Add files using upload-large-folder tool

README.md

@@ -0,0 +1,200 @@
+---
+library_name: vllm
+license: apache-2.0
+base_model: mistralai/Mistral-7B-v0.3
+extra_gated_description: >-
+  If you want to learn more about how we process your personal data, please read
+  our <a href="https://mistral.ai/terms/">Privacy Policy</a>.
+tags:
+- vllm
+- mistral-common
+---
+
+# Model Card for Mistral-7B-Instruct-v0.3
+
+The Mistral-7B-Instruct-v0.3 Large Language Model (LLM) is an instruct fine-tuned version of the Mistral-7B-v0.3.
+
+Mistral-7B-v0.3 has the following changes compared to [Mistral-7B-v0.2](https://huggingface.co/mistralai/Mistral-7B-Instruct-v0.2/edit/main/README.md)
+- Extended vocabulary to 32768
+- Supports v3 Tokenizer
+- Supports function calling
+
+## Installation
+
+It is recommended to use `mistralai/Mistral-7B-Instruct-v0.3` with [mistral-inference](https://github.com/mistralai/mistral-inference). For HF transformers code snippets, please keep scrolling.
+
+```
+pip install mistral_inference
+```
+
+## Download
+
+```py
+from huggingface_hub import snapshot_download
+from pathlib import Path
+
+mistral_models_path = Path.home().joinpath('mistral_models', '7B-Instruct-v0.3')
+mistral_models_path.mkdir(parents=True, exist_ok=True)
+
+snapshot_download(repo_id="mistralai/Mistral-7B-Instruct-v0.3", allow_patterns=["params.json", "consolidated.safetensors", "tokenizer.model.v3"], local_dir=mistral_models_path)
+```
+
+### Chat
+
+After installing `mistral_inference`, a `mistral-chat` CLI command should be available in your environment. You can chat with the model using
+
+```
+mistral-chat $HOME/mistral_models/7B-Instruct-v0.3 --instruct --max_tokens 256
+```
+
+### Instruct following
+
+```py
+from mistral_inference.transformer import Transformer
+from mistral_inference.generate import generate
+
+from mistral_common.tokens.tokenizers.mistral import MistralTokenizer
+from mistral_common.protocol.instruct.messages import UserMessage
+from mistral_common.protocol.instruct.request import ChatCompletionRequest
+
+
+tokenizer = MistralTokenizer.from_file(f"{mistral_models_path}/tokenizer.model.v3")
+model = Transformer.from_folder(mistral_models_path)
+
+completion_request = ChatCompletionRequest(messages=[UserMessage(content="Explain Machine Learning to me in a nutshell.")])
+
+tokens = tokenizer.encode_chat_completion(completion_request).tokens
+
+out_tokens, _ = generate([tokens], model, max_tokens=64, temperature=0.0, eos_id=tokenizer.instruct_tokenizer.tokenizer.eos_id)
+result = tokenizer.instruct_tokenizer.tokenizer.decode(out_tokens[0])
+
+print(result)
+```
+
+### Function calling
+
+```py
+from mistral_common.protocol.instruct.tool_calls import Function, Tool
+from mistral_inference.transformer import Transformer
+from mistral_inference.generate import generate
+
+from mistral_common.tokens.tokenizers.mistral import MistralTokenizer
+from mistral_common.protocol.instruct.messages import UserMessage
+from mistral_common.protocol.instruct.request import ChatCompletionRequest
+
+
+tokenizer = MistralTokenizer.from_file(f"{mistral_models_path}/tokenizer.model.v3")
+model = Transformer.from_folder(mistral_models_path)
+
+completion_request = ChatCompletionRequest(
+    tools=[
+        Tool(
+            function=Function(
+                name="get_current_weather",
+                description="Get the current weather",
+                parameters={
+                    "type": "object",
+                    "properties": {
+                        "location": {
+                            "type": "string",
+                            "description": "The city and state, e.g. San Francisco, CA",
+                        },
+                        "format": {
+                            "type": "string",
+                            "enum": ["celsius", "fahrenheit"],
+                            "description": "The temperature unit to use. Infer this from the users location.",
+                        },
+                    },
+                    "required": ["location", "format"],
+                },
+            )
+        )
+    ],
+    messages=[
+        UserMessage(content="What's the weather like today in Paris?"),
+        ],
+)
+
+tokens = tokenizer.encode_chat_completion(completion_request).tokens
+
+out_tokens, _ = generate([tokens], model, max_tokens=64, temperature=0.0, eos_id=tokenizer.instruct_tokenizer.tokenizer.eos_id)
+result = tokenizer.instruct_tokenizer.tokenizer.decode(out_tokens[0])
+
+print(result)
+```
+
+## Generate with `transformers`
+
+If you want to use Hugging Face `transformers` to generate text, you can do something like this.
+
+```py
+from transformers import pipeline
+
+messages = [
+    {"role": "system", "content": "You are a pirate chatbot who always responds in pirate speak!"},
+    {"role": "user", "content": "Who are you?"},
+]
+chatbot = pipeline("text-generation", model="mistralai/Mistral-7B-Instruct-v0.3")
+chatbot(messages)
+```
+
+
+## Function calling with `transformers`
+
+To use this example, you'll need `transformers` version 4.42.0 or higher. Please see the 
+[function calling guide](https://huggingface.co/docs/transformers/main/chat_templating#advanced-tool-use--function-calling)
+in the `transformers` docs for more information.
+
+```python
+from transformers import AutoModelForCausalLM, AutoTokenizer
+import torch
+
+model_id = "mistralai/Mistral-7B-Instruct-v0.3"
+tokenizer = AutoTokenizer.from_pretrained(model_id)
+
+def get_current_weather(location: str, format: str):
+    """
+    Get the current weather
+
+    Args:
+        location: The city and state, e.g. San Francisco, CA
+        format: The temperature unit to use. Infer this from the users location. (choices: ["celsius", "fahrenheit"])
+    """
+    pass
+
+conversation = [{"role": "user", "content": "What's the weather like in Paris?"}]
+tools = [get_current_weather]
+
+
+# format and tokenize the tool use prompt 
+inputs = tokenizer.apply_chat_template(
+            conversation,
+            tools=tools,
+            add_generation_prompt=True,
+            return_dict=True,
+            return_tensors="pt",
+)
+
+model = AutoModelForCausalLM.from_pretrained(model_id, torch_dtype=torch.bfloat16, device_map="auto")
+
+inputs.to(model.device)
+outputs = model.generate(**inputs, max_new_tokens=1000)
+print(tokenizer.decode(outputs[0], skip_special_tokens=True))
+```
+
+Note that, for reasons of space, this example does not show a complete cycle of calling a tool and adding the tool call and tool
+results to the chat history so that the model can use them in its next generation. For a full tool calling example, please
+see the [function calling guide](https://huggingface.co/docs/transformers/main/chat_templating#advanced-tool-use--function-calling), 
+and note that Mistral **does** use tool call IDs, so these must be included in your tool calls and tool results. They should be
+exactly 9 alphanumeric characters.
