+  "neuron_config": {
+    "activation_quantization_type": null,
+    "allow_input_truncation": false,
+    "apply_seq_ids_mask": false,
+    "async_mode": false,
+    "attention_dp_degree": 1,
+    "attention_dtype": null,
+    "attn_block_cte_nki_kernel_enabled": false,
+    "attn_block_tkg_nki_kernel_cache_update": false,
+    "attn_block_tkg_nki_kernel_enabled": false,
+    "attn_cls": {
+      "__module__": "neuronx_distributed_inference.models.qwen3.modeling_qwen3",
+      "__name__": "NeuronQwen3Attention"
+    },
+    "attn_kernel_enabled": null,
+    "attn_tkg_builtin_kernel_enabled": false,
+    "attn_tkg_nki_kernel_enabled": false,
+    "batch_size": 1,
+    "bucket_n_active_tokens": true,
+    "buckets": [
+      512
+    ],
+    "cast_type": "config",
+    "cc_pipeline_tiling_factor": 2,
+    "chunked_prefill_config": null,
+    "context_encoding_buckets": [
+      512
+    ],
+    "cp_degree": 1,
+    "ctx_batch_size": 1,
+    "disable_kv_cache_tiling": false,
+    "draft_model_modules_to_not_convert": null,
+    "enable_bucketing": true,
+    "enable_eagle_draft_input_norm": false,
+    "enable_eagle_speculation": false,
+    "enable_fused_speculation": false,
+    "enable_long_context_mode": false,
+    "enable_output_completion_notifications": false,
+    "enable_spill_reload_dge": false,
+    "enable_token_tree": false,
+    "ep_degree": 1,
+    "expert_mlp_nki_kernel_enabled": null,
+    "flash_decoding_enabled": false,
+    "fused_qkv": false,
+    "fused_rmsnorm_skip_gamma": false,
+    "is_block_kv_layout": null,
+    "is_chunked_prefill": false,
+    "is_continuous_batching": true,
+    "is_eagle_draft": false,
+    "is_medusa": false,
+    "is_prefill_stage": true,
+    "is_prefix_caching": false,
+    "k_cache_transposed": false,
+    "kv_cache_batch_size": 1,
+    "kv_cache_padding_size": 0,
+    "kv_cache_quant": false,
+    "kv_cache_tiling": false,
+    "layer_boundary_markers": false,
+    "lm_head_pad": false,
+    "lm_head_pad_alignment_size": 1,
+    "local_ranks_size": 2,
+    "logical_nc_config": 1,
+    "lora_config": null,
+    "max_batch_size": 1,
+    "max_context_length": 1024,
+    "max_length": 1024,
+    "max_new_tokens": null,
+    "medusa_speculation_length": 0,
+    "medusa_tree": null,
+    "mlp_kernel_enabled": false,
+    "mlp_kernel_fuse_residual_add": false,
+    "modules_to_not_convert": null,
+    "moe_fused_nki_kernel_enabled": null,
+    "n_active_tokens": 1024,
+    "n_positions": 1024,
+    "num_medusa_heads": 0,
+    "on_cpu": false,
+    "on_device_sampling_config": {
+      "deterministic": false,
+      "do_sample": false,
+      "dynamic": true,
+      "global_topk": 256,
+      "on_device_sampling_config": true,
+      "temperature": 1.0,
+      "top_k": 1,
+      "top_k_kernel_enabled": false,
+      "top_p": 1.0
+    },
+    "output_logits": false,
+    "overrides_torch_dtype": true,
+    "pa_block_size": 1024,
+    "pa_num_blocks": 1,
+    "padding_side": "right",
+    "pp_degree": 1,
+    "prefix_buckets": null,
+    "qk_layernorm": false,
+    "qkv_kernel_enabled": false,
+    "qkv_kernel_fuse_residual_add": false,
+    "qkv_kernel_nbsd_layout": false,
+    "quantization_dtype": "int8",
+    "quantization_type": "per_tensor_symmetric",
+    "quantize_clamp_bound": Infinity,
+    "quantized": false,
+    "quantized_checkpoints_path": null,
+    "quantized_mlp_kernel_enabled": false,
+    "rmsnorm_quantize_kernel_enabled": false,
+    "router_topk_nki_kernel_enabled": null,
+    "rpl_reduce_dtype": null,
+    "save_sharded_checkpoint": true,
+    "scratchpad_page_size": null,
+    "seq_len": 1024,
+    "seq_len_threshold_for_cc_tiling": 16384,
+    "sequence_parallel_enabled": false,
+    "shared_mlp_nki_kernel_enabled": null,
+    "skip_sharding": false,
+    "skip_warmup": false,
+    "spec_batch_size": 1,
+    "speculation_length": 0,
+    "start_rank_id": 0,
+    "target": null,
+    "tile_cc": false,
+    "tkg_batch_size": 1,
+    "token_generation_buckets": null,
+    "token_tree_config": null,
+    "torch_dtype": "bfloat16",
+    "tp_degree": 2,
+    "vocab_parallel": false,
+    "weight_gather_seq_len_threshold": 32768,
+    "weights_to_skip_layout_optimization": [],
+    "world_size": 2
+  },
+  "no_repeat_ngram_size": 0,
+  "num_attention_heads": 32,
+  "num_beam_groups": 1,
+  "num_beams": 1,
+  "num_cores_per_group": 1,
+  "num_hidden_layers": 36,
+  "num_key_value_heads": 8,
+  "num_return_sequences": 1,
+  "output_attentions": false,
+  "output_hidden_states": false,
+  "output_scores": false,
+  "pad_token_id": 0,
+  "prefix": null,
+  "problem_type": null,
+  "pruned_heads": {},
+  "remove_invalid_values": false,
+  "repetition_penalty": 1.0,
+  "return_dict": true,
+  "return_dict_in_generate": false,
+  "rms_norm_eps": 1e-06,
+  "rope_scaling": null,
+  "rope_theta": 1000000,
+  "sep_token_id": null,
+  "sliding_window": null,
+  "suppress_tokens": null,
+  "task_specific_params": null,
+  "temperature": 1.0,
+  "tf_legacy_loss": false,
+  "tie_encoder_decoder": false,
+  "tie_word_embeddings": false,
+  "tokenizer_class": null,
+  "top_k": 50,
+  "top_p": 1.0,
+  "torchscript": false,
+  "transformers_version": "4.51.0",
+  "typical_p": 1.0,
+  "use_bfloat16": false,
+  "use_cache": true,
+  "use_sliding_window": false,
+  "vocab_size": 151936
+}

context_encoding_model/_tp0_bk3/command.txt

@@ -0,0 +1 @@
+neuronx-cc compile --framework=XLA model.MODULE_b3ddbc97e5f0d1d64c82+155de413.hlo_module.pb --output model.MODULE_b3ddbc97e5f0d1d64c82+155de413.neff --target=trn1 --auto-cast=none --model-type=transformer '--tensorizer-options=--enable-ccop-compute-overlap --cc-pipeline-tiling-factor=2 --vectorize-strided-dma ' --lnc=1 -O1 '--internal-hlo2tensorizer-options= --modular-flow-mac-threshold=10  --verify-hlo=true' --logfile=log-neuron-cc.txt --verbose=35

context_encoding_model/_tp0_bk3/compile_flags.MODULE_b3ddbc97e5f0d1d64c82+155de413.json

@@ -0,0 +1 @@
+["--target=trn1", "--auto-cast=none", "--model-type=transformer", "--tensorizer-options=--enable-ccop-compute-overlap --cc-pipeline-tiling-factor=2 --vectorize-strided-dma ", "--lnc=1", "-O1", "--internal-hlo2tensorizer-options= --modular-flow-mac-threshold=10  --verify-hlo=true", "--logfile=/home/ubuntu/qwen3/context_encoding_model/_tp0_bk3/log-neuron-cc.txt"]