+
+
+## Limitations
+
+The Mistral 7B Instruct model is a quick demonstration that the base model can be easily fine-tuned to achieve compelling performance. 
+It does not have any moderation mechanisms. We're looking forward to engaging with the community on ways to
+make the model finely respect guardrails, allowing for deployment in environments requiring moderated outputs.
+
+## The Mistral AI Team
+
+Albert Jiang, Alexandre Sablayrolles, Alexis Tacnet, Antoine Roux, Arthur Mensch, Audrey Herblin-Stoop, Baptiste Bout, Baudouin de Monicault, Blanche Savary, Bam4d, Caroline Feldman, Devendra Singh Chaplot, Diego de las Casas, Eleonore Arcelin, Emma Bou Hanna, Etienne Metzger, Gianna Lengyel, Guillaume Bour, Guillaume Lample, Harizo Rajaona, Jean-Malo Delignon, Jia Li, Justus Murke, Louis Martin, Louis Ternon, Lucile Saulnier, Lélio Renard Lavaud, Margaret Jennings, Marie Pellat, Marie Torelli, Marie-Anne Lachaux, Nicolas Schuhl, Patrick von Platen, Pierre Stock, Sandeep Subramanian, Sophia Yang, Szymon Antoniak, Teven Le Scao, Thibaut Lavril, Timothée Lacroix, Théophile Gervet, Thomas Wang, Valera Nemychnikova, William El Sayed, William Marshall

context_encoding_model/_tp0_bk0/command.txt

@@ -0,0 +1 @@
+neuronx-cc compile --framework=XLA model.MODULE_c6824be80aab0b095843+cc19d8a1.hlo_module.pb --output model.MODULE_c6824be80aab0b095843+cc19d8a1.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_c6824be80aab0b095843+cc19d8a1.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=/models/mistral-7b-v0.3-instruct-neuronx/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": 99.61843872070313,
+      "StaticProfiler::AveragePartitionUtilization": 98.97808837890625,
+      "StaticProfiler::AveragePeUtilization": 98.48223876953125,
+      "StaticProfiler::LocalizationEfficiency": 97.97124481201172,
+      "StaticProfiler::LocalizationEfficiencyIgnoreNonlocal": 100.67024993896484,
+      "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": 0.04956221580505371,
+      "AffinePredicateResolution": 0.002319812774658203,
+      "AliasDependencyElimination": 0.00019097328186035156,
+      "AliasDependencyInduction": 0.01746201515197754,
+      "AliasDependencyReset": 0.04065680503845215,
+      "BFComputeCutting": 0.0013394355773925781,
+      "BirCodeGenLoop": 0.11406064033508301,
+      "CCOpFusion": 0.03461766242980957,
+      "CanonicalizeConv": 4.900000203633681e-05,
+      "CanonicalizeDAGForPGTiling": 0.004343748092651367,
+      "CanonicalizeForTensorizer": 4.600000102072954e-05,
+      "CanonicalizeIR": 0.003444671630859375,
+      "Canonicalizer": 0.0009449999779462814,
+      "CoalesceCCOp": 0.04709577560424805,
+      "CommuteConcat": 0.0010838508605957031,
+      "DMALocalityOpt": 0.002311229705810547,
+      "DMAProfiler": 0.006924152374267578,
+      "DMATilingProfiler": 0.004760265350341797,
+      "DataLocalityOpt": 0.11827898025512695,
+      "DataStreaming": 0.006251335144042969,
+      "DeConcat": 0.0006871223449707031,
+      "DeadCodeElimination": 0.0018725395202636719,
+      "DeadStoreElimination": 0.00615692138671875,
+      "DelinearIndices": 0.004624128341674805,
+      "Delinearization": 0.0023467540740966797,
+      "DoNothing": 0.0003032684326171875,
+      "DramToDramTranspose": 0.015562057495117188,
+      "DumpGraphAndMetadata": 0.021119117736816406,
+      "EliminateDivs": 0.0027742385864257813,
+      "ExpandBatchNorm": 0.0035254955291748047,
+      "ExpandISAMacro": 0.005174160003662109,
+      "FactorizeBlkDims": 0.00865316390991211,
+      "FactorizeThreadAxesInFreeDims": 0.0011937618255615234,
+      "FlattenMacroLoop": 0.002167940139770508,
+      "GenericAccessSimplifier": 0.0008442401885986328,
+      "HoistCompute": 7.000000096013537e-06,
+      "IdentifyCrossPassTensors": 3.899999865097925e-05,
+      "InferInitValue": 0.024956941604614258,
+      "InferIntrinsicOnCC": 0.015187978744506836,
+      "InferNeuronTensor": 0.01759958267211914,
+      "InferNonlocalTensors": 0.014631509780883789,
+      "InferPSumTensor": 0.03635692596435547,
+      "InlineNativeKernels": 0.0049402713775634766,
+      "InsertIOTransposes": 0.012220144271850586,
+      "InsertLocalTransposes": 0.0043489933013916016,
+      "InsertOffloadedTransposes": 0.0029261112213134766,
+      "LICM": 0.002863168716430664,
+      "LateLegalizeInst": 0.00691986083984375,
+      "LateLegalizePostSplit": 0.004354715347290039,
+      "LateLowerReshapeOp": 0.0013675689697265625,
+      "LateLowerTensorOp": 0.0017099380493164063,
+      "LateNeuronInstComb": 0.008911371231079102,
+      "LayoutPreprocessing": 0.024985551834106445,
+      "LayoutPreprocessingAndAnalysis": 0.05915069580078125,
+      "LayoutRequirementAnalysis": 0.004924297332763672,
+      "LegalizeCCOpLayout": 0.003811359405517578,
+      "LegalizeOpLevelAlias": 0.0012297630310058594,
+      "LegalizePartitionReduce": 0.001873016357421875,
+      "LegalizeSundaAccess": 0.06697678565979004,
+      "LegalizeSundaMacro": 0.010838985443115234,
+      "LegalizeType": 0.007873058319091797,
+      "LocalLayoutOpt": 0.01294255256652832,
+      "LoopFusion": 0.005301713943481445,
+      "LoopSplitting": 0.0006639957427978516,
+      "LowerBroadcast": 0.00147247314453125,
+      "LowerCCOpBlockAxis": 0.003854513168334961,
+      "LowerComplexBroadcast": 0.0023345947265625,
+      "LowerIntrinsics": 0.31359052658081055,
+      "LowerTensorOp": 0.013880491256713867,
+      "LowerTranspose": 0.007946014404296875,
+      "MacroGeneration": 0.04993605613708496,
+      "MaskPropagation": 0.0027527809143066406,
+      "MemcastMotion": 3.600000127335079e-05,
+      "MemcpyElimination": 0.02739572525024414,
+      "MutateDataType": 0.0013971328735351563,
+      "NeuronAliasDependencyInduction": 0.0002524852752685547,
+      "NeuronAliasDependencyReset": 0.046864986419677734,
+      "NeuronInstComb": 0.007117271423339844,
+      "NeuronLICM": 0.035208702087402344,
+      "NeuronLoopFusion": 0.008702993392944336,
+      "NeuronLoopInterchange": 0.0014033317565917969,
+      "NeuronSimplifier": 0.0068361759185791016,
+      "NeuronSimplifyPredicates": 0.006758928298950195,
+      "NeuronValueNumbering": 0.003446817398071289,
+      "OptimizeAliasedCopyChain": 0.00058746337890625,
+      "OptimizeNKIKernels": 0.5256326198577881,
+      "PAGLayoutOpt": 0.07949376106262207,
+      "PComputeCutting": 0.005449533462524414,
+      "PGLayoutTilingPipeline": 0.5191597938537598,
+      "PGTiling": 0.1754302978515625,
+      "PadElimination": 0.0005822181701660156,
+      "ParAxesAnnotation": 0.0432279109954834,
+      "PartialLoopFusion": 0.0034580230712890625,
+      "PartialSimdFusion": 0.003893136978149414,
+      "PenguinizeFunctions": 4.099999932805076e-05,
+      "PerfectLoopNest": 0.0017859935760498047,
+      "PruneFunctions": 3.300000025774352e-05,
+      "RecognizeOpIdiom": 0.003717660903930664,
+      "Recompute": 0.0006694793701171875,
+      "RelaxPredicates": 0.004067182540893555,
+      "Rematerialization": 0.002377748489379883,
+      "RemoveOptimizationBarriers": 4.199999966658652e-05,
+      "ReshapeWeights": 0.0010006427764892578,
+      "ResolveAccessConflict": 0.003880739212036133,
+      "ResolveComplicatePredicates": 0.0018019676208496094,
+      "RewriteReplicationMatmul": 0.0015823841094970703,
+      "RewriteWeights": 0.002522706985473633,
+      "SFKVectorizer": 0.12057852745056152,
+      "ScatterMotion": 3.999999989900971e-06,
+      "SimpleAllReduceTiling": 0.003117799758911133,
+      "Simplifier": 0.0035266876220703125,
+      "SimplifyMacroPredicates": 0.005109071731567383,
+      "SimplifyNeuronTensor": 0.2755560874938965,
+      "SimplifySlice": 0.0010409355163574219,
+      "SimplifyTensor": 0.025161027908325195,
+      "SpillPSum": 0.05051732063293457,
+      "SplitAPUnionSets": 0.013710737228393555,
+      "SplitAccGrp": 0.001070261001586914,
+      "StaticProfiler": 0.005475759506225586,
+      "StaticTransposeLocalTensor": 0.0038781166076660156,
+      "SundaISel": 0.04041314125061035,
+      "TCTransform": 0.0008709430694580078,
+      "TensorInitialization": 0.004585742950439453,
+      "TensorOpSimplifier": 0.005696535110473633,
+      "TensorOpTransform": 0.021961212158203125,
+      "TensorizerLegalizationPass": 5.199999577598646e-05,
+      "TileCCOps": 0.010190010070800781,
+      "TilingProfiler": 0.04765009880065918,
+      "TransformConvOp": 0.0030345916748046875,
+      "TritiumFusion": 0.07207250595092773,
+      "ValueNumbering": 0.002035379409790039,
+      "VectorizeDMA": 0.006017446517944336,
+      "VectorizeMatMult": 0.0027844905853271484,
+      "VerifySupportedOps": 3.199999991920777e-05,
+      "WeightCoalescing": 0.0033817291259765625,
+      "ZeroSizeTensorElimination": 0.00019311904907226563,
+      "algsimp": 0.002274000085890293,
+      "batchnorm_expander": 4.3000000005122274e-05,
+      "boundary-marker-removal": 1.2000000424450263e-05,
+      "call-inliner": 0.00041299997246824205,
+      "canonicalize-boundary-marker": 1.3999999282532372e-05,
+      "collective-stream-id-checker": 0.00012900000729132444,
+      "comparison-expander": 0.0003989999822806567,
+      "computation-deduplicator": 5.999999848427251e-05,
+      "conditional-to-select": 1.8999999156221747e-05,
+      "config-lowering": 0.00018099999579135329,
+      "constant-statistics": 0.00039400000241585076,
+      "constant_folding": 0.0002840000088326633,
+      "cse": 5.699999746866524e-05,
+      "dce": 7.900000491645187e-05,
+      "dot_decomposer": 0.0011419999646022916,
+      "dynamic-slice-transpose": 1.4999999621068127e-05,
+      "eliminate-redundant-compare": 0.0002559999702498317,
+      "emit-offloaded-dropout": 6.399999983841553e-05,
+      "flatten-call-graph": 0.0006720000528730452,
+      "fuse-send-recv": 6.70000008540228e-05,
+      "hilo::LegalizeAlias": 1.2999999853491317e-05,
+      "hilo::NeuronInstCombine": 0.00018000000272877514,
+      "hilo::NeuronOpFusion": 7.100000220816582e-05,
+      "hilo::ReplaceTokenTypeWithU8Pass": 1.8000000636675395e-05,
+      "hilo::ScheduleFusion": 1.2000000424450263e-05,
+      "hilo::SixtyFourHack": 6.299999949987978e-05,
+      "hilo::VerifyAliasing": 6.000000212225132e-06,
+      "hlo-mac-count": 0.0011450001038610935,
+      "hlo-verifier": 0.00938900001347065,
+      "instruction-histogram": 0.001053999993018806,
+      "io-con-pipe-begin": 9.000000318337698e-06,
+      "io-con-pipe-end": 9.999999974752427e-07,
+      "io-layout-normalization": 0.0016039999900385737,
+      "io-statistics": 6.900000153109431e-05,
+      "legalize-ccops": 3.999999989900971e-06,
+      "legalize-compare": 1.1000000085914508e-05,
+      "lower-argminmax-custom-call": 1.2000000424450263e-05,
+      "map-inline": 0.0006389999762177467,
+      "metadata-naming": 6.299999949987978e-05,
+      "mlir::detail::OpToOpPassAdaptor": 9.600000339560211e-05,
+      "mlir::hlo::MhloToPyPenguin": 0.07229000329971313,
+      "mlir::mhlo::LowerComplexExtraPass": 0.0003330000035930425,
+      "mlir::mhlo::LowerComplexPass": 0.0003640000068116933,
+      "native-to-custom-softmax": 0.0007670000195503235,
+      "native-to-custom-softmax-dx": 0.00069500005338341,
+      "operand_upcaster": 5.0000002374872565e-05,
+      "opt-barrier-removal": 0.0005799999926239252,
+      "post-par-pipe-begin": 3.000000106112566e-06,
+      "post-par-pipe-end": 0.0,