context_encoding_model/_tp0_bk3/global_metric_store.json

@@ -0,0 +1,1079 @@
+{
+  "Average": {
+    "tensorizer": {
+      "StaticProfiler::AverageFractalPeUtilization": 99.7004623413086,
+      "StaticProfiler::AveragePartitionUtilization": 97.94140625,
+      "StaticProfiler::AveragePeUtilization": 98.78884887695313,
+      "StaticProfiler::LocalizationEfficiency": 91.59693145751953,
+      "StaticProfiler::LocalizationEfficiencyIgnoreNonlocal": 95.863037109375,
+      "TilingProfiler::AveragePartitionUtilizationAfterTiling": 0.0,
+      "TilingProfiler::AveragePeUtilizationAfterTiling": 0.0
+    }
+  },
+  "Count": {
+    "tensorizer": {
+      "StaticProfiler::AverageFractalPeUtilization": 1.0,
+      "StaticProfiler::AveragePartitionUtilization": 1.0,
+      "StaticProfiler::AveragePeUtilization": 1.0,
+      "StaticProfiler::LocalizationEfficiency": 1.0,
+      "StaticProfiler::LocalizationEfficiencyIgnoreNonlocal": 1.0,
+      "TilingProfiler::AveragePartitionUtilizationAfterTiling": 1.0,
+      "TilingProfiler::AveragePeUtilizationAfterTiling": 1.0
+    }
+  },
+  "Sum": {
+    "compiletime": {
+      "AGOrderingAnalysisPass": 0.01837611198425293,
+      "AffinePredicateResolution": 0.0011184215545654297,
+      "AliasDependencyElimination": 0.00015664100646972656,
+      "AliasDependencyInduction": 0.005170583724975586,
+      "AliasDependencyReset": 0.027508020401000977,
+      "BFComputeCutting": 0.0036101341247558594,
+      "BirCodeGenLoop": 0.4774467945098877,
+      "CCOpFusion": 0.033265113830566406,
+      "CanonicalizeConv": 2.300000051036477e-05,
+      "CanonicalizeDAGForPGTiling": 0.004282712936401367,
+      "CanonicalizeForTensorizer": 4.600000102072954e-05,
+      "CanonicalizeIR": 0.0024569034576416016,
+      "Canonicalizer": 0.0009039999567903578,
+      "CoalesceCCOp": 0.014229059219360352,
+      "CommuteConcat": 0.0017316341400146484,
+      "DMALocalityOpt": 0.005630016326904297,
+      "DMAProfiler": 0.012981653213500977,
+      "DMATilingProfiler": 0.0037560462951660156,
+      "DataLocalityOpt": 0.07645320892333984,
+      "DataStreaming": 0.03730320930480957,
+      "DeConcat": 0.0018520355224609375,
+      "DeadCodeElimination": 0.0020148754119873047,
+      "DeadStoreElimination": 0.006912708282470703,
+      "DelinearIndices": 0.004647254943847656,
+      "Delinearization": 0.003908872604370117,
+      "DoNothing": 0.0001888275146484375,
+      "DramToDramTranspose": 0.02015542984008789,
+      "DumpGraphAndMetadata": 0.08691883087158203,
+      "EliminateDivs": 0.0025060176849365234,
+      "ExpandBatchNorm": 0.0027189254760742188,
+      "ExpandISAMacro": 0.011646032333374023,
+      "FactorizeBlkDims": 0.010123252868652344,
+      "FactorizeThreadAxesInFreeDims": 0.0023202896118164063,
+      "FlattenMacroLoop": 0.00232696533203125,
+      "GenericAccessSimplifier": 0.0008094310760498047,
+      "HoistCompute": 5.999999757477781e-06,
+      "IdentifyCrossPassTensors": 5.2999999752501026e-05,
+      "InferInitValue": 0.02833867073059082,
+      "InferIntrinsicOnCC": 0.008923768997192383,
+      "InferNeuronTensor": 0.025766372680664063,
+      "InferNonlocalTensors": 0.014599800109863281,
+      "InferPSumTensor": 0.28418898582458496,
+      "InlineNativeKernels": 0.00860905647277832,
+      "InsertIOTransposes": 0.01989889144897461,
+      "InsertLocalTransposes": 0.004229307174682617,
+      "InsertOffloadedTransposes": 0.0029871463775634766,
+      "LICM": 0.0030870437622070313,
+      "LateLegalizeInst": 0.014106035232543945,
+      "LateLegalizePostSplit": 0.014872312545776367,
+      "LateLowerReshapeOp": 0.0010464191436767578,
+      "LateLowerTensorOp": 0.002707242965698242,
+      "LateNeuronInstComb": 0.010563373565673828,
+      "LayoutPreprocessing": 0.026853561401367188,
+      "LayoutPreprocessingAndAnalysis": 0.0556035041809082,
+      "LayoutRequirementAnalysis": 0.004946470260620117,
+      "LegalizeCCOpLayout": 0.0025353431701660156,
+      "LegalizeOpLevelAlias": 0.0018966197967529297,
+      "LegalizePartitionReduce": 0.0017490386962890625,
+      "LegalizeSundaAccess": 0.07800722122192383,
+      "LegalizeSundaMacro": 0.012125253677368164,
+      "LegalizeType": 0.012685060501098633,
+      "LocalLayoutOpt": 0.013860225677490234,
+      "LoopFusion": 0.005201578140258789,
+      "LoopSplitting": 0.0003204345703125,
+      "LowerBroadcast": 0.002086162567138672,
+      "LowerCCOpBlockAxis": 0.0040171146392822266,
+      "LowerComplexBroadcast": 0.002280712127685547,
+      "LowerIntrinsics": 0.3143951892852783,
+      "LowerTensorOp": 0.01141357421875,
+      "LowerTranspose": 0.012923002243041992,
+      "MacroGeneration": 0.034410953521728516,
+      "MaskPropagation": 0.0028192996978759766,
+      "MemcastMotion": 1.8000000636675395e-05,
+      "MemcpyElimination": 0.02788853645324707,
+      "MutateDataType": 0.0012311935424804688,
+      "NeuronAliasDependencyInduction": 0.0001773834228515625,
+      "NeuronAliasDependencyReset": 0.024976015090942383,
+      "NeuronInstComb": 0.005156517028808594,
+      "NeuronLICM": 0.036696434020996094,
+      "NeuronLoopFusion": 0.008457422256469727,
+      "NeuronLoopInterchange": 0.001413106918334961,
+      "NeuronSimplifier": 0.007856369018554688,
+      "NeuronSimplifyPredicates": 0.12235808372497559,
+      "NeuronValueNumbering": 0.004765748977661133,
+      "OptimizeAliasedCopyChain": 0.0006341934204101563,
+      "OptimizeNKIKernels": 0.38834357261657715,
+      "PAGLayoutOpt": 0.0889735221862793,
+      "PComputeCutting": 0.005109071731567383,
+      "PGLayoutTilingPipeline": 0.6248171329498291,
+      "PGTiling": 0.1645822525024414,
+      "PadElimination": 0.0003485679626464844,
+      "ParAxesAnnotation": 0.05196070671081543,
+      "PartialLoopFusion": 0.011112451553344727,
+      "PartialSimdFusion": 0.012138128280639648,
+      "PenguinizeFunctions": 4.3000000005122274e-05,
+      "PerfectLoopNest": 0.002288341522216797,
+      "PruneFunctions": 4.099999932805076e-05,
+      "RecognizeOpIdiom": 0.0041277408599853516,
+      "Recompute": 0.00026416778564453125,
+      "RelaxPredicates": 0.01356959342956543,
+      "Rematerialization": 0.0024864673614501953,
+      "RemoveOptimizationBarriers": 4.900000203633681e-05,
+      "ReshapeWeights": 0.0007522106170654297,
+      "ResolveAccessConflict": 0.0048482418060302734,
+      "ResolveComplicatePredicates": 0.0015094280242919922,
+      "RewriteReplicationMatmul": 0.0015668869018554688,
+      "RewriteWeights": 0.0027174949645996094,
+      "SFKVectorizer": 0.2781519889831543,
+      "ScatterMotion": 4.70000013592653e-05,
+      "SimpleAllReduceTiling": 0.009549379348754883,
+      "Simplifier": 0.003630399703979492,
+      "SimplifyMacroPredicates": 0.011396646499633789,
+      "SimplifyNeuronTensor": 1.0561063289642334,
+      "SimplifySlice": 0.0023348331451416016,
+      "SimplifyTensor": 0.005601167678833008,
+      "SpillPSum": 0.013618230819702148,
+      "SplitAPUnionSets": 0.11336159706115723,
+      "SplitAccGrp": 0.001394510269165039,
+      "StaticProfiler": 0.014252662658691406,
+      "StaticTransposeLocalTensor": 0.003930330276489258,
+      "SundaISel": 0.04436635971069336,
+      "TCTransform": 0.0008757114410400391,
+      "TensorInitialization": 0.01558232307434082,
+      "TensorOpSimplifier": 0.004608869552612305,
+      "TensorOpTransform": 0.01923346519470215,
+      "TensorizerLegalizationPass": 5.2999999752501026e-05,
+      "TileCCOps": 0.005507707595825195,
+      "TilingProfiler": 0.007405757904052734,
+      "TransformConvOp": 0.0030219554901123047,
+      "TritiumFusion": 0.05425119400024414,
+      "ValueNumbering": 0.0020017623901367188,
+      "VectorizeDMA": 0.002228975296020508,
+      "VectorizeMatMult": 0.006806135177612305,
+      "VerifySupportedOps": 3.5000000934815034e-05,
+      "WeightCoalescing": 0.008660554885864258,
+      "ZeroSizeTensorElimination": 0.00014281272888183594,
+      "algsimp": 0.0027209999971091747,
+      "batchnorm_expander": 4.099999932805076e-05,
+      "boundary-marker-removal": 1.2999998943996616e-05,
+      "call-inliner": 0.0004540000227279961,
+      "canonicalize-boundary-marker": 1.700000029813964e-05,
+      "collective-stream-id-checker": 8.000000525498763e-05,
+      "comparison-expander": 0.0005869999877177179,
+      "computation-deduplicator": 7.500000356230885e-05,
+      "conditional-to-select": 1.700000029813964e-05,
+      "config-lowering": 8.800000068731606e-05,
+      "constant-statistics": 0.0005440000095404685,
+      "constant_folding": 0.00032700004521757364,
+      "cse": 3.7000001611886546e-05,
+      "dce": 9.100000170292333e-05,
+      "dot_decomposer": 0.0013370000524446368,
+      "dynamic-slice-transpose": 1.2000000424450263e-05,
+      "eliminate-redundant-compare": 0.0003020000003743917,
+      "emit-offloaded-dropout": 3.9999998989515007e-05,
+      "flatten-call-graph": 0.0009239999344572425,
+      "fuse-send-recv": 7.79999973019585e-05,
+      "hilo::LegalizeAlias": 1.1999999514955562e-05,
+      "hilo::NeuronInstCombine": 0.00018899999849963933,
+      "hilo::NeuronOpFusion": 4.5000000682193786e-05,
+      "hilo::ReplaceTokenTypeWithU8Pass": 5.7999997807201e-05,
+      "hilo::ScheduleFusion": 0.00016099998902063817,
+      "hilo::SixtyFourHack": 6.70000008540228e-05,
+      "hilo::VerifyAliasing": 4.999999873689376e-06,
+      "hlo-mac-count": 0.0013409999664872885,
+      "hlo-verifier": 0.007716999854892492,
+      "instruction-histogram": 0.0007719999994151294,
+      "io-con-pipe-begin": 4.999999873689376e-06,
+      "io-con-pipe-end": 9.999999974752427e-07,
+      "io-layout-normalization": 0.00139999995008111,
+      "io-statistics": 6.299999949987978e-05,
+      "legalize-ccops": 3.999999989900971e-06,
+      "legalize-compare": 1.1000000085914508e-05,
+      "lower-argminmax-custom-call": 1.1000000085914508e-05,
+      "map-inline": 0.0008809999562799931,
+      "metadata-naming": 6.70000008540228e-05,
+      "mlir::detail::OpToOpPassAdaptor": 0.00020599999697878957,
+      "mlir::hlo::MhloToPyPenguin": 0.00291300006210804,
+      "mlir::mhlo::LowerComplexExtraPass": 0.00027200000477023423,
+      "mlir::mhlo::LowerComplexPass": 0.0003980000037699938,
+      "native-to-custom-softmax": 0.0007730000070296228,
+      "native-to-custom-softmax-dx": 0.0006189999985508621,
+      "operand_upcaster": 6.299999949987978e-05,
+      "opt-barrier-removal": 0.0005789999850094318,
+      "post-par-pipe-begin": 7.999999979801942e-06,
+      "post-par-pipe-end": 0.0,
+      "post-partition-simplification": 0.0017419999931007624,
+      "pre-par-pipe-begin": 1.9999999949504854e-06,
+      "pre-par-pipe-end": 0.0,
+      "pre-partition-simplification": 0.1384889930486679,
+      "replace-minimum-constant": 0.0004579999949783087,
+      "reshape-mover": 0.00011000000085914508,
+      "simplify-concat": 0.00014099999680183828,
+      "simplify-while-loops": 9.40000027185306e-05,
+      "transform-variadic-reduce": 8.100000559352338e-05,
+      "tuple-simplifier": 0.00030600003083236516,
+      "unpack-nested-aws-ntwsr": 0.000438000017311424,
+      "unroll-while-loop": 1.8999999156221747e-05,
+      "zero_sized_hlo_elimination": 0.0008750000270083547
+    },
+    "hilo": {
+      "ConstantSize": 2368805.0,
+      "HloInputCount": 475.0,
+      "HloMacCount": 206469595136.0,
+      "HloOutputCount": 73.0,
+      "IfmapSize": 8266549248.0,
+      "OfmapSize": 75497472.0,
+      "OutputsReadFromCount": 0.0,
+      "PassthroughTensorsCount": 0.0,
+      "RedundantOutputCount": 0.0,
+      "Traffic": 1751252352.0
+    },
+    "tensorizer": {
+      "DMATilingProfiler::TotalInstructionsAfterTiling": 49538.0,
+      "StaticProfiler::AifUb": 304.240234375,
+      "StaticProfiler::ArithmeticIntensityTensorizer": 278.67474365234375,
+      "StaticProfiler::AverageDmaLength": 1974.1033935546875,
+      "StaticProfiler::DDRTransferBytes": 862646080.0,
+      "StaticProfiler::InternalTransferBytes": 669456896.0,
+      "StaticProfiler::LoadExpanded": 390679.0,
+      "StaticProfiler::StoreExpanded": 7261.0,
+      "StaticProfiler::TotalDMAExpanded": 397940.0,
+      "StaticProfiler::TotalDynamicInstancesCount": 59578.0,
+      "StaticProfiler::TotalDynamicInstancesWithMmPackedCount": 59132.0,
+      "StaticProfiler::TotalLNCComm": 0.0,
+      "StaticProfiler::TotalLNCCommTransfer": 0.0,
+      "TilingProfiler::BatchnormInstructionsAfterTiling": 0.0,
+      "TilingProfiler::DmaInstructionsAfterTiling": 0.0,
+      "TilingProfiler::GenericInstructionsAfterTiling": 4.0,
+      "TilingProfiler::MatMultInstructionsAfterTiling": 28224.0,
+      "TilingProfiler::NumPfTransposes": 5.0,
+      "TilingProfiler::NumPfTransposesForIo": 1.0,
+      "TilingProfiler::NumPfTransposesForLocal": 1.0,
+      "TilingProfiler::NumPfTransposesForNonlocal": 3.0,
+      "TilingProfiler::PfTransposeInstructions": 19777.0,
+      "TilingProfiler::PfTransposeInstructionsForIo": 19008.0,
+      "TilingProfiler::PfTransposeInstructionsForLocal": 1.0,
+      "TilingProfiler::PfTransposeInstructionsForNonlocal": 768.0,
+      "TilingProfiler::ReduceInstructionsAfterTiling": 6.0,
+      "TilingProfiler::SimdInstructionsAfterTiling": 303.0,
+      "TilingProfiler::TotalInstructionsAfterTiling": 0.0,
+      "TransformConvOp::Conv1d_depthwise_bf01_oi01_bf01": 0.0,
+      "TransformConvOp::Conv2d_dw_fb01_io01_01bf_rep_nhwc_Pcinh": 0.0,
+      "TransformConvOp::Conv2d_pbp_0f1b_0i1o_01fb_experimental_1": 0.0,
+      "TransformConvOp::Conv2d_pbp_fb01_io01_01bf_experimental_1": 0.0,
+      "TransformConvOp::conv2d_column_packing": 0.0,
+      "TransformConvOp::conv2d_column_packing_1": 0.0,
+      "TransformConvOp::conv2d_column_packing_io10": 0.0,
+      "TransformConvOp::conv2d_depthwise_f01b_o01i_bf01": 0.0
+    }
+  },
+  "all": {
+    "compiletime": {
+      "algsimp": 0.002532999962568283,
+      "call-inliner": 0.00042600001324899495,
+      "collective-stream-id-checker": 6.70000008540228e-05,
+      "comparison-expander": 0.0005719999899156392,
+      "constant-statistics": 0.0005440000095404685,
+      "constant_folding": 0.0003000000142492354,
+      "dce": 8.800000068731606e-05,
+      "dot_decomposer": 0.0013370000524446368,
+      "eliminate-redundant-compare": 0.000291000003926456,
+      "flatten-call-graph": 0.0008929999894462526,
+      "hlo-mac-count": 0.0010870000114664435,
+      "hlo-verifier": 0.007048000115901232,
+      "instruction-histogram": 0.0007719999994151294,
+      "io-con-pipe-begin": 4.999999873689376e-06,
+      "io-con-pipe-end": 9.999999974752427e-07,
+      "io-layout-normalization": 0.00139999995008111,
+      "io-statistics": 6.299999949987978e-05,
+      "map-inline": 0.0008459999808110297,
+      "native-to-custom-softmax": 0.0006709999870508909,
+      "native-to-custom-softmax-dx": 0.0005300000193528831,
+      "opt-barrier-removal": 0.0005789999850094318,
+      "pre-par-pipe-begin": 1.9999999949504854e-06,
+      "pre-par-pipe-end": 0.0,
+      "pre-partition-simplification": 0.1384889930486679,
+      "replace-minimum-constant": 0.00041700000292621553,
+      "reshape-mover": 9.999999747378752e-05,
+      "simplify-while-loops": 8.800000068731606e-05,
+      "tuple-simplifier": 0.000291000003926456,
+      "unpack-nested-aws-ntwsr": 0.00042600001324899495,
+      "unroll-while-loop": 1.8999999156221747e-05,
+      "zero_sized_hlo_elimination": 0.0008750000270083547
+    }
+  },
+  "cumsum": {
+    "compiletime": {
+      "CoalesceCCOp": 0.00023508071899414063,
+      "DMALocalityOpt": 0.00017404556274414063,
+      "DMAProfiler": 0.0008785724639892578,
+      "DataStreaming": 0.0002880096435546875,
+      "DoNothing": 0.00011467933654785156,
+      "ExpandISAMacro": 0.0006787776947021484,
+      "FactorizeBlkDims": 0.0004444122314453125,
+      "InferPSumTensor": 0.0004467964172363281,
+      "LateLegalizeInst": 0.000461578369140625,
+      "LateNeuronInstComb": 0.0004818439483642578,
+      "LegalizeSundaAccess": 0.0016222000122070313,
+      "LegalizeType": 0.0002703666687011719,
+      "LowerBroadcast": 0.00025391578674316406,
+      "LowerIntrinsics": 0.00021457672119140625,
+      "LowerTranspose": 0.00024318695068359375,
+      "NeuronInstComb": 0.00048065185546875,
+      "NeuronLICM": 0.00038552284240722656,
+      "NeuronSimplifyPredicates": 0.0027823448181152344,
+      "NeuronValueNumbering": 0.00043129920959472656,
+      "SFKVectorizer": 0.003134012222290039,
+      "SimpleAllReduceTiling": 0.00022721290588378906,
+      "SimplifyNeuronTensor": 0.0005092620849609375,
+      "SpillPSum": 0.0005443096160888672,
+      "WeightCoalescing": 0.00020051002502441406
+    }
+  },
+  "sg00": {
+    "compiletime": {
+      "CanonicalizeConv": 9.999999974752427e-07,
+      "CanonicalizeForTensorizer": 1.700000029813964e-05,
+      "Canonicalizer": 0.00033599999733269215,
+      "HoistCompute": 3.000000106112566e-06,
+      "IdentifyCrossPassTensors": 1.5999999959603883e-05,
+      "MemcastMotion": 1.1000000085914508e-05,
+      "PenguinizeFunctions": 1.8000000636675395e-05,
+      "PruneFunctions": 1.4000000192027073e-05,
+      "RemoveOptimizationBarriers": 1.2999999853491317e-05,
+      "ScatterMotion": 2.4000000848900527e-05,
+      "TensorizerLegalizationPass": 2.700000004551839e-05,
+      "VerifySupportedOps": 1.2000000424450263e-05,
+      "algsimp": 6.500000017695129e-05,
+      "batchnorm_expander": 1.4000000192027073e-05,
+      "boundary-marker-removal": 3.999999989900971e-06,
+      "call-inliner": 9.000000318337698e-06,
+      "canonicalize-boundary-marker": 6.000000212225132e-06,
+      "collective-stream-id-checker": 3.999999989900971e-06,
+      "comparison-expander": 4.999999873689376e-06,
+      "computation-deduplicator": 2.300000051036477e-05,
+      "conditional-to-select": 4.999999873689376e-06,
+      "config-lowering": 3.9999998989515007e-05,
+      "constant_folding": 9.000000318337698e-06,
+      "cse": 1.2999999853491317e-05,
+      "dce": 9.999999974752427e-07,
+      "dynamic-slice-transpose": 3.999999989900971e-06,
+      "eliminate-redundant-compare": 3.999999989900971e-06,
+      "emit-offloaded-dropout": 1.4000000192027073e-05,
+      "flatten-call-graph": 9.999999747378752e-06,
+      "fuse-send-recv": 2.8000000384054147e-05,
+      "hilo::LegalizeAlias": 4.999999873689376e-06,
+      "hilo::NeuronInstCombine": 8.499999967170879e-05,
+      "hilo::NeuronOpFusion": 2.700000004551839e-05,
+      "hilo::ReplaceTokenTypeWithU8Pass": 1.4999999621068127e-05,
+      "hilo::ScheduleFusion": 9.999999974752427e-07,
+      "hilo::SixtyFourHack": 1.2999999853491317e-05,
+      "hilo::VerifyAliasing": 1.9999999949504854e-06,
+      "hlo-mac-count": 3.099999958067201e-05,
+      "hlo-verifier": 0.0002530000056140125,
+      "legalize-ccops": 9.999999974752427e-07,
+      "legalize-compare": 3.999999989900971e-06,
+      "lower-argminmax-custom-call": 3.999999989900971e-06,
+      "map-inline": 1.2000000424450263e-05,
+      "metadata-naming": 2.4000000848900527e-05,
+      "mlir::detail::OpToOpPassAdaptor": 2.2000000171829015e-05,
+      "mlir::hlo::MhloToPyPenguin": 0.0010389999952167273,
+      "mlir::mhlo::LowerComplexExtraPass": 8.800000068731606e-05,
+      "mlir::mhlo::LowerComplexPass": 0.00014200000441633165,
+      "native-to-custom-softmax": 9.000000136438757e-05,
+      "native-to-custom-softmax-dx": 4.3000000005122274e-05,
+      "operand_upcaster": 2.300000051036477e-05,
+      "post-par-pipe-begin": 3.000000106112566e-06,
+      "post-par-pipe-end": 0.0,
+      "post-partition-simplification": 0.0006249999860301614,
+      "replace-minimum-constant": 2.5999999706982635e-05,
+      "reshape-mover": 3.999999989900971e-06,
+      "simplify-concat": 4.8000001697801054e-05,
+      "simplify-while-loops": 1.9999999949504854e-06,
+      "transform-variadic-reduce": 9.000000318337698e-06,
+      "tuple-simplifier": 4.999999873689376e-06,
+      "unpack-nested-aws-ntwsr": 3.999999989900971e-06,
+      "unroll-while-loop": 0.0
+    },
+    "hilo": {
+      "ArithmeticIntensity": 73.02900695800781,
+      "ConstantSize": 2368805.0,
+      "HloInputCount": 475.0,
+      "HloMacCount": 25769803776.0,
+      "HloOutputCount": 73.0,
+      "IfmapSize": 8266549248.0,
+      "OfmapSize": 75497472.0,
+      "OutputsReadFromCount": 0.0,
+      "PassthroughTensorsCount": 0.0,
+      "RedundantOutputCount": 0.0,
+      "Traffic": 705741632.0
+    }
+  },
+  "sg0000": {
+    "compiletime": {
+      "AGOrderingAnalysisPass": 0.0818486213684082,
+      "AffinePredicateResolution": 0.001665353775024414,
+      "AliasDependencyElimination": 0.00012683868408203125,
+      "AliasDependencyInduction": 0.008559942245483398,
+      "AliasDependencyReset": 0.03254294395446777,
+      "BFComputeCutting": 0.003969907760620117,
+      "BirCodeGenLoop": 0.06339025497436523,
+      "CCOpFusion": 0.029911041259765625,
+      "CanonicalizeDAGForPGTiling": 0.003092050552368164,
+      "CanonicalizeIR": 0.002637147903442383,
+      "CoalesceCCOp": 0.0051479339599609375,
+      "CommuteConcat": 0.001478433609008789,
+      "DMALocalityOpt": 0.0016412734985351563,
+      "DMAProfiler": 0.004613637924194336,
+      "DMATilingProfiler": 0.004850864410400391,
+      "DataLocalityOpt": 0.11357831954956055,
+      "DataStreaming": 0.0061092376708984375,
+      "DeConcat": 0.0013332366943359375,
+      "DeadCodeElimination": 0.0018727779388427734,
+      "DeadStoreElimination": 0.03094482421875,
+      "DelinearIndices": 0.008640289306640625,
+      "Delinearization": 0.0035429000854492188,
+      "DoNothing": 8.106231689453125e-05,
+      "DramToDramTranspose": 0.03549051284790039,
+      "DumpGraphAndMetadata": 0.005577564239501953,
+      "EliminateDivs": 0.003966331481933594,
+      "ExpandBatchNorm": 0.0017447471618652344,
+      "ExpandISAMacro": 0.002687692642211914,
+      "FactorizeBlkDims": 0.026469945907592773,
+      "FactorizeThreadAxesInFreeDims": 0.0014863014221191406,
+      "FlattenMacroLoop": 0.00392913818359375,
+      "GenericAccessSimplifier": 0.0018973350524902344,
+      "InferInitValue": 0.03517007827758789,
+      "InferIntrinsicOnCC": 0.010237932205200195,
+      "InferNeuronTensor": 0.051462411880493164,
+      "InferNonlocalTensors": 0.14991235733032227,
+      "InferPSumTensor": 0.053685903549194336,
+      "InlineNativeKernels": 0.002433300018310547,
+      "InsertIOTransposes": 0.015550613403320313,
+      "InsertLocalTransposes": 0.007843017578125,
+      "InsertOffloadedTransposes": 0.002854585647583008,
+      "LICM": 0.003381490707397461,
+      "LateLegalizeInst": 0.0069310665130615234,
+      "LateLegalizePostSplit": 0.00308990478515625,
+      "LateLowerReshapeOp": 0.0017940998077392578,
+      "LateLowerTensorOp": 0.005001068115234375,
+      "LateNeuronInstComb": 0.016704320907592773,
+      "LayoutPreprocessing": 0.033296823501586914,
+      "LayoutPreprocessingAndAnalysis": 0.12302517890930176,
+      "LayoutRequirementAnalysis": 0.007364988327026367,
+      "LegalizeCCOpLayout": 0.0029296875,
+      "LegalizeOpLevelAlias": 0.0016987323760986328,
+      "LegalizePartitionReduce": 0.0014727115631103516,
+      "LegalizeSundaAccess": 0.04025077819824219,
+      "LegalizeSundaMacro": 0.009906291961669922,
+      "LegalizeType": 0.004493236541748047,
+      "LocalLayoutOpt": 0.017308473587036133,
+      "LoopFusion": 0.005831241607666016,
+      "LoopSplitting": 0.00037789344787597656,
+      "LowerBroadcast": 0.0016851425170898438,
+      "LowerCCOpBlockAxis": 0.005655765533447266,
+      "LowerComplexBroadcast": 0.0020987987518310547,
+      "LowerIntrinsics": 0.040236473083496094,
+      "LowerTensorOp": 0.012641191482543945,
+      "LowerTranspose": 0.0125579833984375,
+      "MacroGeneration": 0.08074021339416504,
+      "MaskPropagation": 0.005038022994995117,
+      "MemcpyElimination": 0.10875082015991211,
+      "MutateDataType": 0.0013315677642822266,
+      "NeuronAliasDependencyInduction": 0.00025200843811035156,
+      "NeuronAliasDependencyReset": 0.021958112716674805,
+      "NeuronInstComb": 0.009703636169433594,
+      "NeuronLICM": 0.011526823043823242,
+      "NeuronLoopFusion": 0.017663955688476563,
+      "NeuronLoopInterchange": 0.002567291259765625,
+      "NeuronSimplifier": 0.011670589447021484,
+      "NeuronSimplifyPredicates": 0.017385244369506836,
+      "NeuronValueNumbering": 0.004181623458862305,
+      "OptimizeAliasedCopyChain": 0.0017867088317871094,
+      "OptimizeNKIKernels": 0.0020456314086914063,
+      "PAGLayoutOpt": 0.3681519031524658,
+      "PComputeCutting": 0.008620262145996094,
+      "PGLayoutTilingPipeline": 1.3210320472717285,
+      "PGTiling": 0.27039527893066406,
+      "PadElimination": 0.0003745555877685547,
+      "ParAxesAnnotation": 0.33005595207214355,
+      "PartialLoopFusion": 0.026912212371826172,
+      "PartialSimdFusion": 0.03544425964355469,
+      "PerfectLoopNest": 0.0021703243255615234,
+      "RecognizeOpIdiom": 0.004334926605224609,
+      "Recompute": 0.0002522468566894531,
+      "RelaxPredicates": 0.004270076751708984,
+      "Rematerialization": 0.005487918853759766,
+      "ReshapeWeights": 0.0006825923919677734,
+      "ResolveAccessConflict": 0.003779888153076172,
+      "ResolveComplicatePredicates": 0.0018131732940673828,
+      "RewriteReplicationMatmul": 0.002633333206176758,
+      "RewriteWeights": 0.0036499500274658203,
+      "SFKVectorizer": 0.2772994041442871,
+      "SimpleAllReduceTiling": 0.002454519271850586,
+      "Simplifier": 0.0045070648193359375,
+      "SimplifyMacroPredicates": 0.016190290451049805,
+      "SimplifyNeuronTensor": 0.01452183723449707,
+      "SimplifySlice": 0.0010039806365966797,
+      "SimplifyTensor": 0.00657200813293457,
+      "SpillPSum": 0.02208685874938965,
+      "SplitAPUnionSets": 0.04095458984375,
+      "SplitAccGrp": 0.0018160343170166016,
+      "StaticProfiler": 0.004816770553588867,
+      "StaticTransposeLocalTensor": 0.004886150360107422,
+      "SundaISel": 0.04611611366271973,
+      "TCTransform": 0.001667022705078125,
+      "TensorInitialization": 0.022374629974365234,
+      "TensorOpSimplifier": 0.006697177886962891,
+      "TensorOpTransform": 0.02793574333190918,
+      "TileCCOps": 0.007641792297363281,
+      "TilingProfiler": 0.015750885009765625,
+      "TransformConvOp": 0.0026845932006835938,
+      "TritiumFusion": 0.08186149597167969,
+      "ValueNumbering": 0.0026755332946777344,
+      "VectorizeDMA": 0.007223367691040039,
+      "VectorizeMatMult": 0.018305540084838867,
+      "WeightCoalescing": 0.003328561782836914,
+      "ZeroSizeTensorElimination": 0.00011229515075683594
+    },
+    "tensorizer": {
+      "DMATilingProfiler::TotalInstructionsAfterTiling": 5862.0,
+      "StaticProfiler::AifUb": 88.59026336669922,
+      "StaticProfiler::ArithmeticIntensityTensorizer": 582.7418823242188,
+      "StaticProfiler::AverageDmaLength": 2248.2685546875,
+      "StaticProfiler::AverageFractalPeUtilization": 99.96076202392578,
+      "StaticProfiler::AveragePartitionUtilization": 99.90216827392578,