+      "post-partition-simplification": 0.0016440000617876649,
+      "pre-par-pipe-begin": 1.9999999949504854e-06,
+      "pre-par-pipe-end": 0.0,
+      "pre-partition-simplification": 0.2085230052471161,
+      "replace-minimum-constant": 0.0003719999804161489,
+      "reshape-mover": 0.0001140000022132881,
+      "simplify-concat": 0.00015900000289548188,
+      "simplify-while-loops": 9.700000373413786e-05,
+      "transform-variadic-reduce": 6.900000153109431e-05,
+      "tuple-simplifier": 0.000295000005280599,
+      "unpack-nested-aws-ntwsr": 0.00043299997923895717,
+      "unroll-while-loop": 1.700000029813964e-05,
+      "zero_sized_hlo_elimination": 0.0006750000175088644
+    },
+    "hilo": {
+      "ConstantSize": 271333.0,
+      "HloInputCount": 359.0,
+      "HloMacCount": 28118614016.0,
+      "HloOutputCount": 65.0,
+      "IfmapSize": 7785161728.0,
+      "OfmapSize": 536870912.0,
+      "OutputsReadFromCount": 0.0,
+      "PassthroughTensorsCount": 0.0,
+      "RedundantOutputCount": 0.0,
+      "Traffic": 752578432.0
+    },
+    "tensorizer": {
+      "DMATilingProfiler::TotalInstructionsAfterTiling": 9240.0,
+      "StaticProfiler::AifUb": 74.75788116455078,
+      "StaticProfiler::ArithmeticIntensityTensorizer": 73.24122619628906,
+      "StaticProfiler::AverageDmaLength": 6860.4052734375,
+      "StaticProfiler::DDRTransferBytes": 318999072.0,
+      "StaticProfiler::InternalTransferBytes": 6136852.0,
+      "StaticProfiler::LoadExpanded": 40210.0,
+      "StaticProfiler::StoreExpanded": 1195.0,
+      "StaticProfiler::TotalDMAExpanded": 41405.0,
+      "StaticProfiler::TotalDynamicInstancesCount": 9321.0,
+      "StaticProfiler::TotalDynamicInstancesWithMmPackedCount": 9225.0,
+      "StaticProfiler::TotalLNCComm": 0.0,
+      "StaticProfiler::TotalLNCCommTransfer": 0.0,
+      "TilingProfiler::BatchnormInstructionsAfterTiling": 0.0,
+      "TilingProfiler::DmaInstructionsAfterTiling": 0.0,
+      "TilingProfiler::GenericInstructionsAfterTiling": 4.0,
+      "TilingProfiler::MatMultInstructionsAfterTiling": 8128.0,
+      "TilingProfiler::NumPfTransposes": 4.0,
+      "TilingProfiler::NumPfTransposesForIo": 0.0,
+      "TilingProfiler::NumPfTransposesForLocal": 1.0,
+      "TilingProfiler::NumPfTransposesForNonlocal": 3.0,
+      "TilingProfiler::PfTransposeInstructions": 97.0,
+      "TilingProfiler::PfTransposeInstructionsForIo": 0.0,
+      "TilingProfiler::PfTransposeInstructionsForLocal": 1.0,
+      "TilingProfiler::PfTransposeInstructionsForNonlocal": 96.0,
+      "TilingProfiler::ReduceInstructionsAfterTiling": 4.0,
+      "TilingProfiler::SimdInstructionsAfterTiling": 166.0,
+      "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": 0.0020540000405162573,
+      "call-inliner": 0.0003819999983534217,
+      "collective-stream-id-checker": 0.00010900000052060932,
+      "comparison-expander": 0.00038499999209307134,
+      "constant-statistics": 0.00039400000241585076,
+      "constant_folding": 0.0002589999930933118,
+      "dce": 7.500000356230885e-05,
+      "dot_decomposer": 0.0011419999646022916,
+      "eliminate-redundant-compare": 0.00024399999529123306,
+      "flatten-call-graph": 0.0006419999990612268,
+      "hlo-mac-count": 0.0008709999965503812,
+      "hlo-verifier": 0.008843000046908855,
+      "instruction-histogram": 0.001053999993018806,
+      "io-con-pipe-begin": 9.000000318337698e-06,
+      "io-con-pipe-end": 9.999999974752427e-07,
+      "io-layout-normalization": 0.0016039999900385737,
+      "io-statistics": 6.900000153109431e-05,
+      "map-inline": 0.0006040000007487833,
+      "native-to-custom-softmax": 0.0007200000109151006,
+      "native-to-custom-softmax-dx": 0.0005460000247694552,
+      "opt-barrier-removal": 0.0005799999926239252,
+      "pre-par-pipe-begin": 1.9999999949504854e-06,
+      "pre-par-pipe-end": 0.0,
+      "pre-partition-simplification": 0.2085230052471161,
+      "replace-minimum-constant": 0.0003429999924264848,
+      "reshape-mover": 0.00010299999848939478,
+      "simplify-while-loops": 9.100000170292333e-05,
+      "tuple-simplifier": 0.0002800000074785203,
+      "unpack-nested-aws-ntwsr": 0.00042299999040551484,
+      "unroll-while-loop": 1.700000029813964e-05,
+      "zero_sized_hlo_elimination": 0.0006750000175088644
+    }
+  },
+  "cumsum": {
+    "compiletime": {
+      "CoalesceCCOp": 0.0002715587615966797,
+      "DMALocalityOpt": 0.00022292137145996094,
+      "DMAProfiler": 0.001041412353515625,
+      "DataStreaming": 0.0003218650817871094,
+      "DoNothing": 0.0001633167266845703,
+      "ExpandISAMacro": 0.0005700588226318359,
+      "FactorizeBlkDims": 0.00047659873962402344,
+      "InferPSumTensor": 0.0006239414215087891,
+      "LateLegalizeInst": 0.00047588348388671875,
+      "LateNeuronInstComb": 0.0006630420684814453,
+      "LegalizeSundaAccess": 0.0018274784088134766,
+      "LegalizeType": 0.0003712177276611328,
+      "LowerBroadcast": 0.0002789497375488281,
+      "LowerIntrinsics": 0.00029349327087402344,
+      "LowerTranspose": 0.00027632713317871094,
+      "NeuronInstComb": 0.0005421638488769531,
+      "NeuronLICM": 0.0004737377166748047,
+      "NeuronSimplifyPredicates": 0.0035429000854492188,
+      "NeuronValueNumbering": 0.000461578369140625,
+      "SFKVectorizer": 0.0029914379119873047,
+      "SimpleAllReduceTiling": 0.0002620220184326172,
+      "SimplifyNeuronTensor": 0.0004684925079345703,
+      "SpillPSum": 0.0006010532379150391,
+      "WeightCoalescing": 0.00029349327087402344
+    }
+  },
+  "sg00": {
+    "compiletime": {
+      "CanonicalizeConv": 3.7000001611886546e-05,
+      "CanonicalizeForTensorizer": 1.9999999494757503e-05,
+      "Canonicalizer": 0.000371000001905486,
+      "HoistCompute": 3.999999989900971e-06,
+      "IdentifyCrossPassTensors": 1.5999999959603883e-05,
+      "MemcastMotion": 1.4000000192027073e-05,
+      "PenguinizeFunctions": 1.8999999156221747e-05,
+      "PruneFunctions": 1.2999999853491317e-05,
+      "RemoveOptimizationBarriers": 1.4000000192027073e-05,
+      "ScatterMotion": 0.0,
+      "TensorizerLegalizationPass": 2.9999999242136255e-05,
+      "VerifySupportedOps": 1.1000000085914508e-05,
+      "algsimp": 7.000000186963007e-05,
+      "batchnorm_expander": 1.2999999853491317e-05,
+      "boundary-marker-removal": 3.000000106112566e-06,
+      "call-inliner": 9.999999747378752e-06,
+      "canonicalize-boundary-marker": 4.999999873689376e-06,
+      "collective-stream-id-checker": 3.000000106112566e-06,
+      "comparison-expander": 3.999999989900971e-06,
+      "computation-deduplicator": 1.4999999621068127e-05,
+      "conditional-to-select": 4.999999873689376e-06,
+      "config-lowering": 5.2999999752501026e-05,
+      "constant_folding": 7.999999979801942e-06,
+      "cse": 1.4999999621068127e-05,
+      "dce": 9.999999974752427e-07,
+      "dynamic-slice-transpose": 3.999999989900971e-06,
+      "eliminate-redundant-compare": 3.000000106112566e-06,
+      "emit-offloaded-dropout": 2.099999983329326e-05,
+      "flatten-call-graph": 9.999999747378752e-06,
+      "fuse-send-recv": 1.9999999494757503e-05,
+      "hilo::LegalizeAlias": 6.000000212225132e-06,
+      "hilo::NeuronInstCombine": 6.399999983841553e-05,
+      "hilo::NeuronOpFusion": 4.199999966658652e-05,
+      "hilo::ReplaceTokenTypeWithU8Pass": 6.000000212225132e-06,
+      "hilo::ScheduleFusion": 9.000000318337698e-06,
+      "hilo::SixtyFourHack": 1.4999999621068127e-05,
+      "hilo::VerifyAliasing": 3.000000106112566e-06,
+      "hlo-mac-count": 3.400000059627928e-05,
+      "hlo-verifier": 0.00014699999883305281,
+      "legalize-ccops": 1.9999999949504854e-06,
+      "legalize-compare": 3.999999989900971e-06,
+      "lower-argminmax-custom-call": 3.000000106112566e-06,
+      "map-inline": 9.999999747378752e-06,
+      "metadata-naming": 1.5999999959603883e-05,
+      "mlir::detail::OpToOpPassAdaptor": 3.300000025774352e-05,
+      "mlir::hlo::MhloToPyPenguin": 0.03224699944257736,
+      "mlir::mhlo::LowerComplexExtraPass": 0.00011999999696854502,
+      "mlir::mhlo::LowerComplexPass": 0.00016799999866634607,
+      "native-to-custom-softmax": 2.099999983329326e-05,
+      "native-to-custom-softmax-dx": 8.600000001024455e-05,
+      "operand_upcaster": 1.4000000192027073e-05,
+      "post-par-pipe-begin": 9.999999974752427e-07,
+      "post-par-pipe-end": 0.0,
+      "post-partition-simplification": 0.0005360000068321824,
+      "replace-minimum-constant": 7.999999979801942e-06,
+      "reshape-mover": 3.999999989900971e-06,
+      "simplify-concat": 5.199999941396527e-05,
+      "simplify-while-loops": 1.9999999949504854e-06,
+      "transform-variadic-reduce": 9.000000318337698e-06,
+      "tuple-simplifier": 3.999999989900971e-06,
+      "unpack-nested-aws-ntwsr": 3.000000106112566e-06,
+      "unroll-while-loop": 0.0
+    },
+    "hilo": {
+      "ArithmeticIntensity": 27.82740020751953,
+      "ConstantSize": 271333.0,
+      "HloInputCount": 359.0,
+      "HloMacCount": 2751463424.0,
+      "HloOutputCount": 65.0,
+      "IfmapSize": 7785161728.0,
+      "OfmapSize": 536870912.0,
+      "OutputsReadFromCount": 0.0,
+      "PassthroughTensorsCount": 0.0,
+      "RedundantOutputCount": 0.0,
+      "Traffic": 197752096.0
+    }
+  },
+  "sg0000": {
+    "compiletime": {
+      "AGOrderingAnalysisPass": 0.06467413902282715,
+      "AffinePredicateResolution": 0.002021312713623047,
+      "AliasDependencyElimination": 0.00022792816162109375,
+      "AliasDependencyInduction": 0.009432077407836914,
+      "AliasDependencyReset": 0.055196523666381836,
+      "BFComputeCutting": 0.0026645660400390625,
+      "BirCodeGenLoop": 0.04398298263549805,
+      "CCOpFusion": 0.013358116149902344,
+      "CanonicalizeDAGForPGTiling": 0.005773782730102539,
+      "CanonicalizeIR": 0.002256631851196289,
+      "CoalesceCCOp": 0.0045413970947265625,
+      "CommuteConcat": 0.0009562969207763672,
+      "DMALocalityOpt": 0.0011527538299560547,
+      "DMAProfiler": 0.00419306755065918,
+      "DMATilingProfiler": 0.003909587860107422,
+      "DataLocalityOpt": 0.18979620933532715,
+      "DataStreaming": 0.003125905990600586,
+      "DeConcat": 0.0009047985076904297,
+      "DeadCodeElimination": 0.0010743141174316406,
+      "DeadStoreElimination": 0.04163408279418945,
+      "DelinearIndices": 0.006937980651855469,
+      "Delinearization": 0.0029773712158203125,
+      "DoNothing": 0.0001316070556640625,
+      "DramToDramTranspose": 0.10994625091552734,
+      "DumpGraphAndMetadata": 0.012879133224487305,
+      "EliminateDivs": 0.00676274299621582,
+      "ExpandBatchNorm": 0.0029730796813964844,
+      "ExpandISAMacro": 0.003587961196899414,
+      "FactorizeBlkDims": 0.007817268371582031,
+      "FactorizeThreadAxesInFreeDims": 0.001369476318359375,
+      "FlattenMacroLoop": 0.002986907958984375,
+      "GenericAccessSimplifier": 0.0009241104125976563,
+      "InferInitValue": 0.06494665145874023,
+      "InferIntrinsicOnCC": 0.010074853897094727,
+      "InferNeuronTensor": 0.05807805061340332,
+      "InferNonlocalTensors": 0.07637619972229004,
+      "InferPSumTensor": 0.06549668312072754,
+      "InlineNativeKernels": 0.001531362533569336,
+      "InsertIOTransposes": 0.03947162628173828,
+      "InsertLocalTransposes": 0.011230707168579102,
+      "InsertOffloadedTransposes": 0.0025644302368164063,
+      "LICM": 0.0028715133666992188,