+      "StaticProfiler::AveragePeUtilization": 99.8394546508789,
+      "StaticProfiler::DDRTransferBytes": 104424704.0,
+      "StaticProfiler::InternalTransferBytes": 122421248.0,
+      "StaticProfiler::LoadExpanded": 25346.0,
+      "StaticProfiler::LocalizationEfficiency": 657.7944946289063,
+      "StaticProfiler::LocalizationEfficiencyIgnoreNonlocal": 915.0787353515625,
+      "StaticProfiler::StoreExpanded": 10753.0,
+      "StaticProfiler::TotalDMAExpanded": 36099.0,
+      "StaticProfiler::TotalDynamicInstancesCount": 8866.0,
+      "StaticProfiler::TotalDynamicInstancesWithMmPackedCount": 8860.0,
+      "StaticProfiler::TotalLNCComm": 0.0,
+      "StaticProfiler::TotalLNCCommTransfer": 0.0,
+      "TilingProfiler::AveragePartitionUtilizationAfterTiling": 0.0,
+      "TilingProfiler::AveragePeUtilizationAfterTiling": 0.0,
+      "TilingProfiler::BatchnormInstructionsAfterTiling": 0.0,
+      "TilingProfiler::DmaInstructionsAfterTiling": 0.0,
+      "TilingProfiler::GenericInstructionsAfterTiling": 96.0,
+      "TilingProfiler::MatMultInstructionsAfterTiling": 3080.0,
+      "TilingProfiler::NumPfTransposes": 8.0,
+      "TilingProfiler::NumPfTransposesForIo": 1.0,
+      "TilingProfiler::NumPfTransposesForLocal": 6.0,
+      "TilingProfiler::NumPfTransposesForNonlocal": 1.0,
+      "TilingProfiler::PfTransposeInstructions": 1760.0,
+      "TilingProfiler::PfTransposeInstructionsForIo": 256.0,
+      "TilingProfiler::PfTransposeInstructionsForLocal": 1376.0,
+      "TilingProfiler::PfTransposeInstructionsForNonlocal": 128.0,
+      "TilingProfiler::ReduceInstructionsAfterTiling": 0.0,
+      "TilingProfiler::SimdInstructionsAfterTiling": 649.0,
+      "TilingProfiler::TotalInstructionsAfterTiling": 0.0,
+      "TransformConvOp::Conv1d_depthwise_bf01_oi01_bf01": 0.0,
+      "TransformConvOp::Conv2d_dw_fb01_io01_01bf_rep_nhwc_Pcinh": 0.0,
+      "TransformConvOp::Conv2d_pbp_0f1b_0i1o_01fb_experimental_1": 0.0,
+      "TransformConvOp::Conv2d_pbp_fb01_io01_01bf_experimental_1": 0.0,
+      "TransformConvOp::conv2d_column_packing": 0.0,
+      "TransformConvOp::conv2d_column_packing_1": 0.0,
+      "TransformConvOp::conv2d_column_packing_io10": 0.0,
+      "TransformConvOp::conv2d_depthwise_f01b_o01i_bf01": 0.0
+    }
+  },
+  "sg0001": {
+    "compiletime": {
+      "AGOrderingAnalysisPass": 0.03383040428161621,
+      "AffinePredicateResolution": 0.0015320777893066406,
+      "AliasDependencyElimination": 0.0001316070556640625,
+      "AliasDependencyInduction": 0.00819253921508789,
+      "AliasDependencyReset": 0.02862405776977539,
+      "BFComputeCutting": 0.004217624664306641,
+      "BirCodeGenLoop": 0.0443270206451416,
+      "CCOpFusion": 0.04336118698120117,
+      "CanonicalizeDAGForPGTiling": 0.0031616687774658203,
+      "CanonicalizeIR": 0.0021500587463378906,
+      "CoalesceCCOp": 0.005389690399169922,
+      "CommuteConcat": 0.0024237632751464844,
+      "DMALocalityOpt": 0.002274751663208008,
+      "DMAProfiler": 0.003973484039306641,
+      "DMATilingProfiler": 0.005924701690673828,
+      "DataLocalityOpt": 0.15027260780334473,
+      "DataStreaming": 0.004762887954711914,
+      "DeConcat": 0.0018739700317382813,
+      "DeadCodeElimination": 0.001882314682006836,
+      "DeadStoreElimination": 0.03486776351928711,
+      "DelinearIndices": 0.009628534317016602,
+      "Delinearization": 0.0037381649017333984,
+      "DoNothing": 6.985664367675781e-05,
+      "DramToDramTranspose": 0.04212188720703125,
+      "DumpGraphAndMetadata": 0.004312038421630859,
+      "EliminateDivs": 0.005432844161987305,
+      "ExpandBatchNorm": 0.002119302749633789,
+      "ExpandISAMacro": 0.0024309158325195313,
+      "FactorizeBlkDims": 0.02235579490661621,
+      "FactorizeThreadAxesInFreeDims": 0.0018169879913330078,
+      "FlattenMacroLoop": 0.0030968189239501953,
+      "GenericAccessSimplifier": 0.0016777515411376953,
+      "InferInitValue": 0.043079376220703125,
+      "InferIntrinsicOnCC": 0.009890556335449219,
+      "InferNeuronTensor": 0.05600404739379883,
+      "InferNonlocalTensors": 0.03101515769958496,
+      "InferPSumTensor": 0.04645681381225586,
+      "InlineNativeKernels": 0.0015399456024169922,
+      "InsertIOTransposes": 0.02417731285095215,
+      "InsertLocalTransposes": 0.0070497989654541016,
+      "InsertOffloadedTransposes": 0.003525972366333008,
+      "LICM": 0.0035805702209472656,
+      "LateLegalizeInst": 0.0041539669036865234,
+      "LateLegalizePostSplit": 0.0027403831481933594,
+      "LateLowerReshapeOp": 0.0014560222625732422,
+      "LateLowerTensorOp": 0.004617452621459961,
+      "LateNeuronInstComb": 0.015344619750976563,
+      "LayoutPreprocessing": 0.030884981155395508,
+      "LayoutPreprocessingAndAnalysis": 0.06435275077819824,
+      "LayoutRequirementAnalysis": 0.007463693618774414,
+      "LegalizeCCOpLayout": 0.002064943313598633,
+      "LegalizeOpLevelAlias": 0.0011925697326660156,
+      "LegalizePartitionReduce": 0.0026116371154785156,
+      "LegalizeSundaAccess": 0.015822887420654297,
+      "LegalizeSundaMacro": 0.012560844421386719,
+      "LegalizeType": 0.004744291305541992,
+      "LocalLayoutOpt": 0.023772239685058594,
+      "LoopFusion": 0.0066835880279541016,
+      "LoopSplitting": 0.0003638267517089844,
+      "LowerBroadcast": 0.002238750457763672,
+      "LowerCCOpBlockAxis": 0.005678653717041016,
+      "LowerComplexBroadcast": 0.0019271373748779297,
+      "LowerIntrinsics": 0.042801856994628906,
+      "LowerTensorOp": 0.012106895446777344,
+      "LowerTranspose": 0.012960433959960938,
+      "MacroGeneration": 0.12800955772399902,
+      "MaskPropagation": 0.0031516551971435547,
+      "MemcpyElimination": 0.10379505157470703,
+      "MutateDataType": 0.0014393329620361328,
+      "NeuronAliasDependencyInduction": 0.00022101402282714844,
+      "NeuronAliasDependencyReset": 0.020102262496948242,
+      "NeuronInstComb": 0.009283781051635742,
+      "NeuronLICM": 0.009867429733276367,
+      "NeuronLoopFusion": 0.022713661193847656,
+      "NeuronLoopInterchange": 0.002709627151489258,
+      "NeuronSimplifier": 0.01328134536743164,
+      "NeuronSimplifyPredicates": 0.001683950424194336,
+      "NeuronValueNumbering": 0.0033235549926757813,
+      "OptimizeAliasedCopyChain": 0.0007724761962890625,
+      "OptimizeNKIKernels": 0.001729726791381836,
+      "PAGLayoutOpt": 0.13172507286071777,
+      "PComputeCutting": 0.007474422454833984,
+      "PGLayoutTilingPipeline": 0.9329550266265869,
+      "PGTiling": 0.4518747329711914,
+      "PadElimination": 0.00040411949157714844,
+      "ParAxesAnnotation": 0.0915369987487793,
+      "PartialLoopFusion": 0.020573854446411133,
+      "PartialSimdFusion": 0.04284977912902832,
+      "PerfectLoopNest": 0.002377033233642578,
+      "RecognizeOpIdiom": 0.0049991607666015625,
+      "Recompute": 0.00026345252990722656,
+      "RelaxPredicates": 0.0034220218658447266,
+      "Rematerialization": 0.0021615028381347656,
+      "ReshapeWeights": 0.0007557868957519531,
+      "ResolveAccessConflict": 0.004181861877441406,
+      "ResolveComplicatePredicates": 0.0015151500701904297,
+      "RewriteReplicationMatmul": 0.0020759105682373047,
+      "RewriteWeights": 0.0036649703979492188,
+      "SFKVectorizer": 0.20148277282714844,
+      "SimpleAllReduceTiling": 0.003732442855834961,
+      "Simplifier": 0.004697084426879883,
+      "SimplifyMacroPredicates": 0.007361888885498047,
+      "SimplifyNeuronTensor": 0.009825944900512695,
+      "SimplifySlice": 0.0017888545989990234,
+      "SimplifyTensor": 0.006832122802734375,
+      "SpillPSum": 0.022799968719482422,
+      "SplitAPUnionSets": 0.020108938217163086,
+      "SplitAccGrp": 0.0015766620635986328,
+      "StaticProfiler": 0.004146099090576172,
+      "StaticTransposeLocalTensor": 0.004926919937133789,
+      "SundaISel": 0.04472494125366211,
+      "TCTransform": 0.0018138885498046875,
+      "TensorInitialization": 0.004791736602783203,
+      "TensorOpSimplifier": 0.0064849853515625,
+      "TensorOpTransform": 0.0333099365234375,
+      "TileCCOps": 0.0056035518646240234,
+      "TilingProfiler": 0.01600933074951172,
+      "TransformConvOp": 0.002446413040161133,
+      "TritiumFusion": 0.1239166259765625,
+      "ValueNumbering": 0.0030901432037353516,
+      "VectorizeDMA": 0.0017311573028564453,
+      "VectorizeMatMult": 0.018932580947875977,
+      "WeightCoalescing": 0.0027513504028320313,
+      "ZeroSizeTensorElimination": 0.00011587142944335938
+    },
+    "tensorizer": {
+      "DMATilingProfiler::TotalInstructionsAfterTiling": 15811.0,
+      "StaticProfiler::AifUb": 934.4357299804688,
+      "StaticProfiler::ArithmeticIntensityTensorizer": 708.8487548828125,
+      "StaticProfiler::AverageDmaLength": 1109.3380126953125,
+      "StaticProfiler::AverageFractalPeUtilization": 100.0,
+      "StaticProfiler::AveragePartitionUtilization": 99.8372802734375,
+      "StaticProfiler::AveragePeUtilization": 100.0,
+      "StaticProfiler::DDRTransferBytes": 306283520.0,
+      "StaticProfiler::InternalTransferBytes": 104595456.0,
+      "StaticProfiler::LoadExpanded": 257536.0,
+      "StaticProfiler::LocalizationEfficiency": 75.85848236083984,
+      "StaticProfiler::LocalizationEfficiencyIgnoreNonlocal": 85.1915054321289,
+      "StaticProfiler::StoreExpanded": 10241.0,
+      "StaticProfiler::TotalDMAExpanded": 267777.0,
+      "StaticProfiler::TotalDynamicInstancesCount": 19667.0,
+      "StaticProfiler::TotalDynamicInstancesWithMmPackedCount": 19667.0,
+      "StaticProfiler::TotalLNCComm": 0.0,
+      "StaticProfiler::TotalLNCCommTransfer": 0.0,
+      "TilingProfiler::AveragePartitionUtilizationAfterTiling": 0.0,
+      "TilingProfiler::AveragePeUtilizationAfterTiling": 0.0,
+      "TilingProfiler::BatchnormInstructionsAfterTiling": 0.0,
+      "TilingProfiler::DmaInstructionsAfterTiling": 0.0,
+      "TilingProfiler::GenericInstructionsAfterTiling": 64.0,
+      "TilingProfiler::MatMultInstructionsAfterTiling": 12288.0,
+      "TilingProfiler::NumPfTransposes": 9.0,
+      "TilingProfiler::NumPfTransposesForIo": 3.0,
+      "TilingProfiler::NumPfTransposesForLocal": 4.0,
+      "TilingProfiler::NumPfTransposesForNonlocal": 2.0,
+      "TilingProfiler::PfTransposeInstructions": 1904.0,
+      "TilingProfiler::PfTransposeInstructionsForIo": 272.0,
+      "TilingProfiler::PfTransposeInstructionsForLocal": 1120.0,
+      "TilingProfiler::PfTransposeInstructionsForNonlocal": 512.0,
+      "TilingProfiler::ReduceInstructionsAfterTiling": 0.0,
+      "TilingProfiler::SimdInstructionsAfterTiling": 704.0,
+      "TilingProfiler::TotalInstructionsAfterTiling": 0.0,
+      "TransformConvOp::Conv1d_depthwise_bf01_oi01_bf01": 0.0,
+      "TransformConvOp::Conv2d_dw_fb01_io01_01bf_rep_nhwc_Pcinh": 0.0,
+      "TransformConvOp::Conv2d_pbp_0f1b_0i1o_01fb_experimental_1": 0.0,
+      "TransformConvOp::Conv2d_pbp_fb01_io01_01bf_experimental_1": 0.0,
+      "TransformConvOp::conv2d_column_packing": 0.0,
+      "TransformConvOp::conv2d_column_packing_1": 0.0,
+      "TransformConvOp::conv2d_column_packing_io10": 0.0,
+      "TransformConvOp::conv2d_depthwise_f01b_o01i_bf01": 0.0
+    }
+  },
+  "sg0002": {
+    "compiletime": {
+      "AGOrderingAnalysisPass": 0.01837611198425293,
+      "AffinePredicateResolution": 0.0011184215545654297,
+      "AliasDependencyElimination": 0.00015664100646972656,
+      "AliasDependencyInduction": 0.005170583724975586,
+      "AliasDependencyReset": 0.027508020401000977,
+      "BFComputeCutting": 0.0036101341247558594,
+      "BirCodeGenLoop": 0.4774467945098877,
+      "CCOpFusion": 0.033265113830566406,
+      "CanonicalizeDAGForPGTiling": 0.004282712936401367,
+      "CanonicalizeIR": 0.0024569034576416016,
+      "CoalesceCCOp": 0.013993978500366211,
+      "CommuteConcat": 0.0017316341400146484,
+      "DMALocalityOpt": 0.005455970764160156,
+      "DMAProfiler": 0.012103080749511719,
+      "DMATilingProfiler": 0.0037560462951660156,
+      "DataLocalityOpt": 0.07645320892333984,
+      "DataStreaming": 0.03701519966125488,
+      "DeConcat": 0.0018520355224609375,
+      "DeadCodeElimination": 0.0020148754119873047,
+      "DeadStoreElimination": 0.006912708282470703,
+      "DelinearIndices": 0.004647254943847656,
+      "Delinearization": 0.003908872604370117,
+      "DoNothing": 7.414817810058594e-05,
+      "DramToDramTranspose": 0.02015542984008789,
+      "DumpGraphAndMetadata": 0.08691883087158203,
+      "EliminateDivs": 0.0025060176849365234,
+      "ExpandBatchNorm": 0.0027189254760742188,
+      "ExpandISAMacro": 0.010967254638671875,
+      "FactorizeBlkDims": 0.009678840637207031,
+      "FactorizeThreadAxesInFreeDims": 0.0023202896118164063,
+      "FlattenMacroLoop": 0.00232696533203125,
+      "GenericAccessSimplifier": 0.0008094310760498047,
+      "InferInitValue": 0.02833867073059082,
+      "InferIntrinsicOnCC": 0.008923768997192383,
+      "InferNeuronTensor": 0.025766372680664063,
+      "InferNonlocalTensors": 0.014599800109863281,
+      "InferPSumTensor": 0.28374218940734863,
+      "InlineNativeKernels": 0.00860905647277832,
+      "InsertIOTransposes": 0.01989889144897461,
+      "InsertLocalTransposes": 0.004229307174682617,
+      "InsertOffloadedTransposes": 0.0029871463775634766,
+      "LICM": 0.0030870437622070313,
+      "LateLegalizeInst": 0.01364445686340332,
+      "LateLegalizePostSplit": 0.014872312545776367,
+      "LateLowerReshapeOp": 0.0010464191436767578,
+      "LateLowerTensorOp": 0.002707242965698242,
+      "LateNeuronInstComb": 0.01008152961730957,
+      "LayoutPreprocessing": 0.026853561401367188,
+      "LayoutPreprocessingAndAnalysis": 0.0556035041809082,
+      "LayoutRequirementAnalysis": 0.004946470260620117,
+      "LegalizeCCOpLayout": 0.0025353431701660156,
+      "LegalizeOpLevelAlias": 0.0018966197967529297,
+      "LegalizePartitionReduce": 0.0017490386962890625,
+      "LegalizeSundaAccess": 0.0763850212097168,
+      "LegalizeSundaMacro": 0.012125253677368164,
+      "LegalizeType": 0.012414693832397461,
+      "LocalLayoutOpt": 0.013860225677490234,
+      "LoopFusion": 0.005201578140258789,
+      "LoopSplitting": 0.0003204345703125,
+      "LowerBroadcast": 0.0018322467803955078,
+      "LowerCCOpBlockAxis": 0.0040171146392822266,
+      "LowerComplexBroadcast": 0.002280712127685547,
+      "LowerIntrinsics": 0.3141806125640869,
+      "LowerTensorOp": 0.01141357421875,
+      "LowerTranspose": 0.012679815292358398,
+      "MacroGeneration": 0.034410953521728516,
+      "MaskPropagation": 0.0028192996978759766,
+      "MemcpyElimination": 0.02788853645324707,
+      "MutateDataType": 0.0012311935424804688,
+      "NeuronAliasDependencyInduction": 0.0001773834228515625,
+      "NeuronAliasDependencyReset": 0.024976015090942383,
+      "NeuronInstComb": 0.004675865173339844,
+      "NeuronLICM": 0.03631091117858887,
+      "NeuronLoopFusion": 0.008457422256469727,
+      "NeuronLoopInterchange": 0.001413106918334961,
+      "NeuronSimplifier": 0.007856369018554688,
+      "NeuronSimplifyPredicates": 0.11957573890686035,
+      "NeuronValueNumbering": 0.004334449768066406,
+      "OptimizeAliasedCopyChain": 0.0006341934204101563,
+      "OptimizeNKIKernels": 0.38834357261657715,
+      "PAGLayoutOpt": 0.0889735221862793,
+      "PComputeCutting": 0.005109071731567383,
+      "PGLayoutTilingPipeline": 0.6248171329498291,
+      "PGTiling": 0.1645822525024414,
+      "PadElimination": 0.0003485679626464844,
+      "ParAxesAnnotation": 0.05196070671081543,
+      "PartialLoopFusion": 0.011112451553344727,
+      "PartialSimdFusion": 0.012138128280639648,
+      "PerfectLoopNest": 0.002288341522216797,
+      "RecognizeOpIdiom": 0.0041277408599853516,
+      "Recompute": 0.00026416778564453125,
+      "RelaxPredicates": 0.01356959342956543,
+      "Rematerialization": 0.0024864673614501953,
+      "ReshapeWeights": 0.0007522106170654297,
+      "ResolveAccessConflict": 0.0048482418060302734,
+      "ResolveComplicatePredicates": 0.0015094280242919922,
+      "RewriteReplicationMatmul": 0.0015668869018554688,
+      "RewriteWeights": 0.0027174949645996094,
+      "SFKVectorizer": 0.27501797676086426,
+      "SimpleAllReduceTiling": 0.009322166442871094,
+      "Simplifier": 0.003630399703979492,
+      "SimplifyMacroPredicates": 0.011396646499633789,
+      "SimplifyNeuronTensor": 1.0555970668792725,
+      "SimplifySlice": 0.0023348331451416016,
+      "SimplifyTensor": 0.005601167678833008,
+      "SpillPSum": 0.013073921203613281,
+      "SplitAPUnionSets": 0.11336159706115723,
+      "SplitAccGrp": 0.001394510269165039,
+      "StaticProfiler": 0.014252662658691406,
+      "StaticTransposeLocalTensor": 0.003930330276489258,
+      "SundaISel": 0.04436635971069336,
+      "TCTransform": 0.0008757114410400391,
+      "TensorInitialization": 0.01558232307434082,
+      "TensorOpSimplifier": 0.004608869552612305,
+      "TensorOpTransform": 0.01923346519470215,
+      "TileCCOps": 0.005507707595825195,
+      "TilingProfiler": 0.007405757904052734,
+      "TransformConvOp": 0.0030219554901123047,
+      "TritiumFusion": 0.05425119400024414,
+      "ValueNumbering": 0.0020017623901367188,
+      "VectorizeDMA": 0.002228975296020508,
+      "VectorizeMatMult": 0.006806135177612305,
+      "WeightCoalescing": 0.008460044860839844,
+      "ZeroSizeTensorElimination": 0.00014281272888183594
+    },
+    "tensorizer": {
+      "DMATilingProfiler::TotalInstructionsAfterTiling": 49538.0,
+      "StaticProfiler::AifUb": 304.240234375,
+      "StaticProfiler::ArithmeticIntensityTensorizer": 278.67474365234375,
+      "StaticProfiler::AverageDmaLength": 1974.1033935546875,
+      "StaticProfiler::AverageFractalPeUtilization": 99.7004623413086,
+      "StaticProfiler::AveragePartitionUtilization": 97.94140625,
+      "StaticProfiler::AveragePeUtilization": 98.78884887695313,
+      "StaticProfiler::DDRTransferBytes": 862646080.0,
+      "StaticProfiler::InternalTransferBytes": 669456896.0,
+      "StaticProfiler::LoadExpanded": 390679.0,
+      "StaticProfiler::LocalizationEfficiency": 91.59693145751953,
+      "StaticProfiler::LocalizationEfficiencyIgnoreNonlocal": 95.863037109375,
+      "StaticProfiler::StoreExpanded": 7261.0,
+      "StaticProfiler::TotalDMAExpanded": 397940.0,
+      "StaticProfiler::TotalDynamicInstancesCount": 59578.0,
+      "StaticProfiler::TotalDynamicInstancesWithMmPackedCount": 59132.0,
+      "StaticProfiler::TotalLNCComm": 0.0,
+      "StaticProfiler::TotalLNCCommTransfer": 0.0,
+      "TilingProfiler::AveragePartitionUtilizationAfterTiling": 0.0,
+      "TilingProfiler::AveragePeUtilizationAfterTiling": 0.0,
+      "TilingProfiler::BatchnormInstructionsAfterTiling": 0.0,
+      "TilingProfiler::DmaInstructionsAfterTiling": 0.0,
+      "TilingProfiler::GenericInstructionsAfterTiling": 4.0,
+      "TilingProfiler::MatMultInstructionsAfterTiling": 28224.0,
+      "TilingProfiler::NumPfTransposes": 5.0,
+      "TilingProfiler::NumPfTransposesForIo": 1.0,
+      "TilingProfiler::NumPfTransposesForLocal": 1.0,
+      "TilingProfiler::NumPfTransposesForNonlocal": 3.0,
+      "TilingProfiler::PfTransposeInstructions": 19777.0,
+      "TilingProfiler::PfTransposeInstructionsForIo": 19008.0,
+      "TilingProfiler::PfTransposeInstructionsForLocal": 1.0,
+      "TilingProfiler::PfTransposeInstructionsForNonlocal": 768.0,
+      "TilingProfiler::ReduceInstructionsAfterTiling": 6.0,
+      "TilingProfiler::SimdInstructionsAfterTiling": 303.0,
+      "TilingProfiler::TotalInstructionsAfterTiling": 0.0,
+      "TransformConvOp::Conv1d_depthwise_bf01_oi01_bf01": 0.0,
+      "TransformConvOp::Conv2d_dw_fb01_io01_01bf_rep_nhwc_Pcinh": 0.0,
+      "TransformConvOp::Conv2d_pbp_0f1b_0i1o_01fb_experimental_1": 0.0,
+      "TransformConvOp::Conv2d_pbp_fb01_io01_01bf_experimental_1": 0.0,
+      "TransformConvOp::conv2d_column_packing": 0.0,
+      "TransformConvOp::conv2d_column_packing_1": 0.0,
+      "TransformConvOp::conv2d_column_packing_io10": 0.0,
+      "TransformConvOp::conv2d_depthwise_f01b_o01i_bf01": 0.0
+    }
+  },
+  "sg01": {
+    "compiletime": {
+      "CanonicalizeConv": 2.2000000171829015e-05,
+      "CanonicalizeForTensorizer": 1.4999999621068127e-05,
+      "Canonicalizer": 0.00025499999173916876,
+      "HoistCompute": 1.9999999949504854e-06,
+      "IdentifyCrossPassTensors": 2.499999936844688e-05,
+      "MemcastMotion": 7.000000096013537e-06,
+      "PenguinizeFunctions": 1.4999999621068127e-05,
+      "PruneFunctions": 1.8999999156221747e-05,
+      "RemoveOptimizationBarriers": 2.700000004551839e-05,
+      "ScatterMotion": 1.9999999494757503e-05,
+      "TensorizerLegalizationPass": 1.9999999494757503e-05,
+      "VerifySupportedOps": 1.1000000085914508e-05,
+      "algsimp": 6.299999949987978e-05,
+      "batchnorm_expander": 1.4000000192027073e-05,
+      "boundary-marker-removal": 4.999999873689376e-06,
+      "call-inliner": 9.000000318337698e-06,
+      "canonicalize-boundary-marker": 6.000000212225132e-06,
+      "collective-stream-id-checker": 4.999999873689376e-06,
+      "comparison-expander": 4.999999873689376e-06,
+      "computation-deduplicator": 2.5999999706982635e-05,
+      "conditional-to-select": 4.999999873689376e-06,
+      "config-lowering": 2.2000000171829015e-05,
+      "constant_folding": 9.000000318337698e-06,
+      "cse": 1.2000000424450263e-05,
+      "dce": 9.999999974752427e-07,
+      "dynamic-slice-transpose": 3.999999989900971e-06,
+      "eliminate-redundant-compare": 3.999999989900971e-06,
+      "emit-offloaded-dropout": 1.2999999853491317e-05,
+      "flatten-call-graph": 9.000000318337698e-06,
+      "fuse-send-recv": 2.9999999242136255e-05,
+      "hilo::LegalizeAlias": 4.999999873689376e-06,
+      "hilo::NeuronInstCombine": 3.600000127335079e-05,
+      "hilo::NeuronOpFusion": 1.4000000192027073e-05,
+      "hilo::ReplaceTokenTypeWithU8Pass": 2.099999983329326e-05,
+      "hilo::ScheduleFusion": 9.999999974752427e-07,
+      "hilo::SixtyFourHack": 1.4000000192027073e-05,
+      "hilo::VerifyAliasing": 1.9999999949504854e-06,
+      "hlo-mac-count": 4.600000102072954e-05,
+      "hlo-verifier": 0.00023299999884329736,
+      "legalize-ccops": 9.999999974752427e-07,
+      "legalize-compare": 3.999999989900971e-06,
+      "lower-argminmax-custom-call": 3.999999989900971e-06,
+      "map-inline": 1.1000000085914508e-05,
+      "metadata-naming": 2.700000004551839e-05,
+      "mlir::detail::OpToOpPassAdaptor": 0.00017299999308306724,
+      "mlir::hlo::MhloToPyPenguin": 0.0009840000420808792,
+      "mlir::mhlo::LowerComplexExtraPass": 9.600000339560211e-05,
+      "mlir::mhlo::LowerComplexPass": 0.00013600000238511711,
+      "native-to-custom-softmax": 6.000000212225132e-06,
+      "native-to-custom-softmax-dx": 2.2000000171829015e-05,
+      "operand_upcaster": 2.4000000848900527e-05,
+      "post-par-pipe-begin": 3.000000106112566e-06,
+      "post-par-pipe-end": 0.0,
+      "post-partition-simplification": 0.0005660000024363399,
+      "replace-minimum-constant": 6.000000212225132e-06,
+      "reshape-mover": 3.000000106112566e-06,
+      "simplify-concat": 4.8999998398358e-05,
+      "simplify-while-loops": 1.9999999949504854e-06,
+      "transform-variadic-reduce": 9.000000318337698e-06,
+      "tuple-simplifier": 4.999999873689376e-06,
+      "unpack-nested-aws-ntwsr": 3.999999989900971e-06,
+      "unroll-while-loop": 0.0
+    },
+    "hilo": {
+      "ArithmeticIntensity": 834.6854858398438,
+      "HloMacCount": 103079215104.0,
+      "Traffic": 246989344.0
+    }
+  },
+  "sg02": {
+    "compiletime": {
+      "CanonicalizeConv": 0.0,
+      "CanonicalizeForTensorizer": 1.4000000192027073e-05,
+      "Canonicalizer": 0.0003129999968223274,
+      "HoistCompute": 9.999999974752427e-07,
+      "IdentifyCrossPassTensors": 1.2000000424450263e-05,
+      "MemcastMotion": 0.0,
+      "PenguinizeFunctions": 9.999999747378752e-06,
+      "PruneFunctions": 7.999999979801942e-06,
+      "RemoveOptimizationBarriers": 9.000000318337698e-06,
+      "ScatterMotion": 3.000000106112566e-06,
+      "TensorizerLegalizationPass": 6.000000212225132e-06,
+      "VerifySupportedOps": 1.2000000424450263e-05,
+      "algsimp": 5.999999848427251e-05,
+      "batchnorm_expander": 1.2999999853491317e-05,
+      "boundary-marker-removal": 3.999999989900971e-06,
+      "call-inliner": 9.999999747378752e-06,
+      "canonicalize-boundary-marker": 4.999999873689376e-06,
+      "collective-stream-id-checker": 3.999999989900971e-06,
+      "comparison-expander": 4.999999873689376e-06,
+      "computation-deduplicator": 2.5999999706982635e-05,
+      "conditional-to-select": 7.000000096013537e-06,
+      "config-lowering": 2.5999999706982635e-05,
+      "constant_folding": 9.000000318337698e-06,
+      "cse": 1.2000000424450263e-05,
+      "dce": 9.999999974752427e-07,
+      "dynamic-slice-transpose": 3.999999989900971e-06,
+      "eliminate-redundant-compare": 3.000000106112566e-06,
+      "emit-offloaded-dropout": 1.2999999853491317e-05,
+      "flatten-call-graph": 1.2000000424450263e-05,
+      "fuse-send-recv": 1.9999999494757503e-05,
+      "hilo::LegalizeAlias": 1.9999999949504854e-06,
+      "hilo::NeuronInstCombine": 6.800000119255856e-05,
+      "hilo::NeuronOpFusion": 3.999999989900971e-06,
+      "hilo::ReplaceTokenTypeWithU8Pass": 2.2000000171829015e-05,
+      "hilo::ScheduleFusion": 0.00015900000289548188,
+      "hilo::SixtyFourHack": 3.9999998989515007e-05,
+      "hilo::VerifyAliasing": 9.999999974752427e-07,
+      "hlo-mac-count": 0.00017699999443721026,
+      "hlo-verifier": 0.0001829999964684248,
+      "legalize-ccops": 1.9999999949504854e-06,
+      "legalize-compare": 3.000000106112566e-06,
+      "lower-argminmax-custom-call": 3.000000106112566e-06,
+      "map-inline": 1.2000000424450263e-05,
+      "metadata-naming": 1.5999999959603883e-05,
+      "mlir::detail::OpToOpPassAdaptor": 1.1000000085914508e-05,
+      "mlir::hlo::MhloToPyPenguin": 0.0008900000248104334,
+      "mlir::mhlo::LowerComplexExtraPass": 8.800000068731606e-05,
+      "mlir::mhlo::LowerComplexPass": 0.00011999999696854502,
+      "native-to-custom-softmax": 6.000000212225132e-06,
+      "native-to-custom-softmax-dx": 2.4000000848900527e-05,
+      "operand_upcaster": 1.5999999959603883e-05,
+      "post-par-pipe-begin": 1.9999999949504854e-06,
+      "post-par-pipe-end": 0.0,
+      "post-partition-simplification": 0.0005510000046342611,
+      "replace-minimum-constant": 9.000000318337698e-06,
+      "reshape-mover": 3.000000106112566e-06,
+      "simplify-concat": 4.400000034365803e-05,
+      "simplify-while-loops": 1.9999999949504854e-06,
+      "transform-variadic-reduce": 6.299999949987978e-05,
+      "tuple-simplifier": 4.999999873689376e-06,
+      "unpack-nested-aws-ntwsr": 3.999999989900971e-06,
+      "unroll-while-loop": 0.0
+    },
+    "hilo": {
+      "ArithmeticIntensity": 194.41075134277344,
+      "HloMacCount": 77620576256.0,
+      "Traffic": 798521408.0
+    }
+  }
+}