+      "LateLegalizeInst": 0.005754709243774414,
+      "LateLegalizePostSplit": 0.0025861263275146484,
+      "LateLowerReshapeOp": 0.0012161731719970703,
+      "LateLowerTensorOp": 0.010146856307983398,
+      "LateNeuronInstComb": 0.053855180740356445,
+      "LayoutPreprocessing": 0.02685070037841797,
+      "LayoutPreprocessingAndAnalysis": 0.050562143325805664,
+      "LayoutRequirementAnalysis": 0.0064203739166259766,
+      "LegalizeCCOpLayout": 0.002747058868408203,
+      "LegalizeOpLevelAlias": 0.0020356178283691406,
+      "LegalizePartitionReduce": 0.0014362335205078125,
+      "LegalizeSundaAccess": 0.07279443740844727,
+      "LegalizeSundaMacro": 0.0071544647216796875,
+      "LegalizeType": 0.003951549530029297,
+      "LocalLayoutOpt": 0.04533553123474121,
+      "LoopFusion": 0.005699634552001953,
+      "LoopSplitting": 0.0004475116729736328,
+      "LowerBroadcast": 0.002071380615234375,
+      "LowerCCOpBlockAxis": 0.0065212249755859375,
+      "LowerComplexBroadcast": 0.002064228057861328,
+      "LowerIntrinsics": 0.024514198303222656,
+      "LowerTensorOp": 0.01275944709777832,
+      "LowerTranspose": 0.007394313812255859,
+      "MacroGeneration": 0.11238408088684082,
+      "MaskPropagation": 0.03646087646484375,
+      "MemcpyElimination": 0.2132580280303955,
+      "MutateDataType": 0.001348257064819336,
+      "NeuronAliasDependencyInduction": 0.00030231475830078125,
+      "NeuronAliasDependencyReset": 0.047522783279418945,
+      "NeuronInstComb": 0.05169367790222168,
+      "NeuronLICM": 0.006502866744995117,
+      "NeuronLoopFusion": 0.05488872528076172,
+      "NeuronLoopInterchange": 0.0018024444580078125,
+      "NeuronSimplifier": 0.008179903030395508,
+      "NeuronSimplifyPredicates": 0.003264904022216797,
+      "NeuronValueNumbering": 0.003409862518310547,
+      "OptimizeAliasedCopyChain": 0.0009076595306396484,
+      "OptimizeNKIKernels": 0.001665353775024414,
+      "PAGLayoutOpt": 0.563495397567749,
+      "PComputeCutting": 0.009307861328125,
+      "PGLayoutTilingPipeline": 1.5276920795440674,
+      "PGTiling": 0.28220415115356445,
+      "PadElimination": 0.0005035400390625,
+      "ParAxesAnnotation": 0.5283262729644775,
+      "PartialLoopFusion": 0.008737325668334961,
+      "PartialSimdFusion": 0.0075261592864990234,
+      "PerfectLoopNest": 0.0016448497772216797,
+      "RecognizeOpIdiom": 0.0037012100219726563,
+      "Recompute": 0.0005586147308349609,
+      "RelaxPredicates": 0.0031774044036865234,
+      "Rematerialization": 0.0065402984619140625,
+      "ReshapeWeights": 0.000885009765625,
+      "ResolveAccessConflict": 0.0038557052612304688,
+      "ResolveComplicatePredicates": 0.0025780200958251953,
+      "RewriteReplicationMatmul": 0.0013949871063232422,
+      "RewriteWeights": 0.003153562545776367,
+      "SFKVectorizer": 0.08175945281982422,
+      "SimpleAllReduceTiling": 0.0017082691192626953,
+      "Simplifier": 0.015218257904052734,
+      "SimplifyMacroPredicates": 0.010445356369018555,
+      "SimplifyNeuronTensor": 0.006752490997314453,
+      "SimplifySlice": 0.0012371540069580078,
+      "SimplifyTensor": 0.005439281463623047,
+      "SpillPSum": 0.01030278205871582,
+      "SplitAPUnionSets": 0.03318023681640625,
+      "SplitAccGrp": 0.0014679431915283203,
+      "StaticProfiler": 0.003851652145385742,
+      "StaticTransposeLocalTensor": 0.008411169052124023,
+      "SundaISel": 0.08733582496643066,
+      "TCTransform": 0.0011320114135742188,
+      "TensorInitialization": 0.009679555892944336,
+      "TensorOpSimplifier": 0.008788108825683594,
+      "TensorOpTransform": 0.04109930992126465,
+      "TileCCOps": 0.006173372268676758,
+      "TilingProfiler": 0.009527444839477539,
+      "TransformConvOp": 0.002694368362426758,
+      "TritiumFusion": 0.01809835433959961,
+      "ValueNumbering": 0.002146005630493164,
+      "VectorizeDMA": 0.004664897918701172,
+      "VectorizeMatMult": 0.008771419525146484,
+      "WeightCoalescing": 0.0025610923767089844,
+      "ZeroSizeTensorElimination": 0.00019073486328125
+    },
+    "tensorizer": {
+      "DMATilingProfiler::TotalInstructionsAfterTiling": 1100.0,
+      "StaticProfiler::AifUb": 27.658906936645508,
+      "StaticProfiler::ArithmeticIntensityTensorizer": 73.42570495605469,
+      "StaticProfiler::AverageDmaLength": 9659.0947265625,
+      "StaticProfiler::AverageFractalPeUtilization": 99.76654052734375,
+      "StaticProfiler::AveragePartitionUtilization": 91.43461608886719,
+      "StaticProfiler::AveragePeUtilization": 99.76654052734375,
+      "StaticProfiler::DDRTransferBytes": 79836424.0,
+      "StaticProfiler::InternalTransferBytes": 8093696.0,
+      "StaticProfiler::LoadExpanded": 10755.0,
+      "StaticProfiler::LocalizationEfficiency": 265.46856689453125,
+      "StaticProfiler::LocalizationEfficiencyIgnoreNonlocal": 482.6210021972656,
+      "StaticProfiler::StoreExpanded": 3073.0,
+      "StaticProfiler::TotalDMAExpanded": 13828.0,
+      "StaticProfiler::TotalDynamicInstancesCount": 910.0,
+      "StaticProfiler::TotalDynamicInstancesWithMmPackedCount": 910.0,
+      "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": 437.0,
+      "TilingProfiler::NumPfTransposes": 4.0,
+      "TilingProfiler::NumPfTransposesForIo": 0.0,
+      "TilingProfiler::NumPfTransposesForLocal": 3.0,
+      "TilingProfiler::NumPfTransposesForNonlocal": 1.0,
+      "TilingProfiler::PfTransposeInstructions": 52.0,
+      "TilingProfiler::PfTransposeInstructionsForIo": 0.0,
+      "TilingProfiler::PfTransposeInstructionsForLocal": 36.0,
+      "TilingProfiler::PfTransposeInstructionsForNonlocal": 16.0,
+      "TilingProfiler::ReduceInstructionsAfterTiling": 0.0,