context_encoding_model/_tp0_bk3/graph.neff

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:3602ab29177b01531c0dbdb62bc869556ef53a934ba98dd3bd846e75e171cc3a
+size 2561024

context_encoding_model/_tp0_bk3/metaneff.pb

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:3ea141404110996ab61ca5ba70e86499e6c4390e0b31c1ef947cf95911917766
+size 1816103

context_encoding_model/_tp0_bk3/model.MODULE_b3ddbc97e5f0d1d64c82+155de413.hlo_module.pb

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:9068f3ba4f55e1b8b35adde74efc6a9e617baa344783aaee62353f9181c3092c
+size 1893189

context_encoding_model/_tp0_bk3/model.MODULE_b3ddbc97e5f0d1d64c82+155de413.neff

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:3602ab29177b01531c0dbdb62bc869556ef53a934ba98dd3bd846e75e171cc3a
+size 2561024

context_encoding_model/_tp0_bk3/neuron_config.json

@@ -0,0 +1,220 @@
+{
+  "_attn_implementation_autoset": false,
+  "_name_or_path": "Qwen/Qwen3-8B",
+  "add_cross_attention": false,
+  "architectures": [
+    "Qwen3ForCausalLM"
+  ],
+  "attention_bias": false,
+  "attention_dropout": 0.0,
+  "attribute_map": {},
+  "bad_words_ids": null,
+  "begin_suppress_tokens": null,
+  "bos_token_id": 151643,
+  "chunk_size_feed_forward": 0,
+  "cross_attention_hidden_size": null,
+  "decoder_start_token_id": null,
+  "diversity_penalty": 0.0,
+  "do_sample": false,
+  "early_stopping": false,
+  "encoder_no_repeat_ngram_size": 0,
+  "eos_token_id": 151645,
+  "exponential_decay_length_penalty": null,
+  "finetuning_task": null,
+  "forced_bos_token_id": null,
+  "forced_eos_token_id": null,
+  "fused_spec_config": null,
+  "head_dim": 128,
+  "hidden_act": "silu",
+  "hidden_size": 4096,
+  "id2label": {
+    "0": "LABEL_0",
+    "1": "LABEL_1"
+  },
+  "initializer_range": 0.02,
+  "intermediate_size": 12288,
+  "is_decoder": false,
+  "is_encoder_decoder": false,
+  "label2id": {
+    "LABEL_0": 0,
+    "LABEL_1": 1
+  },
+  "length_penalty": 1.0,
+  "max_length": 20,
+  "max_position_embeddings": 40960,
+  "max_window_layers": 36,
+  "metadata": null,
+  "min_length": 0,
+  "model_type": "qwen3",
+  "neuron_config": {
+    "activation_quantization_type": null,
+    "allow_input_truncation": false,
+    "apply_seq_ids_mask": false,
+    "async_mode": false,
+    "attention_dp_degree": 1,
+    "attention_dtype": null,
+    "attn_block_cte_nki_kernel_enabled": false,
+    "attn_block_tkg_nki_kernel_cache_update": false,
+    "attn_block_tkg_nki_kernel_enabled": false,
+    "attn_cls": {
+      "__module__": "neuronx_distributed_inference.models.qwen3.modeling_qwen3",
+      "__name__": "NeuronQwen3Attention"
+    },
+    "attn_kernel_enabled": null,
+    "attn_tkg_builtin_kernel_enabled": false,
+    "attn_tkg_nki_kernel_enabled": false,
+    "batch_size": 1,
+    "bucket_n_active_tokens": true,
+    "buckets": [
+      1024
+    ],
+    "cast_type": "config",
+    "cc_pipeline_tiling_factor": 2,
+    "chunked_prefill_config": null,
+    "context_encoding_buckets": [
+      1024
+    ],
+    "cp_degree": 1,
+    "ctx_batch_size": 1,
+    "disable_kv_cache_tiling": false,
+    "draft_model_modules_to_not_convert": null,
+    "enable_bucketing": true,
+    "enable_eagle_draft_input_norm": false,
+    "enable_eagle_speculation": false,
+    "enable_fused_speculation": false,
+    "enable_long_context_mode": false,
+    "enable_output_completion_notifications": false,
+    "enable_spill_reload_dge": false,
+    "enable_token_tree": false,
+    "ep_degree": 1,
+    "expert_mlp_nki_kernel_enabled": null,
+    "flash_decoding_enabled": false,
+    "fused_qkv": false,
+    "fused_rmsnorm_skip_gamma": false,
+    "is_block_kv_layout": null,
+    "is_chunked_prefill": false,
+    "is_continuous_batching": true,
+    "is_eagle_draft": false,
+    "is_medusa": false,
+    "is_prefill_stage": true,
+    "is_prefix_caching": false,
+    "k_cache_transposed": false,
+    "kv_cache_batch_size": 1,
+    "kv_cache_padding_size": 0,
+    "kv_cache_quant": false,
+    "kv_cache_tiling": false,
+    "layer_boundary_markers": false,
+    "lm_head_pad": false,
+    "lm_head_pad_alignment_size": 1,
+    "local_ranks_size": 2,
+    "logical_nc_config": 1,
+    "lora_config": null,
+    "max_batch_size": 1,
+    "max_context_length": 1024,
+    "max_length": 1024,
+    "max_new_tokens": null,
+    "medusa_speculation_length": 0,
+    "medusa_tree": null,
+    "mlp_kernel_enabled": false,
+    "mlp_kernel_fuse_residual_add": false,
+    "modules_to_not_convert": null,
+    "moe_fused_nki_kernel_enabled": null,
+    "n_active_tokens": 1024,
+    "n_positions": 1024,
+    "num_medusa_heads": 0,
+    "on_cpu": false,
+    "on_device_sampling_config": {
+      "deterministic": false,
+      "do_sample": false,
+      "dynamic": true,
+      "global_topk": 256,
+      "on_device_sampling_config": true,
+      "temperature": 1.0,
+      "top_k": 1,
+      "top_k_kernel_enabled": false,
+      "top_p": 1.0
+    },
+    "output_logits": false,
+    "overrides_torch_dtype": true,
+    "pa_block_size": 1024,
+    "pa_num_blocks": 1,
+    "padding_side": "right",
+    "pp_degree": 1,
+    "prefix_buckets": null,
+    "qk_layernorm": false,
+    "qkv_kernel_enabled": false,
+    "qkv_kernel_fuse_residual_add": false,
+    "qkv_kernel_nbsd_layout": false,
+    "quantization_dtype": "int8",
+    "quantization_type": "per_tensor_symmetric",
+    "quantize_clamp_bound": Infinity,
+    "quantized": false,
+    "quantized_checkpoints_path": null,
+    "quantized_mlp_kernel_enabled": false,
+    "rmsnorm_quantize_kernel_enabled": false,
+    "router_topk_nki_kernel_enabled": null,
+    "rpl_reduce_dtype": null,
+    "save_sharded_checkpoint": true,
+    "scratchpad_page_size": null,
+    "seq_len": 1024,
+    "seq_len_threshold_for_cc_tiling": 16384,
+    "sequence_parallel_enabled": false,
+    "shared_mlp_nki_kernel_enabled": null,
+    "skip_sharding": false,
+    "skip_warmup": false,
+    "spec_batch_size": 1,
+    "speculation_length": 0,
+    "start_rank_id": 0,
+    "target": null,
+    "tile_cc": false,
+    "tkg_batch_size": 1,
+    "token_generation_buckets": null,
+    "token_tree_config": null,
+    "torch_dtype": "bfloat16",
+    "tp_degree": 2,
+    "vocab_parallel": false,
+    "weight_gather_seq_len_threshold": 32768,
+    "weights_to_skip_layout_optimization": [],
+    "world_size": 2
+  },
+  "no_repeat_ngram_size": 0,
+  "num_attention_heads": 32,
+  "num_beam_groups": 1,
+  "num_beams": 1,
+  "num_cores_per_group": 1,
+  "num_hidden_layers": 36,
+  "num_key_value_heads": 8,
+  "num_return_sequences": 1,
+  "output_attentions": false,
+  "output_hidden_states": false,
+  "output_scores": false,
+  "pad_token_id": 0,
+  "prefix": null,
+  "problem_type": null,
+  "pruned_heads": {},
+  "remove_invalid_values": false,
+  "repetition_penalty": 1.0,
+  "return_dict": true,
+  "return_dict_in_generate": false,
+  "rms_norm_eps": 1e-06,
+  "rope_scaling": null,
+  "rope_theta": 1000000,
+  "sep_token_id": null,
+  "sliding_window": null,
+  "suppress_tokens": null,
+  "task_specific_params": null,
+  "temperature": 1.0,
+  "tf_legacy_loss": false,
+  "tie_encoder_decoder": false,
+  "tie_word_embeddings": false,
+  "tokenizer_class": null,
+  "top_k": 50,
+  "top_p": 1.0,
+  "torchscript": false,
+  "transformers_version": "4.51.0",
+  "typical_p": 1.0,
+  "use_bfloat16": false,
+  "use_cache": true,
+  "use_sliding_window": false,
+  "vocab_size": 151936
+}

layout_opt/command.txt

@@ -0,0 +1 @@
+neuronx-cc compile graph.hlo --framework XLA --target trn1 --output graph.neff --model-type=transformer -O1 --lnc=1 '--internal-hlo2tensorizer-options=--experimental-unsafe-fp8e4m3fn-as-fp8e4m3 --verify-hlo=false' --logfile=log-neuron-cc.txt --verbose=35

layout_opt/graph.neff

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:eafae43287bda2aa58740df223d211d8e3638af29e402c9cc6cbcadcf302ddde
+size 5786624