+      "TilingProfiler::SimdInstructionsAfterTiling": 99.0,
+      "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": 0.11036324501037598,
+      "AffinePredicateResolution": 0.002354860305786133,
+      "AliasDependencyElimination": 0.0002741813659667969,
+      "AliasDependencyInduction": 0.008609294891357422,
+      "AliasDependencyReset": 0.020910978317260742,
+      "BFComputeCutting": 0.001337289810180664,
+      "BirCodeGenLoop": 0.06340932846069336,
+      "CCOpFusion": 0.018311500549316406,
+      "CanonicalizeDAGForPGTiling": 0.004097461700439453,
+      "CanonicalizeIR": 0.001764535903930664,
+      "CoalesceCCOp": 0.004993915557861328,
+      "CommuteConcat": 0.0019958019256591797,
+      "DMALocalityOpt": 0.0014808177947998047,
+      "DMAProfiler": 0.0038518905639648438,
+      "DMATilingProfiler": 0.0041544437408447266,
+      "DataLocalityOpt": 0.20078325271606445,
+      "DataStreaming": 0.004034519195556641,
+      "DeConcat": 0.0015177726745605469,
+      "DeadCodeElimination": 0.0010747909545898438,
+      "DeadStoreElimination": 0.039188385009765625,
+      "DelinearIndices": 0.008949041366577148,
+      "Delinearization": 0.0033283233642578125,
+      "DoNothing": 0.00011873245239257813,
+      "DramToDramTranspose": 0.0655205249786377,
+      "DumpGraphAndMetadata": 0.012286186218261719,
+      "EliminateDivs": 0.0033621788024902344,
+      "ExpandBatchNorm": 0.0016183853149414063,
+      "ExpandISAMacro": 0.003101825714111328,
+      "FactorizeBlkDims": 0.008794307708740234,
+      "FactorizeThreadAxesInFreeDims": 0.0022835731506347656,
+      "FlattenMacroLoop": 0.0027937889099121094,
+      "GenericAccessSimplifier": 0.0009398460388183594,
+      "InferInitValue": 0.06454229354858398,
+      "InferIntrinsicOnCC": 0.009270429611206055,
+      "InferNeuronTensor": 0.033622026443481445,
+      "InferNonlocalTensors": 0.028959989547729492,
+      "InferPSumTensor": 0.09024310111999512,
+      "InlineNativeKernels": 0.0021810531616210938,
+      "InsertIOTransposes": 0.016814231872558594,
+      "InsertLocalTransposes": 0.009081840515136719,
+      "InsertOffloadedTransposes": 0.003053903579711914,
+      "LICM": 0.003207683563232422,
+      "LateLegalizeInst": 0.0053195953369140625,
+      "LateLegalizePostSplit": 0.0034935474395751953,
+      "LateLowerReshapeOp": 0.0013909339904785156,
+      "LateLowerTensorOp": 0.005151987075805664,
+      "LateNeuronInstComb": 0.026895761489868164,
+      "LayoutPreprocessing": 0.02971482276916504,
+      "LayoutPreprocessingAndAnalysis": 0.05191612243652344,
+      "LayoutRequirementAnalysis": 0.008739709854125977,
+      "LegalizeCCOpLayout": 0.00177764892578125,
+      "LegalizeOpLevelAlias": 0.0013723373413085938,
+      "LegalizePartitionReduce": 0.0010402202606201172,
+      "LegalizeSundaAccess": 0.07228899002075195,
+      "LegalizeSundaMacro": 0.009223222732543945,
+      "LegalizeType": 0.007097959518432617,
+      "LocalLayoutOpt": 0.020928382873535156,
+      "LoopFusion": 0.006183147430419922,
+      "LoopSplitting": 0.0005662441253662109,
+      "LowerBroadcast": 0.002042055130004883,
+      "LowerCCOpBlockAxis": 0.00509333610534668,
+      "LowerComplexBroadcast": 0.001802206039428711,
+      "LowerIntrinsics": 0.08197450637817383,
+      "LowerTensorOp": 0.012150287628173828,
+      "LowerTranspose": 0.009961605072021484,
+      "MacroGeneration": 0.07223057746887207,
+      "MaskPropagation": 0.003228425979614258,
+      "MemcpyElimination": 0.14747166633605957,
+      "MutateDataType": 0.0013821125030517578,
+      "NeuronAliasDependencyInduction": 0.0003688335418701172,
+      "NeuronAliasDependencyReset": 0.025783538818359375,
+      "NeuronInstComb": 0.014778375625610352,
+      "NeuronLICM": 0.029512643814086914,
+      "NeuronLoopFusion": 0.012999534606933594,
+      "NeuronLoopInterchange": 0.0011935234069824219,
+      "NeuronSimplifier": 0.009111881256103516,
+      "NeuronSimplifyPredicates": 0.0014109611511230469,
+      "NeuronValueNumbering": 0.05046653747558594,
+      "OptimizeAliasedCopyChain": 0.0009007453918457031,
+      "OptimizeNKIKernels": 0.0016131401062011719,
+      "PAGLayoutOpt": 0.41016316413879395,
+      "PComputeCutting": 0.035622358322143555,
+      "PGLayoutTilingPipeline": 1.0860846042633057,
+      "PGTiling": 0.28725552558898926,
+      "PadElimination": 0.0004394054412841797,
+      "ParAxesAnnotation": 0.33440423011779785,
+      "PartialLoopFusion": 0.01318979263305664,
+      "PartialSimdFusion": 0.013974428176879883,
+      "PerfectLoopNest": 0.0018496513366699219,
+      "RecognizeOpIdiom": 0.007380247116088867,
+      "Recompute": 0.00032448768615722656,
+      "RelaxPredicates": 0.004439115524291992,
+      "Rematerialization": 0.0023751258850097656,
+      "ReshapeWeights": 0.0008261203765869141,
+      "ResolveAccessConflict": 0.0039043426513671875,
+      "ResolveComplicatePredicates": 0.0013427734375,
+      "RewriteReplicationMatmul": 0.0014929771423339844,
+      "RewriteWeights": 0.0029494762420654297,
+      "SFKVectorizer": 0.257702112197876,
+      "SimpleAllReduceTiling": 0.0015501976013183594,
+      "Simplifier": 0.004350900650024414,
+      "SimplifyMacroPredicates": 0.0053751468658447266,
+      "SimplifyNeuronTensor": 0.0093841552734375,
+      "SimplifySlice": 0.0010516643524169922,
+      "SimplifyTensor": 0.0060961246490478516,
+      "SpillPSum": 0.02084517478942871,
+      "SplitAPUnionSets": 0.010507822036743164,
+      "SplitAccGrp": 0.0010433197021484375,
+      "StaticProfiler": 0.04431486129760742,
+      "StaticTransposeLocalTensor": 0.0048198699951171875,
+      "SundaISel": 0.08904266357421875,
+      "TCTransform": 0.0012717247009277344,