layout_opt/metaneff

@@ -0,0 +1,1198 @@
+
+(
+input0��	�2embed_tokens.weight8
+;
+input1� �2'layers.0.self_attn.o_proj.o_proj.weight8
+=
+input2�� 2)layers.0.self_attn.qkv_proj.v_proj.weight8
+1
+input3� 2layers.0.input_layernorm.weight8
+7
+input4�
2%layers.0.self_attn.k_layernorm.weight8
+=
+input5�� 2)layers.0.self_attn.qkv_proj.k_proj.weight8
+7
+input6�
2%layers.0.self_attn.q_layernorm.weight8
+=
+input7�� 2)layers.0.self_attn.qkv_proj.q_proj.weight8
+1
+input8� �02layers.0.mlp.down_proj.weight8
+/
+input9�0� 2layers.0.mlp.up_proj.weight8
+;
+input10� 2(layers.0.post_attention_layernorm.weight8
+2
+input11�0� 2layers.0.mlp.gate_proj.weight8
+<
+input12� �2'layers.1.self_attn.o_proj.o_proj.weight8
+>
+input13�� 2)layers.1.self_attn.qkv_proj.v_proj.weight8
+2
+input14� 2layers.1.input_layernorm.weight8
+8
+input15�
2%layers.1.self_attn.k_layernorm.weight8
+>
+input16�� 2)layers.1.self_attn.qkv_proj.k_proj.weight8
+8
+input17�
2%layers.1.self_attn.q_layernorm.weight8
+>
+input18�� 2)layers.1.self_attn.qkv_proj.q_proj.weight8
+2
+input19� �02layers.1.mlp.down_proj.weight8
+0
+input20�0� 2layers.1.mlp.up_proj.weight8
+;
+input21� 2(layers.1.post_attention_layernorm.weight8
+2
+input22�0� 2layers.1.mlp.gate_proj.weight8
+<
+input23� �2'layers.2.self_attn.o_proj.o_proj.weight8
+>
+input24�� 2)layers.2.self_attn.qkv_proj.v_proj.weight8
+2
+input25� 2layers.2.input_layernorm.weight8
+8
+input26�
2%layers.2.self_attn.k_layernorm.weight8
+>
+input27�� 2)layers.2.self_attn.qkv_proj.k_proj.weight8
+8
+input28�
2%layers.2.self_attn.q_layernorm.weight8
+>
+input29�� 2)layers.2.self_attn.qkv_proj.q_proj.weight8
+2
+input30� �02layers.2.mlp.down_proj.weight8
+0
+input31�0� 2layers.2.mlp.up_proj.weight8
+;
+input32� 2(layers.2.post_attention_layernorm.weight8
+2
+input33�0� 2layers.2.mlp.gate_proj.weight8
+<
+input34� �2'layers.3.self_attn.o_proj.o_proj.weight8
+>
+input35�� 2)layers.3.self_attn.qkv_proj.v_proj.weight8
+2
+input36� 2layers.3.input_layernorm.weight8
+8
+input37�
2%layers.3.self_attn.k_layernorm.weight8
+>
+input38�� 2)layers.3.self_attn.qkv_proj.k_proj.weight8
+8
+input39�
2%layers.3.self_attn.q_layernorm.weight8
+>
+input40�� 2)layers.3.self_attn.qkv_proj.q_proj.weight8
+2
+input41� �02layers.3.mlp.down_proj.weight8
+0
+input42�0� 2layers.3.mlp.up_proj.weight8
+;
+input43� 2(layers.3.post_attention_layernorm.weight8
+2
+input44�0� 2layers.3.mlp.gate_proj.weight8
+<
+input45� �2'layers.4.self_attn.o_proj.o_proj.weight8
+>
+input46�� 2)layers.4.self_attn.qkv_proj.v_proj.weight8
+2
+input47� 2layers.4.input_layernorm.weight8
+8
+input48�
2%layers.4.self_attn.k_layernorm.weight8
+>
+input49�� 2)layers.4.self_attn.qkv_proj.k_proj.weight8
+8
+input50�
2%layers.4.self_attn.q_layernorm.weight8
+>
+input51�� 2)layers.4.self_attn.qkv_proj.q_proj.weight8
+2
+input52� �02layers.4.mlp.down_proj.weight8
+0
+input53�0� 2layers.4.mlp.up_proj.weight8
+;
+input54� 2(layers.4.post_attention_layernorm.weight8
+2
+input55�0� 2layers.4.mlp.gate_proj.weight8
+<
+input56� �2'layers.5.self_attn.o_proj.o_proj.weight8
+>
+input57�� 2)layers.5.self_attn.qkv_proj.v_proj.weight8
+2
+input58� 2layers.5.input_layernorm.weight8
+8
+input59�
2%layers.5.self_attn.k_layernorm.weight8
+>
+input60�� 2)layers.5.self_attn.qkv_proj.k_proj.weight8
+8
+input61�
2%layers.5.self_attn.q_layernorm.weight8
+>
+input62�� 2)layers.5.self_attn.qkv_proj.q_proj.weight8
+2
+input63� �02layers.5.mlp.down_proj.weight8
+0
+input64�0� 2layers.5.mlp.up_proj.weight8
+;
+input65� 2(layers.5.post_attention_layernorm.weight8
+2
+input66�0� 2layers.5.mlp.gate_proj.weight8
+<
+input67� �2'layers.6.self_attn.o_proj.o_proj.weight8
+>
+input68�� 2)layers.6.self_attn.qkv_proj.v_proj.weight8
+2
+input69� 2layers.6.input_layernorm.weight8
+8
+input70�
2%layers.6.self_attn.k_layernorm.weight8
+>
+input71�� 2)layers.6.self_attn.qkv_proj.k_proj.weight8
+8
+input72�
2%layers.6.self_attn.q_layernorm.weight8
+>
+input73�� 2)layers.6.self_attn.qkv_proj.q_proj.weight8
+2
+input74� �02layers.6.mlp.down_proj.weight8
+0
+input75�0� 2layers.6.mlp.up_proj.weight8
+;
+input76� 2(layers.6.post_attention_layernorm.weight8
+2
+input77�0� 2layers.6.mlp.gate_proj.weight8
+<
+input78� �2'layers.7.self_attn.o_proj.o_proj.weight8
+>
+input79�� 2)layers.7.self_attn.qkv_proj.v_proj.weight8
+2
+input80� 2layers.7.input_layernorm.weight8
+8
+input81�
2%layers.7.self_attn.k_layernorm.weight8
+>
+input82�� 2)layers.7.self_attn.qkv_proj.k_proj.weight8
+8
+input83�
2%layers.7.self_attn.q_layernorm.weight8
+>
+input84�� 2)layers.7.self_attn.qkv_proj.q_proj.weight8
+2
+input85� �02layers.7.mlp.down_proj.weight8
+0
+input86�0� 2layers.7.mlp.up_proj.weight8
+;
+input87� 2(layers.7.post_attention_layernorm.weight8
+2
+input88�0� 2layers.7.mlp.gate_proj.weight8
+<
+input89� �2'layers.8.self_attn.o_proj.o_proj.weight8
+>
+input90�� 2)layers.8.self_attn.qkv_proj.v_proj.weight8
+2
+input91� 2layers.8.input_layernorm.weight8
+8
+input92�
2%layers.8.self_attn.k_layernorm.weight8
+>
+input93�� 2)layers.8.self_attn.qkv_proj.k_proj.weight8
+8
+input94�
2%layers.8.self_attn.q_layernorm.weight8
+>
+input95�� 2)layers.8.self_attn.qkv_proj.q_proj.weight8
+2
+input96� �02layers.8.mlp.down_proj.weight8
+0
+input97�0� 2layers.8.mlp.up_proj.weight8
+;
+input98� 2(layers.8.post_attention_layernorm.weight8
+2
+input99�0� 2layers.8.mlp.gate_proj.weight8
+=
+input100� �2'layers.9.self_attn.o_proj.o_proj.weight8
+?
+input101�� 2)layers.9.self_attn.qkv_proj.v_proj.weight8
+3
+input102� 2layers.9.input_layernorm.weight8
+9
+input103�
2%layers.9.self_attn.k_layernorm.weight8
+?
+input104�� 2)layers.9.self_attn.qkv_proj.k_proj.weight8
+9
+input105�
2%layers.9.self_attn.q_layernorm.weight8
+?
+input106�� 2)layers.9.self_attn.qkv_proj.q_proj.weight8
+3
+input107� �02layers.9.mlp.down_proj.weight8
+1
+input108�0� 2layers.9.mlp.up_proj.weight8
+<
+input109� 2(layers.9.post_attention_layernorm.weight8
+3
+input110�0� 2layers.9.mlp.gate_proj.weight8
+>
+input111� �2(layers.10.self_attn.o_proj.o_proj.weight8
+@
+input112�� 2*layers.10.self_attn.qkv_proj.v_proj.weight8
+4
+input113� 2 layers.10.input_layernorm.weight8
+:
+input114�
2&layers.10.self_attn.k_layernorm.weight8
+@
+input115�� 2*layers.10.self_attn.qkv_proj.k_proj.weight8
+:
+input116�
2&layers.10.self_attn.q_layernorm.weight8
+@
+input117�� 2*layers.10.self_attn.qkv_proj.q_proj.weight8
+4
+input118� �02layers.10.mlp.down_proj.weight8
+2
+input119�0� 2layers.10.mlp.up_proj.weight8
+=
+input120� 2)layers.10.post_attention_layernorm.weight8
+4
+input121�0� 2layers.10.mlp.gate_proj.weight8
+>
+input122� �2(layers.11.self_attn.o_proj.o_proj.weight8
+@
+input123�� 2*layers.11.self_attn.qkv_proj.v_proj.weight8
+4
+input124� 2 layers.11.input_layernorm.weight8
+:
+input125�
2&layers.11.self_attn.k_layernorm.weight8
+@
+input126�� 2*layers.11.self_attn.qkv_proj.k_proj.weight8
+:
+input127�
2&layers.11.self_attn.q_layernorm.weight8
+@
+input128�� 2*layers.11.self_attn.qkv_proj.q_proj.weight8
+4
+input129� �02layers.11.mlp.down_proj.weight8
+2
+input130�0� 2layers.11.mlp.up_proj.weight8
+=
+input131� 2)layers.11.post_attention_layernorm.weight8
+4
+input132�0� 2layers.11.mlp.gate_proj.weight8
+>
+input133� �2(layers.12.self_attn.o_proj.o_proj.weight8
+@
+input134�� 2*layers.12.self_attn.qkv_proj.v_proj.weight8
+4
+input135� 2 layers.12.input_layernorm.weight8
+:
+input136�
2&layers.12.self_attn.k_layernorm.weight8
+@
+input137�� 2*layers.12.self_attn.qkv_proj.k_proj.weight8
+:
+input138�
2&layers.12.self_attn.q_layernorm.weight8
+@
+input139�� 2*layers.12.self_attn.qkv_proj.q_proj.weight8
+4
+input140� �02layers.12.mlp.down_proj.weight8
+2
+input141�0� 2layers.12.mlp.up_proj.weight8
+=
+input142� 2)layers.12.post_attention_layernorm.weight8
+4
+input143�0� 2layers.12.mlp.gate_proj.weight8
+>
+input144� �2(layers.13.self_attn.o_proj.o_proj.weight8
+@
+input145�� 2*layers.13.self_attn.qkv_proj.v_proj.weight8
+4
+input146� 2 layers.13.input_layernorm.weight8
+:
+input147�
2&layers.13.self_attn.k_layernorm.weight8
+@
+input148�� 2*layers.13.self_attn.qkv_proj.k_proj.weight8
+:
+input149�
2&layers.13.self_attn.q_layernorm.weight8
+@
+input150�� 2*layers.13.self_attn.qkv_proj.q_proj.weight8
+4
+input151� �02layers.13.mlp.down_proj.weight8
+2
+input152�0� 2layers.13.mlp.up_proj.weight8
+=
+input153� 2)layers.13.post_attention_layernorm.weight8
+4
+input154�0� 2layers.13.mlp.gate_proj.weight8
+>
+input155� �2(layers.14.self_attn.o_proj.o_proj.weight8
+@
+input156�� 2*layers.14.self_attn.qkv_proj.v_proj.weight8
+4
+input157� 2 layers.14.input_layernorm.weight8
+:
+input158�
2&layers.14.self_attn.k_layernorm.weight8
+@
+input159�� 2*layers.14.self_attn.qkv_proj.k_proj.weight8
+:
+input160�
2&layers.14.self_attn.q_layernorm.weight8
+@
+input161�� 2*layers.14.self_attn.qkv_proj.q_proj.weight8
+4
+input162� �02layers.14.mlp.down_proj.weight8
+2
+input163�0� 2layers.14.mlp.up_proj.weight8
+=
+input164� 2)layers.14.post_attention_layernorm.weight8
+4
+input165�0� 2layers.14.mlp.gate_proj.weight8
+>
+input166� �2(layers.15.self_attn.o_proj.o_proj.weight8
+@
+input167�� 2*layers.15.self_attn.qkv_proj.v_proj.weight8
+4
+input168� 2 layers.15.input_layernorm.weight8
+:
+input169�
2&layers.15.self_attn.k_layernorm.weight8
+@
+input170�� 2*layers.15.self_attn.qkv_proj.k_proj.weight8
+:
+input171�
2&layers.15.self_attn.q_layernorm.weight8
+@
+input172�� 2*layers.15.self_attn.qkv_proj.q_proj.weight8
+4
+input173� �02layers.15.mlp.down_proj.weight8
+2
+input174�0� 2layers.15.mlp.up_proj.weight8
+=
+input175� 2)layers.15.post_attention_layernorm.weight8
+4
+input176�0� 2layers.15.mlp.gate_proj.weight8
+>
+input177� �2(layers.16.self_attn.o_proj.o_proj.weight8
+@
+input178�� 2*layers.16.self_attn.qkv_proj.v_proj.weight8
+4
+input179� 2 layers.16.input_layernorm.weight8
+:
+input180�
2&layers.16.self_attn.k_layernorm.weight8
+@
+input181�� 2*layers.16.self_attn.qkv_proj.k_proj.weight8
+:
+input182�
2&layers.16.self_attn.q_layernorm.weight8
+@
+input183�� 2*layers.16.self_attn.qkv_proj.q_proj.weight8
+4
+input184� �02layers.16.mlp.down_proj.weight8
+2
+input185�0� 2layers.16.mlp.up_proj.weight8
+=
+input186� 2)layers.16.post_attention_layernorm.weight8
+4
+input187�0� 2layers.16.mlp.gate_proj.weight8
+>
+input188� �2(layers.17.self_attn.o_proj.o_proj.weight8
+@
+input189�� 2*layers.17.self_attn.qkv_proj.v_proj.weight8
+4
+input190� 2 layers.17.input_layernorm.weight8
+:
+input191�
2&layers.17.self_attn.k_layernorm.weight8
+@
+input192�� 2*layers.17.self_attn.qkv_proj.k_proj.weight8
+:
+input193�
2&layers.17.self_attn.q_layernorm.weight8
+@
+input194�� 2*layers.17.self_attn.qkv_proj.q_proj.weight8
+4
+input195� �02layers.17.mlp.down_proj.weight8
+2
+input196�0� 2layers.17.mlp.up_proj.weight8
+=
+input197� 2)layers.17.post_attention_layernorm.weight8
+4
+input198�0� 2layers.17.mlp.gate_proj.weight8
+>
+input199� �2(layers.18.self_attn.o_proj.o_proj.weight8
+@
+input200�� 2*layers.18.self_attn.qkv_proj.v_proj.weight8
+4
+input201� 2 layers.18.input_layernorm.weight8
+:
+input202�
2&layers.18.self_attn.k_layernorm.weight8
+@
+input203�� 2*layers.18.self_attn.qkv_proj.k_proj.weight8
+:
+input204�
2&layers.18.self_attn.q_layernorm.weight8
+@
+input205�� 2*layers.18.self_attn.qkv_proj.q_proj.weight8
+4
+input206� �02layers.18.mlp.down_proj.weight8
+2
+input207�0� 2layers.18.mlp.up_proj.weight8
+=
+input208� 2)layers.18.post_attention_layernorm.weight8
+4
+input209�0� 2layers.18.mlp.gate_proj.weight8
+>
+input210� �2(layers.19.self_attn.o_proj.o_proj.weight8
+@
+input211�� 2*layers.19.self_attn.qkv_proj.v_proj.weight8
+4
+input212� 2 layers.19.input_layernorm.weight8
+:
+input213�
2&layers.19.self_attn.k_layernorm.weight8
+@
+input214�� 2*layers.19.self_attn.qkv_proj.k_proj.weight8
+:
+input215�
2&layers.19.self_attn.q_layernorm.weight8
+@
+input216�� 2*layers.19.self_attn.qkv_proj.q_proj.weight8
+4
+input217� �02layers.19.mlp.down_proj.weight8
+2
+input218�0� 2layers.19.mlp.up_proj.weight8
+=
+input219� 2)layers.19.post_attention_layernorm.weight8
+4
+input220�0� 2layers.19.mlp.gate_proj.weight8
+>
+input221� �2(layers.20.self_attn.o_proj.o_proj.weight8
+@
+input222�� 2*layers.20.self_attn.qkv_proj.v_proj.weight8
+4
+input223� 2 layers.20.input_layernorm.weight8
+:
+input224�
2&layers.20.self_attn.k_layernorm.weight8
+@
+input225�� 2*layers.20.self_attn.qkv_proj.k_proj.weight8
+:
+input226�
2&layers.20.self_attn.q_layernorm.weight8
+@
+input227�� 2*layers.20.self_attn.qkv_proj.q_proj.weight8
+4
+input228� �02layers.20.mlp.down_proj.weight8
+2
+input229�0� 2layers.20.mlp.up_proj.weight8
+=
+input230� 2)layers.20.post_attention_layernorm.weight8
+4
+input231�0� 2layers.20.mlp.gate_proj.weight8
+>
+input232� �2(layers.21.self_attn.o_proj.o_proj.weight8
+@
+input233�� 2*layers.21.self_attn.qkv_proj.v_proj.weight8
+4
+input234� 2 layers.21.input_layernorm.weight8
+:
+input235�
2&layers.21.self_attn.k_layernorm.weight8
+@
+input236�� 2*layers.21.self_attn.qkv_proj.k_proj.weight8
+:
+input237�
2&layers.21.self_attn.q_layernorm.weight8
+@
+input238�� 2*layers.21.self_attn.qkv_proj.q_proj.weight8
+4
+input239� �02layers.21.mlp.down_proj.weight8
+2
+input240�0� 2layers.21.mlp.up_proj.weight8
+=
+input241� 2)layers.21.post_attention_layernorm.weight8
+4
+input242�0� 2layers.21.mlp.gate_proj.weight8
+>
+input243� �2(layers.22.self_attn.o_proj.o_proj.weight8
+@
+input244�� 2*layers.22.self_attn.qkv_proj.v_proj.weight8
+4
+input245� 2 layers.22.input_layernorm.weight8
+:
+input246�
2&layers.22.self_attn.k_layernorm.weight8
+@
+input247�� 2*layers.22.self_attn.qkv_proj.k_proj.weight8
+:
+input248�
2&layers.22.self_attn.q_layernorm.weight8
+@
+input249�� 2*layers.22.self_attn.qkv_proj.q_proj.weight8
+4
+input250� �02layers.22.mlp.down_proj.weight8
+2
+input251�0� 2layers.22.mlp.up_proj.weight8
+=
+input252� 2)layers.22.post_attention_layernorm.weight8
+4
+input253�0� 2layers.22.mlp.gate_proj.weight8
+>
+input254� �2(layers.23.self_attn.o_proj.o_proj.weight8
+@
+input255�� 2*layers.23.self_attn.qkv_proj.v_proj.weight8
+4
+input256� 2 layers.23.input_layernorm.weight8
+:
+input257�
2&layers.23.self_attn.k_layernorm.weight8
+@
+input258�� 2*layers.23.self_attn.qkv_proj.k_proj.weight8
+:
+input259�
2&layers.23.self_attn.q_layernorm.weight8
+@
+input260�� 2*layers.23.self_attn.qkv_proj.q_proj.weight8
+4
+input261� �02layers.23.mlp.down_proj.weight8
+2
+input262�0� 2layers.23.mlp.up_proj.weight8
+=
+input263� 2)layers.23.post_attention_layernorm.weight8
+4
+input264�0� 2layers.23.mlp.gate_proj.weight8
+>
+input265� �2(layers.24.self_attn.o_proj.o_proj.weight8
+@
+input266�� 2*layers.24.self_attn.qkv_proj.v_proj.weight8
+4
+input267� 2 layers.24.input_layernorm.weight8
+:
+input268�
2&layers.24.self_attn.k_layernorm.weight8
+@
+input269�� 2*layers.24.self_attn.qkv_proj.k_proj.weight8
+:
+input270�
2&layers.24.self_attn.q_layernorm.weight8
+@
+input271�� 2*layers.24.self_attn.qkv_proj.q_proj.weight8
+4
+input272� �02layers.24.mlp.down_proj.weight8
+2
+input273�0� 2layers.24.mlp.up_proj.weight8
+=
+input274� 2)layers.24.post_attention_layernorm.weight8
+4
+input275�0� 2layers.24.mlp.gate_proj.weight8
+>
+input276� �2(layers.25.self_attn.o_proj.o_proj.weight8
+@
+input277�� 2*layers.25.self_attn.qkv_proj.v_proj.weight8
+4
+input278� 2 layers.25.input_layernorm.weight8
+:
+input279�
2&layers.25.self_attn.k_layernorm.weight8
+@
+input280�� 2*layers.25.self_attn.qkv_proj.k_proj.weight8
+:
+input281�
2&layers.25.self_attn.q_layernorm.weight8
+@
+input282�� 2*layers.25.self_attn.qkv_proj.q_proj.weight8
+4
+input283� �02layers.25.mlp.down_proj.weight8
+2
+input284�0� 2layers.25.mlp.up_proj.weight8
+=
+input285� 2)layers.25.post_attention_layernorm.weight8
+4
+input286�0� 2layers.25.mlp.gate_proj.weight8
+>
+input287� �2(layers.26.self_attn.o_proj.o_proj.weight8
+@
+input288�� 2*layers.26.self_attn.qkv_proj.v_proj.weight8
+4
+input289� 2 layers.26.input_layernorm.weight8
+:
+input290�
2&layers.26.self_attn.k_layernorm.weight8
+@
+input291�� 2*layers.26.self_attn.qkv_proj.k_proj.weight8
+:
+input292�
2&layers.26.self_attn.q_layernorm.weight8
+@
+input293�� 2*layers.26.self_attn.qkv_proj.q_proj.weight8
+4
+input294� �02layers.26.mlp.down_proj.weight8
+2
+input295�0� 2layers.26.mlp.up_proj.weight8
+=
+input296� 2)layers.26.post_attention_layernorm.weight8
+4
+input297�0� 2layers.26.mlp.gate_proj.weight8
+>
+input298� �2(layers.27.self_attn.o_proj.o_proj.weight8
+@
+input299�� 2*layers.27.self_attn.qkv_proj.v_proj.weight8
+4
+input300� 2 layers.27.input_layernorm.weight8
+:
+input301�
2&layers.27.self_attn.k_layernorm.weight8
+@
+input302�� 2*layers.27.self_attn.qkv_proj.k_proj.weight8
+:
+input303�
2&layers.27.self_attn.q_layernorm.weight8
+@
+input304�� 2*layers.27.self_attn.qkv_proj.q_proj.weight8
+4
+input305� �02layers.27.mlp.down_proj.weight8
+2
+input306�0� 2layers.27.mlp.up_proj.weight8
+=
+input307� 2)layers.27.post_attention_layernorm.weight8
+4
+input308�0� 2layers.27.mlp.gate_proj.weight8
+>
+input309� �2(layers.28.self_attn.o_proj.o_proj.weight8
+@
+input310�� 2*layers.28.self_attn.qkv_proj.v_proj.weight8
+4
+input311� 2 layers.28.input_layernorm.weight8
+:
+input312�
2&layers.28.self_attn.k_layernorm.weight8
+@
+input313�� 2*layers.28.self_attn.qkv_proj.k_proj.weight8
+:
+input314�
2&layers.28.self_attn.q_layernorm.weight8
+@
+input315�� 2*layers.28.self_attn.qkv_proj.q_proj.weight8
+4
+input316� �02layers.28.mlp.down_proj.weight8
+2
+input317�0� 2layers.28.mlp.up_proj.weight8
+=
+input318� 2)layers.28.post_attention_layernorm.weight8
+4
+input319�0� 2layers.28.mlp.gate_proj.weight8
+>
+input320� �2(layers.29.self_attn.o_proj.o_proj.weight8
+@
+input321�� 2*layers.29.self_attn.qkv_proj.v_proj.weight8
+4
+input322� 2 layers.29.input_layernorm.weight8
+:
+input323�
2&layers.29.self_attn.k_layernorm.weight8
+@
+input324�� 2*layers.29.self_attn.qkv_proj.k_proj.weight8
+:
+input325�
2&layers.29.self_attn.q_layernorm.weight8
+@
+input326�� 2*layers.29.self_attn.qkv_proj.q_proj.weight8
+4
+input327� �02layers.29.mlp.down_proj.weight8
+2
+input328�0� 2layers.29.mlp.up_proj.weight8
+=
+input329� 2)layers.29.post_attention_layernorm.weight8
+4
+input330�0� 2layers.29.mlp.gate_proj.weight8
+>
+input331� �2(layers.30.self_attn.o_proj.o_proj.weight8
+@
+input332�� 2*layers.30.self_attn.qkv_proj.v_proj.weight8
+4
+input333� 2 layers.30.input_layernorm.weight8
+:
+input334�
2&layers.30.self_attn.k_layernorm.weight8
+@
+input335�� 2*layers.30.self_attn.qkv_proj.k_proj.weight8
+:
+input336�
2&layers.30.self_attn.q_layernorm.weight8
+@
+input337�� 2*layers.30.self_attn.qkv_proj.q_proj.weight8
+4
+input338� �02layers.30.mlp.down_proj.weight8
+2
+input339�0� 2layers.30.mlp.up_proj.weight8
+=
+input340� 2)layers.30.post_attention_layernorm.weight8
+4
+input341�0� 2layers.30.mlp.gate_proj.weight8
+>
+input342� �2(layers.31.self_attn.o_proj.o_proj.weight8
+@
+input343�� 2*layers.31.self_attn.qkv_proj.v_proj.weight8
+4
+input344� 2 layers.31.input_layernorm.weight8
+:
+input345�
2&layers.31.self_attn.k_layernorm.weight8
+@
+input346�� 2*layers.31.self_attn.qkv_proj.k_proj.weight8
+:
+input347�
2&layers.31.self_attn.q_layernorm.weight8
+@
+input348�� 2*layers.31.self_attn.qkv_proj.q_proj.weight8
+4
+input349� �02layers.31.mlp.down_proj.weight8
+2
+input350�0� 2layers.31.mlp.up_proj.weight8
+=
+input351� 2)layers.31.post_attention_layernorm.weight8
+4
+input352�0� 2layers.31.mlp.gate_proj.weight8
+>
+input353� �2(layers.32.self_attn.o_proj.o_proj.weight8
+@
+input354�� 2*layers.32.self_attn.qkv_proj.v_proj.weight8
+4
+input355� 2 layers.32.input_layernorm.weight8
+:
+input356�
2&layers.32.self_attn.k_layernorm.weight8
+@
+input357�� 2*layers.32.self_attn.qkv_proj.k_proj.weight8
+:
+input358�
2&layers.32.self_attn.q_layernorm.weight8
+@
+input359�� 2*layers.32.self_attn.qkv_proj.q_proj.weight8
+4
+input360� �02layers.32.mlp.down_proj.weight8
+2
+input361�0� 2layers.32.mlp.up_proj.weight8
+=
+input362� 2)layers.32.post_attention_layernorm.weight8
+4
+input363�0� 2layers.32.mlp.gate_proj.weight8
+>
+input364� �2(layers.33.self_attn.o_proj.o_proj.weight8
+@
+input365�� 2*layers.33.self_attn.qkv_proj.v_proj.weight8
+4
+input366� 2 layers.33.input_layernorm.weight8
+:
+input367�
2&layers.33.self_attn.k_layernorm.weight8
+@
+input368�� 2*layers.33.self_attn.qkv_proj.k_proj.weight8
+:
+input369�
2&layers.33.self_attn.q_layernorm.weight8
+@
+input370�� 2*layers.33.self_attn.qkv_proj.q_proj.weight8
+4
+input371� �02layers.33.mlp.down_proj.weight8
+2
+input372�0� 2layers.33.mlp.up_proj.weight8
+=
+input373� 2)layers.33.post_attention_layernorm.weight8
+4
+input374�0� 2layers.33.mlp.gate_proj.weight8
+>
+input375� �2(layers.34.self_attn.o_proj.o_proj.weight8
+@
+input376�� 2*layers.34.self_attn.qkv_proj.v_proj.weight8
+4
+input377� 2 layers.34.input_layernorm.weight8
+:
+input378�
2&layers.34.self_attn.k_layernorm.weight8
+@
+input379�� 2*layers.34.self_attn.qkv_proj.k_proj.weight8
+:
+input380�
2&layers.34.self_attn.q_layernorm.weight8
+@
+input381�� 2*layers.34.self_attn.qkv_proj.q_proj.weight8
+4
+input382� �02layers.34.mlp.down_proj.weight8
+2
+input383�0� 2layers.34.mlp.up_proj.weight8
+=
+input384� 2)layers.34.post_attention_layernorm.weight8
+4
+input385�0� 2layers.34.mlp.gate_proj.weight8
+>
+input386� �2(layers.35.self_attn.o_proj.o_proj.weight8
+@
+input387�� 2*layers.35.self_attn.qkv_proj.v_proj.weight8
+4
+input388� 2 layers.35.input_layernorm.weight8
+:
+input389�
2&layers.35.self_attn.k_layernorm.weight8
+@
+input390�� 2*layers.35.self_attn.qkv_proj.k_proj.weight8
+:
+input391�
2&layers.35.self_attn.q_layernorm.weight8
+@
+input392�� 2*layers.35.self_attn.qkv_proj.q_proj.weight8
+4
+input393� �02layers.35.mlp.down_proj.weight8
+2
+input394�0� 2layers.35.mlp.up_proj.weight8
+=
+input395� 2)layers.35.post_attention_layernorm.weight8
+4
+input396�0� 2layers.35.mlp.gate_proj.weight8
+%
+input397��� 2lm_head.weight8
+
+input398� 2norm.weight8'
+output0��	�2embed_tokens.weight>
+output1�
�
2'layers.0.self_attn.o_proj.o_proj.weight>
+output2�
 �
2)layers.0.self_attn.qkv_proj.v_proj.weight1
+output3�
 2layers.0.input_layernorm.weight6
+output4�
2%layers.0.self_attn.k_layernorm.weight>
+output5�
 @2)layers.0.self_attn.qkv_proj.k_proj.weight6
+output6�
2%layers.0.self_attn.q_layernorm.weight?
+output7�
 @2)layers.0.self_attn.qkv_proj.q_proj.weight3
+output8 �
�
2layers.0.mlp.down_proj.weight0
+output90�
 �
2layers.0.mlp.up_proj.weight;
+output10�
 2(layers.0.post_attention_layernorm.weight3
+output110�
 �
2layers.0.mlp.gate_proj.weight?
+output12�
�
2'layers.1.self_attn.o_proj.o_proj.weight?
+output13�
 �
2)layers.1.self_attn.qkv_proj.v_proj.weight2
+output14�
 2layers.1.input_layernorm.weight7
+output15�
2%layers.1.self_attn.k_layernorm.weight?
+output16�
 @2)layers.1.self_attn.qkv_proj.k_proj.weight7
+output17�
2%layers.1.self_attn.q_layernorm.weight@
+output18�
 @2)layers.1.self_attn.qkv_proj.q_proj.weight4
+output19 �
�
2layers.1.mlp.down_proj.weight1
+output200�
 �
2layers.1.mlp.up_proj.weight;
+output21�
 2(layers.1.post_attention_layernorm.weight3
+output220�
 �
2layers.1.mlp.gate_proj.weight?
+output23�
�
2'layers.2.self_attn.o_proj.o_proj.weight?
+output24�
 �
2)layers.2.self_attn.qkv_proj.v_proj.weight2
+output25�
 2layers.2.input_layernorm.weight7
+output26�
2%layers.2.self_attn.k_layernorm.weight?
+output27�
 @2)layers.2.self_attn.qkv_proj.k_proj.weight7
+output28�
2%layers.2.self_attn.q_layernorm.weight@
+output29�
 @2)layers.2.self_attn.qkv_proj.q_proj.weight4
+output30 �
�
2layers.2.mlp.down_proj.weight1
+output310�
 �
2layers.2.mlp.up_proj.weight;
+output32�
 2(layers.2.post_attention_layernorm.weight3
+output330�
 �
2layers.2.mlp.gate_proj.weight?
+output34�
�
2'layers.3.self_attn.o_proj.o_proj.weight?
+output35�
 �
2)layers.3.self_attn.qkv_proj.v_proj.weight2
+output36�
 2layers.3.input_layernorm.weight7
+output37�
2%layers.3.self_attn.k_layernorm.weight?
+output38�
 @2)layers.3.self_attn.qkv_proj.k_proj.weight7
+output39�
2%layers.3.self_attn.q_layernorm.weight@
+output40�
 @2)layers.3.self_attn.qkv_proj.q_proj.weight4
+output41 �
�
2layers.3.mlp.down_proj.weight1
+output420�
 �
2layers.3.mlp.up_proj.weight;
+output43�
 2(layers.3.post_attention_layernorm.weight3
+output440�
 �
2layers.3.mlp.gate_proj.weight?
+output45�
�
2'layers.4.self_attn.o_proj.o_proj.weight?
+output46�
 �
2)layers.4.self_attn.qkv_proj.v_proj.weight2
+output47�
 2layers.4.input_layernorm.weight7
+output48�
2%layers.4.self_attn.k_layernorm.weight?
+output49�
 @2)layers.4.self_attn.qkv_proj.k_proj.weight7
+output50�
2%layers.4.self_attn.q_layernorm.weight@
+output51�
 @2)layers.4.self_attn.qkv_proj.q_proj.weight4
+output52 �
�
2layers.4.mlp.down_proj.weight1
+output530�
 �
2layers.4.mlp.up_proj.weight;
+output54�
 2(layers.4.post_attention_layernorm.weight3
+output550�
 �
2layers.4.mlp.gate_proj.weight?
+output56�
�
2'layers.5.self_attn.o_proj.o_proj.weight?
+output57�
 �
2)layers.5.self_attn.qkv_proj.v_proj.weight2
+output58�
 2layers.5.input_layernorm.weight7
+output59�
2%layers.5.self_attn.k_layernorm.weight?
+output60�
 @2)layers.5.self_attn.qkv_proj.k_proj.weight7
+output61�
2%layers.5.self_attn.q_layernorm.weight@
+output62�
 @2)layers.5.self_attn.qkv_proj.q_proj.weight4
+output63 �
�
2layers.5.mlp.down_proj.weight1
+output640�
 �
2layers.5.mlp.up_proj.weight;
+output65�
 2(layers.5.post_attention_layernorm.weight3
+output660�
 �
2layers.5.mlp.gate_proj.weight?
+output67�
�
2'layers.6.self_attn.o_proj.o_proj.weight?
+output68�
 �
2)layers.6.self_attn.qkv_proj.v_proj.weight2
+output69�
 2layers.6.input_layernorm.weight7
+output70�
2%layers.6.self_attn.k_layernorm.weight?
+output71�
 @2)layers.6.self_attn.qkv_proj.k_proj.weight7
+output72�
2%layers.6.self_attn.q_layernorm.weight@
+output73�
 @2)layers.6.self_attn.qkv_proj.q_proj.weight4
+output74 �
�
2layers.6.mlp.down_proj.weight1
+output750�
 �
2layers.6.mlp.up_proj.weight;
+output76�
 2(layers.6.post_attention_layernorm.weight3
+output770�
 �
2layers.6.mlp.gate_proj.weight?
+output78�
�
2'layers.7.self_attn.o_proj.o_proj.weight?
+output79�
 �
2)layers.7.self_attn.qkv_proj.v_proj.weight2
+output80�
 2layers.7.input_layernorm.weight7
+output81�
2%layers.7.self_attn.k_layernorm.weight?
+output82�
 @2)layers.7.self_attn.qkv_proj.k_proj.weight7
+output83�
2%layers.7.self_attn.q_layernorm.weight@
+output84�
 @2)layers.7.self_attn.qkv_proj.q_proj.weight4
+output85 �
�
2layers.7.mlp.down_proj.weight1
+output860�
 �
2layers.7.mlp.up_proj.weight;
+output87�
 2(layers.7.post_attention_layernorm.weight3
+output880�
 �
2layers.7.mlp.gate_proj.weight?
+output89�
�
2'layers.8.self_attn.o_proj.o_proj.weight?
+output90�
 �
2)layers.8.self_attn.qkv_proj.v_proj.weight2
+output91�
 2layers.8.input_layernorm.weight7
+output92�
2%layers.8.self_attn.k_layernorm.weight?
+output93�
 @2)layers.8.self_attn.qkv_proj.k_proj.weight7
+output94�
2%layers.8.self_attn.q_layernorm.weight@
+output95�
 @2)layers.8.self_attn.qkv_proj.q_proj.weight4
+output96 �
�
2layers.8.mlp.down_proj.weight1
+output970�
 �
2layers.8.mlp.up_proj.weight;
+output98�
 2(layers.8.post_attention_layernorm.weight3
+output990�
 �
2layers.8.mlp.gate_proj.weight@
+	output100�
�
2'layers.9.self_attn.o_proj.o_proj.weight@
+	output101�
 �
2)layers.9.self_attn.qkv_proj.v_proj.weight3
+	output102�
 2layers.9.input_layernorm.weight8
+	output103�
2%layers.9.self_attn.k_layernorm.weight@
+	output104�
 @2)layers.9.self_attn.qkv_proj.k_proj.weight8
+	output105�
2%layers.9.self_attn.q_layernorm.weightA
+	output106�
 @2)layers.9.self_attn.qkv_proj.q_proj.weight5
+	output107 �
�
2layers.9.mlp.down_proj.weight2
+	output1080�
 �
2layers.9.mlp.up_proj.weight<
+	output109�
 2(layers.9.post_attention_layernorm.weight4
+	output1100�
 �
2layers.9.mlp.gate_proj.weightA
+	output111�
�
2(layers.10.self_attn.o_proj.o_proj.weightA
+	output112�
 �
2*layers.10.self_attn.qkv_proj.v_proj.weight4
+	output113�
 2 layers.10.input_layernorm.weight9
+	output114�
2&layers.10.self_attn.k_layernorm.weightA
+	output115�
 @2*layers.10.self_attn.qkv_proj.k_proj.weight9
+	output116�
2&layers.10.self_attn.q_layernorm.weightB
+	output117�
 @2*layers.10.self_attn.qkv_proj.q_proj.weight6
+	output118 �
�
2layers.10.mlp.down_proj.weight3
+	output1190�
 �
2layers.10.mlp.up_proj.weight=
+	output120�
 2)layers.10.post_attention_layernorm.weight5
+	output1210�
 �
2layers.10.mlp.gate_proj.weightA
+	output122�
�
2(layers.11.self_attn.o_proj.o_proj.weightA
+	output123�
 �
2*layers.11.self_attn.qkv_proj.v_proj.weight4
+	output124�
 2 layers.11.input_layernorm.weight9
+	output125�
2&layers.11.self_attn.k_layernorm.weightA
+	output126�
 @2*layers.11.self_attn.qkv_proj.k_proj.weight9
+	output127�
2&layers.11.self_attn.q_layernorm.weightB
+	output128�
 @2*layers.11.self_attn.qkv_proj.q_proj.weight6
+	output129 �
�
2layers.11.mlp.down_proj.weight3
+	output1300�
 �
2layers.11.mlp.up_proj.weight=
+	output131�
 2)layers.11.post_attention_layernorm.weight5
+	output1320�
 �
2layers.11.mlp.gate_proj.weightA
+	output133�
�
2(layers.12.self_attn.o_proj.o_proj.weightA
+	output134�
 �
2*layers.12.self_attn.qkv_proj.v_proj.weight4
+	output135�
 2 layers.12.input_layernorm.weight9
+	output136�
2&layers.12.self_attn.k_layernorm.weightA
+	output137�
 @2*layers.12.self_attn.qkv_proj.k_proj.weight9
+	output138�
2&layers.12.self_attn.q_layernorm.weightB
+	output139�
 @2*layers.12.self_attn.qkv_proj.q_proj.weight6
+	output140 �
�
2layers.12.mlp.down_proj.weight3
+	output1410�
 �
2layers.12.mlp.up_proj.weight=
+	output142�
 2)layers.12.post_attention_layernorm.weight5
+	output1430�
 �
2layers.12.mlp.gate_proj.weightA
+	output144�
�
2(layers.13.self_attn.o_proj.o_proj.weightA
+	output145�
 �
2*layers.13.self_attn.qkv_proj.v_proj.weight4
+	output146�
 2 layers.13.input_layernorm.weight9
+	output147�
2&layers.13.self_attn.k_layernorm.weightA
+	output148�
 @2*layers.13.self_attn.qkv_proj.k_proj.weight9
+	output149�
2&layers.13.self_attn.q_layernorm.weightB
+	output150�
 @2*layers.13.self_attn.qkv_proj.q_proj.weight6
+	output151 �
�
2layers.13.mlp.down_proj.weight3
+	output1520�
 �
2layers.13.mlp.up_proj.weight=
+	output153�
 2)layers.13.post_attention_layernorm.weight5
+	output1540�
 �
2layers.13.mlp.gate_proj.weightA
+	output155�
�
2(layers.14.self_attn.o_proj.o_proj.weightA
+	output156�
 �
2*layers.14.self_attn.qkv_proj.v_proj.weight4
+	output157�
 2 layers.14.input_layernorm.weight9
+	output158�
2&layers.14.self_attn.k_layernorm.weightA
+	output159�
 @2*layers.14.self_attn.qkv_proj.k_proj.weight9
+	output160�
2&layers.14.self_attn.q_layernorm.weightB
+	output161�
 @2*layers.14.self_attn.qkv_proj.q_proj.weight6
+	output162 �
�
2layers.14.mlp.down_proj.weight3
+	output1630�
 �
2layers.14.mlp.up_proj.weight=
+	output164�
 2)layers.14.post_attention_layernorm.weight5
+	output1650�
 �
2layers.14.mlp.gate_proj.weightA
+	output166�
�
2(layers.15.self_attn.o_proj.o_proj.weightA
+	output167�
 �
2*layers.15.self_attn.qkv_proj.v_proj.weight4
+	output168�
 2 layers.15.input_layernorm.weight9
+	output169�
2&layers.15.self_attn.k_layernorm.weightA
+	output170�
 @2*layers.15.self_attn.qkv_proj.k_proj.weight9
+	output171�
2&layers.15.self_attn.q_layernorm.weightB
+	output172�
 @2*layers.15.self_attn.qkv_proj.q_proj.weight6
+	output173 �
�
2layers.15.mlp.down_proj.weight3
+	output1740�
 �
2layers.15.mlp.up_proj.weight=
+	output175�
 2)layers.15.post_attention_layernorm.weight5
+	output1760�
 �
2layers.15.mlp.gate_proj.weightA
+	output177�
�
2(layers.16.self_attn.o_proj.o_proj.weightA
+	output178�
 �
2*layers.16.self_attn.qkv_proj.v_proj.weight4
+	output179�
 2 layers.16.input_layernorm.weight9
+	output180�
2&layers.16.self_attn.k_layernorm.weightA
+	output181�
 @2*layers.16.self_attn.qkv_proj.k_proj.weight9
+	output182�
2&layers.16.self_attn.q_layernorm.weightB
+	output183�
 @2*layers.16.self_attn.qkv_proj.q_proj.weight6
+	output184 �
�
2layers.16.mlp.down_proj.weight3
+	output1850�
 �
2layers.16.mlp.up_proj.weight=
+	output186�
 2)layers.16.post_attention_layernorm.weight5
+	output1870�
 �
2layers.16.mlp.gate_proj.weightA
+	output188�
�
2(layers.17.self_attn.o_proj.o_proj.weightA
+	output189�
 �
2*layers.17.self_attn.qkv_proj.v_proj.weight4
+	output190�
 2 layers.17.input_layernorm.weight9
+	output191�
2&layers.17.self_attn.k_layernorm.weightA
+	output192�
 @2*layers.17.self_attn.qkv_proj.k_proj.weight9
+	output193�
2&layers.17.self_attn.q_layernorm.weightB
+	output194�
 @2*layers.17.self_attn.qkv_proj.q_proj.weight6
+	output195 �
�
2layers.17.mlp.down_proj.weight3
+	output1960�
 �
2layers.17.mlp.up_proj.weight=
+	output197�
 2)layers.17.post_attention_layernorm.weight5
+	output1980�
 �
2layers.17.mlp.gate_proj.weightA
+	output199�
�
2(layers.18.self_attn.o_proj.o_proj.weightA
+	output200�
 �
2*layers.18.self_attn.qkv_proj.v_proj.weight4
+	output201�
 2 layers.18.input_layernorm.weight9
+	output202�
2&layers.18.self_attn.k_layernorm.weightA
+	output203�
 @2*layers.18.self_attn.qkv_proj.k_proj.weight9
+	output204�
2&layers.18.self_attn.q_layernorm.weightB
+	output205�
 @2*layers.18.self_attn.qkv_proj.q_proj.weight6
+	output206 �
�
2layers.18.mlp.down_proj.weight3
+	output2070�
 �
2layers.18.mlp.up_proj.weight=
+	output208�
 2)layers.18.post_attention_layernorm.weight5
+	output2090�
 �
2layers.18.mlp.gate_proj.weightA
+	output210�
�
2(layers.19.self_attn.o_proj.o_proj.weightA
+	output211�
 �
2*layers.19.self_attn.qkv_proj.v_proj.weight4
+	output212�
 2 layers.19.input_layernorm.weight9
+	output213�
2&layers.19.self_attn.k_layernorm.weightA
+	output214�
 @2*layers.19.self_attn.qkv_proj.k_proj.weight9
+	output215�
2&layers.19.self_attn.q_layernorm.weightB
+	output216�
 @2*layers.19.self_attn.qkv_proj.q_proj.weight6
+	output217 �
�
2layers.19.mlp.down_proj.weight3
+	output2180�
 �
2layers.19.mlp.up_proj.weight=
+	output219�
 2)layers.19.post_attention_layernorm.weight5
+	output2200�
 �
2layers.19.mlp.gate_proj.weightA
+	output221�
�
2(layers.20.self_attn.o_proj.o_proj.weightA
+	output222�
 �
2*layers.20.self_attn.qkv_proj.v_proj.weight4
+	output223�
 2 layers.20.input_layernorm.weight9
+	output224�
2&layers.20.self_attn.k_layernorm.weightA
+	output225�
 @2*layers.20.self_attn.qkv_proj.k_proj.weight9
+	output226�
2&layers.20.self_attn.q_layernorm.weightB
+	output227�
 @2*layers.20.self_attn.qkv_proj.q_proj.weight6
+	output228 �
�
2layers.20.mlp.down_proj.weight3
+	output2290�
 �
2layers.20.mlp.up_proj.weight=
+	output230�
 2)layers.20.post_attention_layernorm.weight5
+	output2310�
 �
2layers.20.mlp.gate_proj.weightA
+	output232�
�
2(layers.21.self_attn.o_proj.o_proj.weightA
+	output233�
 �
2*layers.21.self_attn.qkv_proj.v_proj.weight4
+	output234�
 2 layers.21.input_layernorm.weight9
+	output235�
2&layers.21.self_attn.k_layernorm.weightA
+	output236�
 @2*layers.21.self_attn.qkv_proj.k_proj.weight9
+	output237�
2&layers.21.self_attn.q_layernorm.weightB
+	output238�
 @2*layers.21.self_attn.qkv_proj.q_proj.weight6
+	output239 �
�
2layers.21.mlp.down_proj.weight3
+	output2400�
 �
2layers.21.mlp.up_proj.weight=
+	output241�
 2)layers.21.post_attention_layernorm.weight5
+	output2420�
 �
2layers.21.mlp.gate_proj.weightA
+	output243�
�
2(layers.22.self_attn.o_proj.o_proj.weightA
+	output244�
 �
2*layers.22.self_attn.qkv_proj.v_proj.weight4
+	output245�
 2 layers.22.input_layernorm.weight9
+	output246�
2&layers.22.self_attn.k_layernorm.weightA
+	output247�
 @2*layers.22.self_attn.qkv_proj.k_proj.weight9
+	output248�
2&layers.22.self_attn.q_layernorm.weightB
+	output249�
 @2*layers.22.self_attn.qkv_proj.q_proj.weight6
+	output250 �
�
2layers.22.mlp.down_proj.weight3
+	output2510�
 �
2layers.22.mlp.up_proj.weight=
+	output252�
 2)layers.22.post_attention_layernorm.weight5
+	output2530�
 �
2layers.22.mlp.gate_proj.weightA
+	output254�
�
2(layers.23.self_attn.o_proj.o_proj.weightA
+	output255�
 �
2*layers.23.self_attn.qkv_proj.v_proj.weight4
+	output256�
 2 layers.23.input_layernorm.weight9
+	output257�
2&layers.23.self_attn.k_layernorm.weightA
+	output258�
 @2*layers.23.self_attn.qkv_proj.k_proj.weight9
+	output259�
2&layers.23.self_attn.q_layernorm.weightB
+	output260�
 @2*layers.23.self_attn.qkv_proj.q_proj.weight6
+	output261 �
�
2layers.23.mlp.down_proj.weight3
+	output2620�
 �
2layers.23.mlp.up_proj.weight=
+	output263�
 2)layers.23.post_attention_layernorm.weight5
+	output2640�
 �
2layers.23.mlp.gate_proj.weightA
+	output265�
�
2(layers.24.self_attn.o_proj.o_proj.weightA
+	output266�
 �
2*layers.24.self_attn.qkv_proj.v_proj.weight4
+	output267�
 2 layers.24.input_layernorm.weight9
+	output268�
2&layers.24.self_attn.k_layernorm.weightA
+	output269�
 @2*layers.24.self_attn.qkv_proj.k_proj.weight9
+	output270�
2&layers.24.self_attn.q_layernorm.weightB
+	output271�
 @2*layers.24.self_attn.qkv_proj.q_proj.weight6
+	output272 �
�
2layers.24.mlp.down_proj.weight3
+	output2730�
 �
2layers.24.mlp.up_proj.weight=
+	output274�
 2)layers.24.post_attention_layernorm.weight5
+	output2750�
 �
2layers.24.mlp.gate_proj.weightA
+	output276�
�
2(layers.25.self_attn.o_proj.o_proj.weightA
+	output277�
 �
2*layers.25.self_attn.qkv_proj.v_proj.weight4
+	output278�
 2 layers.25.input_layernorm.weight9
+	output279�
2&layers.25.self_attn.k_layernorm.weightA
+	output280�
 @2*layers.25.self_attn.qkv_proj.k_proj.weight9
+	output281�
2&layers.25.self_attn.q_layernorm.weightB
+	output282�
 @2*layers.25.self_attn.qkv_proj.q_proj.weight6
+	output283 �
�
2layers.25.mlp.down_proj.weight3
+	output2840�
 �
2layers.25.mlp.up_proj.weight=
+	output285�
 2)layers.25.post_attention_layernorm.weight5
+	output2860�
 �
2layers.25.mlp.gate_proj.weightA
+	output287�
�
2(layers.26.self_attn.o_proj.o_proj.weightA
+	output288�
 �
2*layers.26.self_attn.qkv_proj.v_proj.weight4
+	output289�
 2 layers.26.input_layernorm.weight9
+	output290�
2&layers.26.self_attn.k_layernorm.weightA
+	output291�
 @2*layers.26.self_attn.qkv_proj.k_proj.weight9
+	output292�
2&layers.26.self_attn.q_layernorm.weightB
+	output293�
 @2*layers.26.self_attn.qkv_proj.q_proj.weight6
+	output294 �
�
2layers.26.mlp.down_proj.weight3
+	output2950�
 �
2layers.26.mlp.up_proj.weight=
+	output296�
 2)layers.26.post_attention_layernorm.weight5
+	output2970�
 �
2layers.26.mlp.gate_proj.weightA
+	output298�
�
2(layers.27.self_attn.o_proj.o_proj.weightA
+	output299�
 �
2*layers.27.self_attn.qkv_proj.v_proj.weight4
+	output300�
 2 layers.27.input_layernorm.weight9
+	output301�
2&layers.27.self_attn.k_layernorm.weightA
+	output302�
 @2*layers.27.self_attn.qkv_proj.k_proj.weight9
+	output303�
2&layers.27.self_attn.q_layernorm.weightB
+	output304�
 @2*layers.27.self_attn.qkv_proj.q_proj.weight6
+	output305 �
�
2layers.27.mlp.down_proj.weight3
+	output3060�
 �
2layers.27.mlp.up_proj.weight=
+	output307�
 2)layers.27.post_attention_layernorm.weight5
+	output3080�
 �
2layers.27.mlp.gate_proj.weightA
+	output309�
�
2(layers.28.self_attn.o_proj.o_proj.weightA
+	output310�
 �
2*layers.28.self_attn.qkv_proj.v_proj.weight4
+	output311�
 2 layers.28.input_layernorm.weight9
+	output312�
2&layers.28.self_attn.k_layernorm.weightA
+	output313�
 @2*layers.28.self_attn.qkv_proj.k_proj.weight9
+	output314�
2&layers.28.self_attn.q_layernorm.weightB
+	output315�
 @2*layers.28.self_attn.qkv_proj.q_proj.weight6
+	output316 �
�
2layers.28.mlp.down_proj.weight3
+	output3170�
 �
2layers.28.mlp.up_proj.weight=
+	output318�
 2)layers.28.post_attention_layernorm.weight5
+	output3190�
 �
2layers.28.mlp.gate_proj.weightA
+	output320�
�
2(layers.29.self_attn.o_proj.o_proj.weightA
+	output321�
 �
2*layers.29.self_attn.qkv_proj.v_proj.weight4
+	output322�
 2 layers.29.input_layernorm.weight9
+	output323�
2&layers.29.self_attn.k_layernorm.weightA
+	output324�
 @2*layers.29.self_attn.qkv_proj.k_proj.weight9
+	output325�
2&layers.29.self_attn.q_layernorm.weightB
+	output326�
 @2*layers.29.self_attn.qkv_proj.q_proj.weight6
+	output327 �
�
2layers.29.mlp.down_proj.weight3
+	output3280�
 �
2layers.29.mlp.up_proj.weight=
+	output329�
 2)layers.29.post_attention_layernorm.weight5
+	output3300�
 �
2layers.29.mlp.gate_proj.weightA
+	output331�
�
2(layers.30.self_attn.o_proj.o_proj.weightA
+	output332�
 �
2*layers.30.self_attn.qkv_proj.v_proj.weight4
+	output333�
 2 layers.30.input_layernorm.weight9
+	output334�
2&layers.30.self_attn.k_layernorm.weightA
+	output335�
 @2*layers.30.self_attn.qkv_proj.k_proj.weight9
+	output336�
2&layers.30.self_attn.q_layernorm.weightB
+	output337�
 @2*layers.30.self_attn.qkv_proj.q_proj.weight6
+	output338 �
�
2layers.30.mlp.down_proj.weight3
+	output3390�
 �
2layers.30.mlp.up_proj.weight=
+	output340�
 2)layers.30.post_attention_layernorm.weight5
+	output3410�
 �
2layers.30.mlp.gate_proj.weightA
+	output342�
�
2(layers.31.self_attn.o_proj.o_proj.weightA
+	output343�
 �
2*layers.31.self_attn.qkv_proj.v_proj.weight4
+	output344�
 2 layers.31.input_layernorm.weight9
+	output345�
2&layers.31.self_attn.k_layernorm.weightA
+	output346�
 @2*layers.31.self_attn.qkv_proj.k_proj.weight9
+	output347�
2&layers.31.self_attn.q_layernorm.weightB
+	output348�
 @2*layers.31.self_attn.qkv_proj.q_proj.weight6
+	output349 �
�
2layers.31.mlp.down_proj.weight3
+	output3500�
 �
2layers.31.mlp.up_proj.weight=
+	output351�
 2)layers.31.post_attention_layernorm.weight5
+	output3520�
 �
2layers.31.mlp.gate_proj.weightA
+	output353�
�
2(layers.32.self_attn.o_proj.o_proj.weightA
+	output354�
 �
2*layers.32.self_attn.qkv_proj.v_proj.weight4
+	output355�
 2 layers.32.input_layernorm.weight9
+	output356�
2&layers.32.self_attn.k_layernorm.weightA
+	output357�
 @2*layers.32.self_attn.qkv_proj.k_proj.weight9
+	output358�
2&layers.32.self_attn.q_layernorm.weightB
+	output359�
 @2*layers.32.self_attn.qkv_proj.q_proj.weight6
+	output360 �
�
2layers.32.mlp.down_proj.weight3
+	output3610�
 �
2layers.32.mlp.up_proj.weight=
+	output362�
 2)layers.32.post_attention_layernorm.weight5
+	output3630�
 �
2layers.32.mlp.gate_proj.weightA
+	output364�
�
2(layers.33.self_attn.o_proj.o_proj.weightA
+	output365�
 �
2*layers.33.self_attn.qkv_proj.v_proj.weight4
+	output366�
 2 layers.33.input_layernorm.weight9
+	output367�
2&layers.33.self_attn.k_layernorm.weightA
+	output368�
 @2*layers.33.self_attn.qkv_proj.k_proj.weight9
+	output369�
2&layers.33.self_attn.q_layernorm.weightB
+	output370�
 @2*layers.33.self_attn.qkv_proj.q_proj.weight6
+	output371 �
�
2layers.33.mlp.down_proj.weight3
+	output3720�
 �
2layers.33.mlp.up_proj.weight=
+	output373�
 2)layers.33.post_attention_layernorm.weight5
+	output3740�
 �
2layers.33.mlp.gate_proj.weightA
+	output375�
�
2(layers.34.self_attn.o_proj.o_proj.weightA
+	output376�
 �
2*layers.34.self_attn.qkv_proj.v_proj.weight4
+	output377�
 2 layers.34.input_layernorm.weight9
+	output378�
2&layers.34.self_attn.k_layernorm.weightA
+	output379�
 @2*layers.34.self_attn.qkv_proj.k_proj.weight9
+	output380�
2&layers.34.self_attn.q_layernorm.weightB
+	output381�
 @2*layers.34.self_attn.qkv_proj.q_proj.weight6
+	output382 �
�
2layers.34.mlp.down_proj.weight3
+	output3830�
 �
2layers.34.mlp.up_proj.weight=
+	output384�
 2)layers.34.post_attention_layernorm.weight5
+	output3850�
 �
2layers.34.mlp.gate_proj.weightA
+	output386�
�
2(layers.35.self_attn.o_proj.o_proj.weightA
+	output387�
 �
2*layers.35.self_attn.qkv_proj.v_proj.weight4
+	output388�
 2 layers.35.input_layernorm.weight9
+	output389�
2&layers.35.self_attn.k_layernorm.weightA
+	output390�
 @2*layers.35.self_attn.qkv_proj.k_proj.weight9
+	output391�
2&layers.35.self_attn.q_layernorm.weightB
+	output392�
 @2*layers.35.self_attn.qkv_proj.q_proj.weight6
+	output393 �
�
2layers.35.mlp.down_proj.weight3
+	output3940�
 �
2layers.35.mlp.up_proj.weight=
+	output395�
 2)layers.35.post_attention_layernorm.weight5
+	output3960�
 �
2layers.35.mlp.gate_proj.weight$
+	output397��� 2lm_head.weight
+	output398�
 2norm.weight

layout_opt/model/graph.hlo

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:12b45b028e502b2dd8c42c1287fbdbea434454143a30d473806853bc18673d98
+size 211060

model.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:c36077018a9f85728962cc73bfcba755ce1d5d5b6f608dacf65d7b95596eb109
+size 47198475

neuron_config.json

@@ -0,0 +1,218 @@
+{
+  "_attn_implementation_autoset": false,
+  "_name_or_path": "Qwen/Qwen3-8B",
+  "add_cross_attention": false,
+  "architectures": [
+    "Qwen3ForCausalLM"
+  ],
+  "attention_bias": false,
+  "attention_dropout": 0.0,
+  "attribute_map": {},
+  "bad_words_ids": null,
+  "begin_suppress_tokens": null,
+  "bos_token_id": 151643,
+  "chunk_size_feed_forward": 0,
+  "cross_attention_hidden_size": null,
+  "decoder_start_token_id": null,
+  "diversity_penalty": 0.0,
+  "do_sample": false,
+  "early_stopping": false,
+  "encoder_no_repeat_ngram_size": 0,
+  "eos_token_id": 151645,
+  "exponential_decay_length_penalty": null,
+  "finetuning_task": null,
+  "forced_bos_token_id": null,
+  "forced_eos_token_id": null,
+  "fused_spec_config": null,
+  "head_dim": 128,
+  "hidden_act": "silu",
+  "hidden_size": 4096,
+  "id2label": {
+    "0": "LABEL_0",
+    "1": "LABEL_1"
+  },
+  "initializer_range": 0.02,
+  "intermediate_size": 12288,
+  "is_decoder": false,
+  "is_encoder_decoder": false,
+  "label2id": {
+    "LABEL_0": 0,
+    "LABEL_1": 1
+  },
+  "length_penalty": 1.0,
+  "max_length": 20,
+  "max_position_embeddings": 40960,
+  "max_window_layers": 36,
+  "metadata": null,
+  "min_length": 0,
+  "model_type": "qwen3",
+  "neuron_config": {
+    "activation_quantization_type": null,
+    "allow_input_truncation": false,
+    "apply_seq_ids_mask": false,
+    "async_mode": false,
+    "attention_dp_degree": 1,
+    "attention_dtype": null,
+    "attn_block_cte_nki_kernel_enabled": false,
+    "attn_block_tkg_nki_kernel_cache_update": false,
+    "attn_block_tkg_nki_kernel_enabled": false,
+    "attn_cls": {
+      "__module__": "neuronx_distributed_inference.models.qwen3.modeling_qwen3",
+      "__name__": "NeuronQwen3Attention"
+    },
+    "attn_kernel_enabled": null,
+    "attn_tkg_builtin_kernel_enabled": false,
+    "attn_tkg_nki_kernel_enabled": false,
+    "batch_size": 1,
+    "bucket_n_active_tokens": false,
+    "buckets": [
+      1024
+    ],
+    "cast_type": "config",
+    "cc_pipeline_tiling_factor": 2,
+    "chunked_prefill_config": null,
+    "context_encoding_buckets": null,
+    "cp_degree": 1,
+    "ctx_batch_size": 1,
+    "disable_kv_cache_tiling": false,
+    "draft_model_modules_to_not_convert": null,
+    "enable_bucketing": true,
+    "enable_eagle_draft_input_norm": false,
+    "enable_eagle_speculation": false,
+    "enable_fused_speculation": false,
+    "enable_long_context_mode": false,
+    "enable_output_completion_notifications": false,
+    "enable_spill_reload_dge": false,
+    "enable_token_tree": false,
+    "ep_degree": 1,
+    "expert_mlp_nki_kernel_enabled": null,
+    "flash_decoding_enabled": false,
+    "fused_qkv": false,
+    "fused_rmsnorm_skip_gamma": false,
+    "is_block_kv_layout": null,
+    "is_chunked_prefill": false,
+    "is_continuous_batching": true,
+    "is_eagle_draft": false,
+    "is_medusa": false,
+    "is_prefill_stage": null,
+    "is_prefix_caching": false,
+    "k_cache_transposed": false,
+    "kv_cache_batch_size": 1,
+    "kv_cache_padding_size": 0,
+    "kv_cache_quant": false,
+    "kv_cache_tiling": false,
+    "layer_boundary_markers": false,
+    "lm_head_pad": false,
+    "lm_head_pad_alignment_size": 1,
+    "local_ranks_size": 2,
+    "logical_nc_config": 1,
+    "lora_config": null,
+    "max_batch_size": 1,
+    "max_context_length": 1024,
+    "max_length": 1024,
+    "max_new_tokens": null,
+    "medusa_speculation_length": 0,
+    "medusa_tree": null,
+    "mlp_kernel_enabled": false,
+    "mlp_kernel_fuse_residual_add": false,
+    "modules_to_not_convert": null,
+    "moe_fused_nki_kernel_enabled": null,
+    "n_active_tokens": 1024,
+    "n_positions": 1024,
+    "num_medusa_heads": 0,
+    "on_cpu": false,
+    "on_device_sampling_config": {
+      "deterministic": false,
+      "do_sample": false,
+      "dynamic": true,
+      "global_topk": 256,
+      "on_device_sampling_config": true,
+      "temperature": 1.0,
+      "top_k": 1,
+      "top_k_kernel_enabled": false,
+      "top_p": 1.0
+    },
+    "output_logits": false,
+    "overrides_torch_dtype": true,
+    "pa_block_size": 1024,
+    "pa_num_blocks": 1,
+    "padding_side": "right",
+    "pp_degree": 1,
+    "prefix_buckets": null,
+    "qk_layernorm": false,
+    "qkv_kernel_enabled": false,
+    "qkv_kernel_fuse_residual_add": false,
+    "qkv_kernel_nbsd_layout": false,
+    "quantization_dtype": "int8",
+    "quantization_type": "per_tensor_symmetric",
+    "quantize_clamp_bound": Infinity,
+    "quantized": false,
+    "quantized_checkpoints_path": null,
+    "quantized_mlp_kernel_enabled": false,
+    "rmsnorm_quantize_kernel_enabled": false,
+    "router_topk_nki_kernel_enabled": null,
+    "rpl_reduce_dtype": null,
+    "save_sharded_checkpoint": true,
+    "scratchpad_page_size": null,
+    "seq_len": 1024,
+    "seq_len_threshold_for_cc_tiling": 16384,
+    "sequence_parallel_enabled": false,
+    "shared_mlp_nki_kernel_enabled": null,
+    "skip_sharding": false,
+    "skip_warmup": false,
+    "spec_batch_size": 1,
+    "speculation_length": 0,
+    "start_rank_id": 0,
+    "target": null,
+    "tile_cc": false,
+    "tkg_batch_size": 1,
+    "token_generation_buckets": null,
+    "token_tree_config": null,
+    "torch_dtype": "bfloat16",
+    "tp_degree": 2,
+    "vocab_parallel": false,
+    "weight_gather_seq_len_threshold": 32768,
+    "weights_to_skip_layout_optimization": [],
+    "world_size": 2
+  },
+  "no_repeat_ngram_size": 0,
+  "num_attention_heads": 32,
+  "num_beam_groups": 1,
+  "num_beams": 1,
+  "num_cores_per_group": 1,
+  "num_hidden_layers": 36,
+  "num_key_value_heads": 8,
+  "num_return_sequences": 1,
+  "output_attentions": false,
+  "output_hidden_states": false,
+  "output_scores": false,
+  "pad_token_id": null,
+  "prefix": null,
+  "problem_type": null,
+  "pruned_heads": {},
+  "remove_invalid_values": false,
+  "repetition_penalty": 1.0,
+  "return_dict": true,
+  "return_dict_in_generate": false,
+  "rms_norm_eps": 1e-06,
+  "rope_scaling": null,
+  "rope_theta": 1000000,
+  "sep_token_id": null,
+  "sliding_window": null,
+  "suppress_tokens": null,
+  "task_specific_params": null,
+  "temperature": 1.0,
+  "tf_legacy_loss": false,
+  "tie_encoder_decoder": false,
+  "tie_word_embeddings": false,
+  "tokenizer_class": null,
+  "top_k": 50,
+  "top_p": 1.0,
+  "torchscript": false,
+  "transformers_version": "4.51.0",
+  "typical_p": 1.0,
+  "use_bfloat16": false,
+  "use_cache": true,
+  "use_sliding_window": false,
+  "vocab_size": 151936
+}

token_generation_model/_tp0_bk0/command.txt

@@ -0,0 +1 @@
+neuronx-cc compile --framework=XLA model.MODULE_6ef5ba8b41fbbe77f080+74ae8282.hlo_module.pb --output model.MODULE_6ef5ba8b41fbbe77f080+74ae8282.neff --target=trn1 --auto-cast=none --model-type=transformer '--tensorizer-options=--enable-ccop-compute-overlap --cc-pipeline-tiling-factor=1 --vectorize-strided-dma ' --lnc=1 -O2 --internal-hlo2tensorizer-options=--verify-hlo=true --logfile=log-neuron-cc.txt --enable-internal-neff-wrapper --verbose=35

token_generation_model/_tp0_bk0/compile_flags.MODULE_6ef5ba8b41fbbe77f080+74ae8282.json

@@ -0,0 +1 @@
+["--target=trn1", "--auto-cast=none", "--model-type=transformer", "--tensorizer-options=--enable-ccop-compute-overlap --cc-pipeline-tiling-factor=1 --vectorize-strided-dma ", "--lnc=1", "-O2", "--internal-hlo2tensorizer-options=--verify-hlo=true", "--logfile=/home/ubuntu/qwen3/token_generation_model/_tp0_bk0/log-neuron-cc.txt", "--enable-internal-neff-wrapper"]

token_generation_model/_tp0_bk0/global_metric_store.json

@@ -0,0 +1,540 @@
+{
+  "Average": {
+    "tensorizer": {
+      "StaticProfiler::AverageFractalPeUtilization": 99.8321762084961,
+      "StaticProfiler::AveragePartitionUtilization": 99.3888168334961,
+      "StaticProfiler::AveragePeUtilization": 99.65400695800781,
+      "StaticProfiler::LocalizationEfficiency": 109.9806137084961,
+      "StaticProfiler::LocalizationEfficiencyIgnoreNonlocal": 110.06793212890625,
+      "TilingProfiler::AveragePartitionUtilizationAfterTiling": 0,
+      "TilingProfiler::AveragePeUtilizationAfterTiling": 0
+    }
+  },
+  "Count": {
+    "tensorizer": {
+      "StaticProfiler::AverageFractalPeUtilization": 1,
+      "StaticProfiler::AveragePartitionUtilization": 1,
+      "StaticProfiler::AveragePeUtilization": 1,
+      "StaticProfiler::LocalizationEfficiency": 1,
+      "StaticProfiler::LocalizationEfficiencyIgnoreNonlocal": 1,
+      "TilingProfiler::AveragePartitionUtilizationAfterTiling": 1,
+      "TilingProfiler::AveragePeUtilizationAfterTiling": 1
+    }
+  },
+  "Sum": {
+    "compiletime": {
+      "AGOrderingAnalysisPass": 1.4457588195800781,
+      "AffinePredicateResolution": 0.05167531967163086,
+      "AliasDependencyElimination": 0.0026276111602783203,
+      "AliasDependencyInduction": 0.44934630393981934,
+      "AliasDependencyReset": 1.2677826881408691,
+      "BFComputeCutting": 0.06423807144165039,
+      "BirCodeGenLoop": 2.421293258666992,
+      "CCOpFusion": 0.41050028800964355,
+      "CanonicalizeConv": 9.999999974752427e-07,
+      "CanonicalizeDAGForPGTiling": 0.21233797073364258,
+      "CanonicalizeForTensorizer": 0.0003640000068116933,
+      "CanonicalizeIR": 0.06626629829406738,
+      "Canonicalizer": 0.007044999860227108,
+      "CoalesceCCOp": 0.19146490097045898,
+      "CommuteConcat": 0.03319668769836426,
+      "DMALocalityOpt": 0.035207271575927734,
+      "DMAProfiler": 0.08866691589355469,
+      "DMATilingProfiler": 0.07109546661376953,
+      "DataLocalityOpt": 1.910703182220459,
+      "DataStreaming": 0.15389323234558105,
+      "DeConcat": 0.012087583541870117,
+      "DeadCodeElimination": 0.035611867904663086,
+      "DeadStoreElimination": 0.37193870544433594,
+      "DelinearIndices": 0.2894127368927002,
+      "Delinearization": 0.1295926570892334,
+      "DoNothing": 0.00019550323486328125,
+      "DramToDramTranspose": 1.0679569244384766,
+      "DumpGraphAndMetadata": 0.24142217636108398,
+      "EliminateDivs": 0.17337489128112793,
+      "ExpandBatchNorm": 0.06027984619140625,
+      "ExpandISAMacro": 0.0909569263458252,
+      "FactorizeBlkDims": 0.24945974349975586,
+      "FactorizeThreadAxesInFreeDims": 0.03613853454589844,
+      "FlattenMacroLoop": 0.26774168014526367,
+      "GenericAccessSimplifier": 0.03175926208496094,
+      "HoistCompute": 4.8000001697801054e-05,
+      "IdentifyCrossPassTensors": 0.00013600000238511711,
+      "InferInitValue": 1.029360294342041,
+      "InferIntrinsicOnCC": 0.34307408332824707,
+      "InferNeuronTensor": 1.7935998439788818,
+      "InferNonlocalTensors": 3.6307339668273926,
+      "InferPSumTensor": 0.9782986640930176,
+      "InlineNativeKernels": 0.05374264717102051,
+      "InsertIOTransposes": 1.162278652191162,
+      "InsertLocalTransposes": 1.0349645614624023,
+      "InsertOffloadedTransposes": 0.0943443775177002,
+      "LICM": 0.1061861515045166,
+      "LateLegalizeInst": 0.22754216194152832,
+      "LateLegalizePostSplit": 0.09247255325317383,
+      "LateLowerReshapeOp": 0.04053616523742676,
+      "LateLowerTensorOp": 0.3356895446777344,
+      "LateNeuronInstComb": 0.4516925811767578,
+      "LayoutPreprocessing": 0.9441671371459961,
+      "LayoutPreprocessingAndAnalysis": 1.2680203914642334,
+      "LayoutRequirementAnalysis": 0.309098482131958,
+      "LegalizeCCOpLayout": 0.07318258285522461,
+      "LegalizeOpLevelAlias": 0.03343796730041504,
+      "LegalizePartitionReduce": 0.034781694412231445,
+      "LegalizeSundaAccess": 1.4558701515197754,
+      "LegalizeSundaMacro": 0.37755250930786133,
+      "LegalizeType": 0.20858454704284668,
+      "LocalLayoutOpt": 0.36218762397766113,
+      "LoopFusion": 0.31240200996398926,
+      "LoopSplitting": 0.013066768646240234,
+      "LowerBroadcast": 0.047890663146972656,
+      "LowerCCOpBlockAxis": 0.23094987869262695,
+      "LowerComplexBroadcast": 0.15572404861450195,
+      "LowerIntrinsics": 1.228858470916748,
+      "LowerTensorOp": 0.4897449016571045,
+      "LowerTranspose": 0.3995330333709717,
+      "MacroGeneration": 2.335334062576294,
+      "MaskPropagation": 0.14433836936950684,
+      "MemcastMotion": 0.00013000000035390258,
+      "MemcpyElimination": 3.9867260456085205,
+      "MutateDataType": 0.04344511032104492,
+      "NeuronAliasDependencyInduction": 0.025929927825927734,
+      "NeuronAliasDependencyReset": 0.04254412651062012,
+      "NeuronInstComb": 0.19350981712341309,
+      "NeuronLICM": 0.2897522449493408,
+      "NeuronLoopFusion": 0.4089043140411377,
+      "NeuronLoopInterchange": 0.04476189613342285,
+      "NeuronSimplifier": 0.30055856704711914,
+      "NeuronSimplifyPredicates": 0.18221426010131836,
+      "NeuronValueNumbering": 0.10663247108459473,
+      "OptimizeAliasedCopyChain": 0.01511383056640625,
+      "OptimizeNKIKernels": 0.4606451988220215,
+      "PAGLayoutOpt": 26.32272720336914,
+      "PComputeCutting": 0.302201509475708,
+      "PGLayoutTilingPipeline": 38.88710403442383,
+      "PGTiling": 4.423768043518066,
+      "PadElimination": 0.008622884750366211,
+      "ParAxesAnnotation": 25.272018432617188,
+      "PartialLoopFusion": 0.2368309497833252,
+      "PartialSimdFusion": 0.20722246170043945,
+      "PenguinizeFunctions": 0.00015999999595806003,
+      "PerfectLoopNest": 0.06273055076599121,
+      "PruneFunctions": 0.00016700000560376793,
+      "RecognizeOpIdiom": 0.20455479621887207,
+      "Recompute": 0.00649714469909668,
+      "RelaxPredicates": 0.154876708984375,
+      "Rematerialization": 0.16764259338378906,
+      "RemoveOptimizationBarriers": 0.00014099999680183828,
+      "ReshapeWeights": 0.021569013595581055,
+      "ResolveAccessConflict": 0.24012255668640137,
+      "ResolveComplicatePredicates": 0.05034017562866211,
+      "RewriteReplicationMatmul": 0.04589343070983887,
+      "RewriteWeights": 0.05840659141540527,
+      "SFKVectorizer": 3.1227571964263916,
+      "ScatterMotion": 0.0041600000113248825,
+      "SimpleAllReduceTiling": 0.06594347953796387,
+      "Simplifier": 0.11366057395935059,
+      "SimplifyMacroPredicates": 0.18840670585632324,
+      "SimplifyNeuronTensor": 1.3299446105957031,
+      "SimplifySlice": 0.03386688232421875,
+      "SimplifyTensor": 0.21405529975891113,
+      "SpillPSum": 0.5441117286682129,
+      "SplitAPUnionSets": 0.3313255310058594,
+      "SplitAccGrp": 0.03839588165283203,
+      "StaticProfiler": 0.13296246528625488,
+      "StaticTransposeLocalTensor": 0.21724367141723633,
+      "SundaISel": 1.6302134990692139,
+      "TCTransform": 0.03438615798950195,
+      "TensorInitialization": 0.13414645195007324,
+      "TensorOpSimplifier": 0.27712535858154297,
+      "TensorOpTransform": 0.8646912574768066,
+      "TensorizerLegalizationPass": 0.000155999994603917,
+      "TileCCOps": 0.263721227645874,
+      "TilingProfiler": 0.39296984672546387,
+      "TransformConvOp": 0.06336498260498047,
+      "TritiumFusion": 1.0901517868041992,
+      "ValueNumbering": 0.09328150749206543,
+      "VectorizeDMA": 0.03394460678100586,
+      "VectorizeMatMult": 0.0209348201751709,
+      "VerifySupportedOps": 0.00023200000578071922,
+      "WeightCoalescing": 0.05484199523925781,
+      "ZeroSizeTensorElimination": 0.0004336833953857422,
+      "algsimp": 0.0020280000753700733,
+      "batchnorm_expander": 0.0007249999907799065,
+      "boundary-marker-removal": 0.0004140000091865659,
+      "call-inliner": 0.0002570000069681555,
+      "canonicalize-boundary-marker": 0.00044800000614486635,
+      "collective-stream-id-checker": 7.000000186963007e-05,
+      "comparison-expander": 0.00041700000292621553,
+      "computation-deduplicator": 0.0004440000047907233,
+      "conditional-to-select": 8.70000003487803e-05,
+      "config-lowering": 0.00020700000459328294,
+      "constant_folding": 0.00016900000628083944,
+      "cse": 0.00043799998820759356,
+      "dce": 3.899999865097925e-05,
+      "dynamic-slice-transpose": 0.00015799999528098851,
+      "eliminate-redundant-compare": 0.0001539999939268455,
+      "emit-offloaded-dropout": 0.0002770000137388706,
+      "flatten-call-graph": 0.000299000006634742,
+      "fuse-send-recv": 0.0015030000358819962,
+      "hilo::LegalizeAlias": 0.003281000070273876,
+      "hilo::NeuronInstCombine": 0.0011020000092685223,
+      "hilo::NeuronOpFusion": 0.0003429999924264848,
+      "hilo::ReplaceTokenTypeWithU8Pass": 0.00018600000475998968,
+      "hilo::ScheduleFusion": 3.5000000934815034e-05,
+      "hilo::SixtyFourHack": 0.00020599999697878957,
+      "hilo::VerifyAliasing": 7.000000186963007e-05,
+      "hlo-mac-count": 0.0006559999892488122,
+      "hlo-verifier": 0.006031000055372715,
+      "io-con-pipe-begin": 4.999999873689376e-06,
+      "io-con-pipe-end": 9.999999974752427e-07,
+      "io-layout-normalization": 0.0009500000160187483,
+      "legalize-ccops": 1.700000029813964e-05,
+      "legalize-compare": 0.00036899998667649925,
+      "lower-argminmax-custom-call": 0.00013800000306218863,
+      "map-inline": 0.0006319999811239541,
+      "metadata-naming": 0.0009749999735504389,
+      "mlir::detail::OpToOpPassAdaptor": 0.00022499999613501132,
+      "mlir::hlo::MhloToPyPenguin": 0.025104999542236328,
+      "mlir::mhlo::LowerComplexExtraPass": 0.002770999912172556,
+      "mlir::mhlo::LowerComplexPass": 0.001180000021122396,
+      "native-to-custom-softmax": 0.00041199999395757914,
+      "native-to-custom-softmax-dx": 0.00042600001324899495,
+      "operand_upcaster": 0.0007089999853633344,
+      "post-par-pipe-begin": 9.999999974752427e-07,
+      "post-par-pipe-end": 0.0,
+      "post-partition-simplification": 0.05639899894595146,
+      "pre-hlo-begin": 4.999999873689376e-06,
+      "pre-hlo-end": 9.999999974752427e-07,
+      "replace-minimum-constant": 0.0002209999947808683,
+      "reshape-mover": 7.400000322377309e-05,
+      "simplify-concat": 0.0018210000125691295,
+      "simplify-while-loops": 5.500000042957254e-05,
+      "transform-variadic-reduce": 0.0006440000142902136,
+      "tuple-simplifier": 0.00016700000560376793,
+      "unpack-nested-aws-ntwsr": 0.00035700001171790063,
+      "unroll-while-loop": 1.1000000085914508e-05
+    },
+    "hilo": {
+      "HloMacCount": 3802996736.0,
+      "Traffic": 8267154432.0
+    },
+    "tensorizer": {
+      "DMATilingProfiler::TotalInstructionsAfterTiling": 262321,
+      "StaticProfiler::AifUb": 10.559271812438965,
+      "StaticProfiler::ArithmeticIntensityTensorizer": 11.613152503967285,
+      "StaticProfiler::AverageDmaLength": 6652.8759765625,
+      "StaticProfiler::DDRTransferBytes": 7587185496,
+      "StaticProfiler::InternalTransferBytes": 632323092,
+      "StaticProfiler::LoadExpanded": 1033407,
+      "StaticProfiler::StoreExpanded": 3422,
+      "StaticProfiler::TotalDMAExpanded": 1036829,
+      "StaticProfiler::TotalDynamicInstancesCount": 275548,
+      "StaticProfiler::TotalDynamicInstancesWithMmPackedCount": 274994,
+      "StaticProfiler::TotalLNCComm": 0,
+      "StaticProfiler::TotalLNCCommTransfer": 0,
+      "TilingProfiler::BatchnormInstructionsAfterTiling": 0,
+      "TilingProfiler::DmaInstructionsAfterTiling": 0,
+      "TilingProfiler::GenericInstructionsAfterTiling": 79,
+      "TilingProfiler::MatMultInstructionsAfterTiling": 231408,
+      "TilingProfiler::NumPfTransposes": 398,
+      "TilingProfiler::NumPfTransposesForIo": 37,
+      "TilingProfiler::NumPfTransposesForLocal": 216,
+      "TilingProfiler::NumPfTransposesForNonlocal": 145,
+      "TilingProfiler::PfTransposeInstructions": 19513,
+      "TilingProfiler::PfTransposeInstructionsForIo": 19152,
+      "TilingProfiler::PfTransposeInstructionsForLocal": 216,
+      "TilingProfiler::PfTransposeInstructionsForNonlocal": 145,
+      "TilingProfiler::ReduceInstructionsAfterTiling": 74,
+      "TilingProfiler::SimdInstructionsAfterTiling": 2999,
+      "TilingProfiler::TotalInstructionsAfterTiling": 0,
+      "TransformConvOp::Conv1d_depthwise_bf01_oi01_bf01": 0,
+      "TransformConvOp::Conv2d_dw_fb01_io01_01bf_rep_nhwc_Pcinh": 0,
+      "TransformConvOp::Conv2d_pbp_0f1b_0i1o_01fb_experimental_1": 0,
+      "TransformConvOp::Conv2d_pbp_fb01_io01_01bf_experimental_1": 0,
+      "TransformConvOp::conv2d_column_packing": 0,
+      "TransformConvOp::conv2d_column_packing_1": 0,
+      "TransformConvOp::conv2d_column_packing_io10": 0,
+      "TransformConvOp::conv2d_depthwise_f01b_o01i_bf01": 0
+    }
+  },
+  "all": {
+    "compiletime": {
+      "CanonicalizeConv": 9.999999974752427e-07,
+      "CanonicalizeForTensorizer": 0.0003640000068116933,
+      "Canonicalizer": 0.007044999860227108,
+      "HoistCompute": 4.8000001697801054e-05,
+      "IdentifyCrossPassTensors": 0.00013600000238511711,
+      "MemcastMotion": 0.00013000000035390258,
+      "PenguinizeFunctions": 0.00015999999595806003,
+      "PruneFunctions": 0.00016700000560376793,
+      "RemoveOptimizationBarriers": 0.00014099999680183828,
+      "ScatterMotion": 0.0041600000113248825,
+      "TensorizerLegalizationPass": 0.000155999994603917,
+      "VerifySupportedOps": 0.00023200000578071922,
+      "algsimp": 0.0020280000753700733,
+      "batchnorm_expander": 0.0007249999907799065,
+      "boundary-marker-removal": 0.0004140000091865659,
+      "call-inliner": 0.0002570000069681555,
+      "canonicalize-boundary-marker": 0.00044800000614486635,
+      "collective-stream-id-checker": 7.000000186963007e-05,
+      "comparison-expander": 0.00041700000292621553,
+      "computation-deduplicator": 0.0004440000047907233,
+      "conditional-to-select": 8.70000003487803e-05,
+      "config-lowering": 0.00020700000459328294,
+      "constant_folding": 0.00016900000628083944,
+      "cse": 0.00043799998820759356,
+      "dce": 3.899999865097925e-05,
+      "dynamic-slice-transpose": 0.00015799999528098851,
+      "eliminate-redundant-compare": 0.0001539999939268455,
+      "emit-offloaded-dropout": 0.0002770000137388706,
+      "flatten-call-graph": 0.000299000006634742,
+      "fuse-send-recv": 0.0015030000358819962,
+      "hilo::LegalizeAlias": 0.003281000070273876,
+      "hilo::NeuronInstCombine": 0.0011020000092685223,
+      "hilo::NeuronOpFusion": 0.0003429999924264848,
+      "hilo::ReplaceTokenTypeWithU8Pass": 0.00018600000475998968,
+      "hilo::ScheduleFusion": 3.5000000934815034e-05,
+      "hilo::SixtyFourHack": 0.00020599999697878957,
+      "hilo::VerifyAliasing": 7.000000186963007e-05,
+      "hlo-mac-count": 0.0006559999892488122,
+      "hlo-verifier": 0.006031000055372715,
+      "io-con-pipe-begin": 4.999999873689376e-06,
+      "io-con-pipe-end": 9.999999974752427e-07,
+      "io-layout-normalization": 0.0009500000160187483,
+      "legalize-ccops": 1.700000029813964e-05,
+      "legalize-compare": 0.00036899998667649925,
+      "lower-argminmax-custom-call": 0.00013800000306218863,
+      "map-inline": 0.0006319999811239541,
+      "metadata-naming": 0.0009749999735504389,
+      "mlir::detail::OpToOpPassAdaptor": 0.00022499999613501132,
+      "mlir::hlo::MhloToPyPenguin": 0.025104999542236328,
+      "mlir::mhlo::LowerComplexExtraPass": 0.002770999912172556,
+      "mlir::mhlo::LowerComplexPass": 0.001180000021122396,
+      "native-to-custom-softmax": 0.00041199999395757914,
+      "native-to-custom-softmax-dx": 0.00042600001324899495,
+      "operand_upcaster": 0.0007089999853633344,
+      "post-par-pipe-begin": 9.999999974752427e-07,
+      "post-par-pipe-end": 0.0,
+      "post-partition-simplification": 0.05639899894595146,
+      "pre-hlo-begin": 4.999999873689376e-06,
+      "pre-hlo-end": 9.999999974752427e-07,
+      "replace-minimum-constant": 0.0002209999947808683,
+      "reshape-mover": 7.400000322377309e-05,
+      "simplify-concat": 0.0018210000125691295,
+      "simplify-while-loops": 5.500000042957254e-05,
+      "transform-variadic-reduce": 0.0006440000142902136,
+      "tuple-simplifier": 0.00016700000560376793,
+      "unpack-nested-aws-ntwsr": 0.00035700001171790063,
+      "unroll-while-loop": 1.1000000085914508e-05
+    }
+  },
+  "cumsum": {
+    "compiletime": {
+      "CoalesceCCOp": 0.0008378028869628906,
+      "DMALocalityOpt": 0.0003306865692138672,
+      "DMAProfiler": 0.0007596015930175781,
+      "DataStreaming": 0.0002918243408203125,
+      "DoNothing": 0.00012636184692382813,
+      "ExpandISAMacro": 0.0005497932434082031,
+      "FactorizeBlkDims": 0.0004723072052001953,
+      "InferPSumTensor": 0.000583648681640625,
+      "LateLegalizeInst": 0.00040459632873535156,
+      "LateNeuronInstComb": 0.0004837512969970703,
+      "LegalizeSundaAccess": 0.0015611648559570313,
+      "LegalizeType": 0.00025010108947753906,
+      "LowerBroadcast": 0.0009808540344238281,
+      "LowerIntrinsics": 0.0002262592315673828,
+      "LowerTranspose": 0.00021767616271972656,
+      "NeuronInstComb": 0.0004963874816894531,
+      "NeuronLICM": 0.0006859302520751953,
+      "NeuronSimplifyPredicates": 0.002815723419189453,
+      "NeuronValueNumbering": 0.0004124641418457031,
+      "SFKVectorizer": 0.0027742385864257813,
+      "SimpleAllReduceTiling": 0.000209808349609375,
+      "SimplifyNeuronTensor": 0.00040721893310546875,
+      "SpillPSum": 0.0009286403656005859,
+      "WeightCoalescing": 0.0002105236053466797
+    }
+  },
+  "sg00": {
+    "hilo": {
+      "ArithmeticIntensity": 0.9200255870819092,
+      "HloMacCount": 3802996736.0,
+      "Traffic": 8267154432.0
+    }
+  },
+  "sg0000": {
+    "compiletime": {
+      "AGOrderingAnalysisPass": 1.4457588195800781,
+      "AffinePredicateResolution": 0.05167531967163086,
+      "AliasDependencyElimination": 0.0026276111602783203,
+      "AliasDependencyInduction": 0.44934630393981934,
+      "AliasDependencyReset": 1.2677826881408691,
+      "BFComputeCutting": 0.06423807144165039,
+      "BirCodeGenLoop": 2.421293258666992,
+      "CCOpFusion": 0.41050028800964355,
+      "CanonicalizeDAGForPGTiling": 0.21233797073364258,
+      "CanonicalizeIR": 0.06626629829406738,
+      "CoalesceCCOp": 0.1906270980834961,
+      "CommuteConcat": 0.03319668769836426,
+      "DMALocalityOpt": 0.03487658500671387,
+      "DMAProfiler": 0.08790731430053711,
+      "DMATilingProfiler": 0.07109546661376953,
+      "DataLocalityOpt": 1.910703182220459,
+      "DataStreaming": 0.15360140800476074,
+      "DeConcat": 0.012087583541870117,
+      "DeadCodeElimination": 0.035611867904663086,
+      "DeadStoreElimination": 0.37193870544433594,
+      "DelinearIndices": 0.2894127368927002,
+      "Delinearization": 0.1295926570892334,
+      "DoNothing": 6.914138793945313e-05,
+      "DramToDramTranspose": 1.0679569244384766,
+      "DumpGraphAndMetadata": 0.24142217636108398,
+      "EliminateDivs": 0.17337489128112793,
+      "ExpandBatchNorm": 0.06027984619140625,
+      "ExpandISAMacro": 0.09040713310241699,
+      "FactorizeBlkDims": 0.24898743629455566,
+      "FactorizeThreadAxesInFreeDims": 0.03613853454589844,
+      "FlattenMacroLoop": 0.26774168014526367,
+      "GenericAccessSimplifier": 0.03175926208496094,
+      "InferInitValue": 1.029360294342041,
+      "InferIntrinsicOnCC": 0.34307408332824707,
+      "InferNeuronTensor": 1.7935998439788818,
+      "InferNonlocalTensors": 3.6307339668273926,
+      "InferPSumTensor": 0.977715015411377,
+      "InlineNativeKernels": 0.05374264717102051,
+      "InsertIOTransposes": 1.162278652191162,
+      "InsertLocalTransposes": 1.0349645614624023,
+      "InsertOffloadedTransposes": 0.0943443775177002,
+      "LICM": 0.1061861515045166,
+      "LateLegalizeInst": 0.22713756561279297,
+      "LateLegalizePostSplit": 0.09247255325317383,
+      "LateLowerReshapeOp": 0.04053616523742676,
+      "LateLowerTensorOp": 0.3356895446777344,
+      "LateNeuronInstComb": 0.45120882987976074,
+      "LayoutPreprocessing": 0.9441671371459961,
+      "LayoutPreprocessingAndAnalysis": 1.2680203914642334,
+      "LayoutRequirementAnalysis": 0.309098482131958,
+      "LegalizeCCOpLayout": 0.07318258285522461,
+      "LegalizeOpLevelAlias": 0.03343796730041504,
+      "LegalizePartitionReduce": 0.034781694412231445,
+      "LegalizeSundaAccess": 1.4543089866638184,
+      "LegalizeSundaMacro": 0.37755250930786133,
+      "LegalizeType": 0.20833444595336914,
+      "LocalLayoutOpt": 0.36218762397766113,
+      "LoopFusion": 0.31240200996398926,
+      "LoopSplitting": 0.013066768646240234,
+      "LowerBroadcast": 0.04690980911254883,
+      "LowerCCOpBlockAxis": 0.23094987869262695,
+      "LowerComplexBroadcast": 0.15572404861450195,
+      "LowerIntrinsics": 1.2286322116851807,
+      "LowerTensorOp": 0.4897449016571045,
+      "LowerTranspose": 0.39931535720825195,
+      "MacroGeneration": 2.335334062576294,
+      "MaskPropagation": 0.14433836936950684,
+      "MemcpyElimination": 3.9867260456085205,
+      "MutateDataType": 0.04344511032104492,
+      "NeuronAliasDependencyInduction": 0.025929927825927734,
+      "NeuronAliasDependencyReset": 0.04254412651062012,
+      "NeuronInstComb": 0.19301342964172363,
+      "NeuronLICM": 0.2890663146972656,
+      "NeuronLoopFusion": 0.4089043140411377,
+      "NeuronLoopInterchange": 0.04476189613342285,
+      "NeuronSimplifier": 0.30055856704711914,
+      "NeuronSimplifyPredicates": 0.1793985366821289,
+      "NeuronValueNumbering": 0.10622000694274902,
+      "OptimizeAliasedCopyChain": 0.01511383056640625,
+      "OptimizeNKIKernels": 0.4606451988220215,
+      "PAGLayoutOpt": 26.32272720336914,
+      "PComputeCutting": 0.302201509475708,
+      "PGLayoutTilingPipeline": 38.88710403442383,
+      "PGTiling": 4.423768043518066,
+      "PadElimination": 0.008622884750366211,
+      "ParAxesAnnotation": 25.272018432617188,
+      "PartialLoopFusion": 0.2368309497833252,
+      "PartialSimdFusion": 0.20722246170043945,
+      "PerfectLoopNest": 0.06273055076599121,
+      "RecognizeOpIdiom": 0.20455479621887207,
+      "Recompute": 0.00649714469909668,
+      "RelaxPredicates": 0.154876708984375,
+      "Rematerialization": 0.16764259338378906,
+      "ReshapeWeights": 0.021569013595581055,
+      "ResolveAccessConflict": 0.24012255668640137,
+      "ResolveComplicatePredicates": 0.05034017562866211,
+      "RewriteReplicationMatmul": 0.04589343070983887,
+      "RewriteWeights": 0.05840659141540527,
+      "SFKVectorizer": 3.119982957839966,
+      "SimpleAllReduceTiling": 0.06573367118835449,
+      "Simplifier": 0.11366057395935059,
+      "SimplifyMacroPredicates": 0.18840670585632324,
+      "SimplifyNeuronTensor": 1.3295373916625977,
+      "SimplifySlice": 0.03386688232421875,
+      "SimplifyTensor": 0.21405529975891113,
+      "SpillPSum": 0.5431830883026123,
+      "SplitAPUnionSets": 0.3313255310058594,
+      "SplitAccGrp": 0.03839588165283203,
+      "StaticProfiler": 0.13296246528625488,
+      "StaticTransposeLocalTensor": 0.21724367141723633,
+      "SundaISel": 1.6302134990692139,
+      "TCTransform": 0.03438615798950195,
+      "TensorInitialization": 0.13414645195007324,
+      "TensorOpSimplifier": 0.27712535858154297,
+      "TensorOpTransform": 0.8646912574768066,
+      "TileCCOps": 0.263721227645874,
+      "TilingProfiler": 0.39296984672546387,
+      "TransformConvOp": 0.06336498260498047,
+      "TritiumFusion": 1.0901517868041992,
+      "ValueNumbering": 0.09328150749206543,
+      "VectorizeDMA": 0.03394460678100586,
+      "VectorizeMatMult": 0.0209348201751709,
+      "WeightCoalescing": 0.05463147163391113,
+      "ZeroSizeTensorElimination": 0.0004336833953857422
+    },
+    "tensorizer": {
+      "DMATilingProfiler::TotalInstructionsAfterTiling": 262321,
+      "StaticProfiler::AifUb": 10.559271812438965,
+      "StaticProfiler::ArithmeticIntensityTensorizer": 11.613152503967285,
+      "StaticProfiler::AverageDmaLength": 6652.8759765625,
+      "StaticProfiler::AverageFractalPeUtilization": 99.8321762084961,
+      "StaticProfiler::AveragePartitionUtilization": 99.3888168334961,
+      "StaticProfiler::AveragePeUtilization": 99.65400695800781,
+      "StaticProfiler::DDRTransferBytes": 7587185496,
+      "StaticProfiler::InternalTransferBytes": 632323092,
+      "StaticProfiler::LoadExpanded": 1033407,
+      "StaticProfiler::LocalizationEfficiency": 109.9806137084961,
+      "StaticProfiler::LocalizationEfficiencyIgnoreNonlocal": 110.06793212890625,
+      "StaticProfiler::StoreExpanded": 3422,
+      "StaticProfiler::TotalDMAExpanded": 1036829,
+      "StaticProfiler::TotalDynamicInstancesCount": 275548,
+      "StaticProfiler::TotalDynamicInstancesWithMmPackedCount": 274994,
+      "StaticProfiler::TotalLNCComm": 0,
+      "StaticProfiler::TotalLNCCommTransfer": 0,
+      "TilingProfiler::AveragePartitionUtilizationAfterTiling": 0,
+      "TilingProfiler::AveragePeUtilizationAfterTiling": 0,
+      "TilingProfiler::BatchnormInstructionsAfterTiling": 0,
+      "TilingProfiler::DmaInstructionsAfterTiling": 0,
+      "TilingProfiler::GenericInstructionsAfterTiling": 79,
+      "TilingProfiler::MatMultInstructionsAfterTiling": 231408,
+      "TilingProfiler::NumPfTransposes": 398,
+      "TilingProfiler::NumPfTransposesForIo": 37,
+      "TilingProfiler::NumPfTransposesForLocal": 216,
+      "TilingProfiler::NumPfTransposesForNonlocal": 145,
+      "TilingProfiler::PfTransposeInstructions": 19513,
+      "TilingProfiler::PfTransposeInstructionsForIo": 19152,
+      "TilingProfiler::PfTransposeInstructionsForLocal": 216,
+      "TilingProfiler::PfTransposeInstructionsForNonlocal": 145,
+      "TilingProfiler::ReduceInstructionsAfterTiling": 74,
+      "TilingProfiler::SimdInstructionsAfterTiling": 2999,
+      "TilingProfiler::TotalInstructionsAfterTiling": 0,
+      "TransformConvOp::Conv1d_depthwise_bf01_oi01_bf01": 0,
+      "TransformConvOp::Conv2d_dw_fb01_io01_01bf_rep_nhwc_Pcinh": 0,
+      "TransformConvOp::Conv2d_pbp_0f1b_0i1o_01fb_experimental_1": 0,
+      "TransformConvOp::Conv2d_pbp_fb01_io01_01bf_experimental_1": 0,
+      "TransformConvOp::conv2d_column_packing": 0,
+      "TransformConvOp::conv2d_column_packing_1": 0,
+      "TransformConvOp::conv2d_column_packing_io10": 0,
+      "TransformConvOp::conv2d_depthwise_f01b_o01i_bf01": 0
+    }
+  }
+}

token_generation_model/_tp0_bk0/graph.neff

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:82be447a0a308a6e83990d1f3d193b4dc43ab835b136e7c27647ecf6cde94383
+size 6001664

token_generation_model/_tp0_bk0/metaneff.pb

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:8f42b279a662fc21e6bb94ab8bdb96ad553535cec385b6c8909a4e7622fad939
+size 985283