+      "TensorInitialization": 0.004460811614990234,
+      "TensorOpSimplifier": 0.006735324859619141,
+      "TensorOpTransform": 0.03174185752868652,
+      "TileCCOps": 0.00526738166809082,
+      "TilingProfiler": 0.010722160339355469,
+      "TransformConvOp": 0.002729177474975586,
+      "TritiumFusion": 0.10425376892089844,
+      "ValueNumbering": 0.0030641555786132813,
+      "VectorizeDMA": 0.003632783889770508,
+      "VectorizeMatMult": 0.0032911300659179688,
+      "WeightCoalescing": 0.0043370723724365234,
+      "ZeroSizeTensorElimination": 0.0002262592315673828
+    },
+    "tensorizer": {
+      "DMATilingProfiler::TotalInstructionsAfterTiling": 5856.0,
+      "StaticProfiler::AifUb": 112.5947036743164,
+      "StaticProfiler::ArithmeticIntensityTensorizer": 128.7735595703125,
+      "StaticProfiler::AverageDmaLength": 7108.31689453125,
+      "StaticProfiler::AverageFractalPeUtilization": 100.0,
+      "StaticProfiler::AveragePartitionUtilization": 99.92850494384766,
+      "StaticProfiler::AveragePeUtilization": 100.0,
+      "StaticProfiler::DDRTransferBytes": 223825920.0,
+      "StaticProfiler::InternalTransferBytes": 9797632.0,
+      "StaticProfiler::LoadExpanded": 29057.0,
+      "StaticProfiler::LocalizationEfficiency": 114.36911010742188,
+      "StaticProfiler::LocalizationEfficiencyIgnoreNonlocal": 116.55321502685547,
+      "StaticProfiler::StoreExpanded": 1153.0,
+      "StaticProfiler::TotalDMAExpanded": 30210.0,
+      "StaticProfiler::TotalDynamicInstancesCount": 5599.0,
+      "StaticProfiler::TotalDynamicInstancesWithMmPackedCount": 5599.0,
+      "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": 4756.0,
+      "TilingProfiler::NumPfTransposes": 7.0,
+      "TilingProfiler::NumPfTransposesForIo": 3.0,
+      "TilingProfiler::NumPfTransposesForLocal": 2.0,
+      "TilingProfiler::NumPfTransposesForNonlocal": 2.0,
+      "TilingProfiler::PfTransposeInstructions": 118.0,
+      "TilingProfiler::PfTransposeInstructionsForIo": 34.0,
+      "TilingProfiler::PfTransposeInstructionsForLocal": 20.0,
+      "TilingProfiler::PfTransposeInstructionsForNonlocal": 64.0,
+      "TilingProfiler::ReduceInstructionsAfterTiling": 0.0,
+      "TilingProfiler::SimdInstructionsAfterTiling": 186.0,
+      "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": 0.04956221580505371,
+      "AffinePredicateResolution": 0.002319812774658203,
+      "AliasDependencyElimination": 0.00019097328186035156,
+      "AliasDependencyInduction": 0.01746201515197754,
+      "AliasDependencyReset": 0.04065680503845215,
+      "BFComputeCutting": 0.0013394355773925781,
+      "BirCodeGenLoop": 0.11406064033508301,
+      "CCOpFusion": 0.03461766242980957,
+      "CanonicalizeDAGForPGTiling": 0.004343748092651367,
+      "CanonicalizeIR": 0.003444671630859375,
+      "CoalesceCCOp": 0.04682421684265137,
+      "CommuteConcat": 0.0010838508605957031,
+      "DMALocalityOpt": 0.002088308334350586,
+      "DMAProfiler": 0.005882740020751953,
+      "DMATilingProfiler": 0.004760265350341797,
+      "DataLocalityOpt": 0.11827898025512695,
+      "DataStreaming": 0.005929470062255859,
+      "DeConcat": 0.0006871223449707031,
+      "DeadCodeElimination": 0.0018725395202636719,
+      "DeadStoreElimination": 0.00615692138671875,
+      "DelinearIndices": 0.004624128341674805,
+      "Delinearization": 0.0023467540740966797,
+      "DoNothing": 0.0001399517059326172,
+      "DramToDramTranspose": 0.015562057495117188,
+      "DumpGraphAndMetadata": 0.021119117736816406,
+      "EliminateDivs": 0.0027742385864257813,
+      "ExpandBatchNorm": 0.0035254955291748047,
+      "ExpandISAMacro": 0.0046041011810302734,
+      "FactorizeBlkDims": 0.008176565170288086,
+      "FactorizeThreadAxesInFreeDims": 0.0011937618255615234,
+      "FlattenMacroLoop": 0.002167940139770508,
+      "GenericAccessSimplifier": 0.0008442401885986328,
+      "InferInitValue": 0.024956941604614258,
+      "InferIntrinsicOnCC": 0.015187978744506836,
+      "InferNeuronTensor": 0.01759958267211914,
+      "InferNonlocalTensors": 0.014631509780883789,
+      "InferPSumTensor": 0.03573298454284668,
+      "InlineNativeKernels": 0.0049402713775634766,
+      "InsertIOTransposes": 0.012220144271850586,
+      "InsertLocalTransposes": 0.0043489933013916016,
+      "InsertOffloadedTransposes": 0.0029261112213134766,
+      "LICM": 0.002863168716430664,
+      "LateLegalizeInst": 0.006443977355957031,
+      "LateLegalizePostSplit": 0.004354715347290039,
+      "LateLowerReshapeOp": 0.0013675689697265625,
+      "LateLowerTensorOp": 0.0017099380493164063,
+      "LateNeuronInstComb": 0.008248329162597656,
+      "LayoutPreprocessing": 0.024985551834106445,
+      "LayoutPreprocessingAndAnalysis": 0.05915069580078125,
+      "LayoutRequirementAnalysis": 0.004924297332763672,
+      "LegalizeCCOpLayout": 0.003811359405517578,
+      "LegalizeOpLevelAlias": 0.0012297630310058594,
+      "LegalizePartitionReduce": 0.001873016357421875,
+      "LegalizeSundaAccess": 0.06514930725097656,
+      "LegalizeSundaMacro": 0.010838985443115234,
+      "LegalizeType": 0.007501840591430664,
+      "LocalLayoutOpt": 0.01294255256652832,
+      "LoopFusion": 0.005301713943481445,
+      "LoopSplitting": 0.0006639957427978516,
+      "LowerBroadcast": 0.0011935234069824219,
+      "LowerCCOpBlockAxis": 0.003854513168334961,
+      "LowerComplexBroadcast": 0.0023345947265625,
+      "LowerIntrinsics": 0.3132970333099365,
+      "LowerTensorOp": 0.013880491256713867,
+      "LowerTranspose": 0.007669687271118164,
+      "MacroGeneration": 0.04993605613708496,
+      "MaskPropagation": 0.0027527809143066406,
+      "MemcpyElimination": 0.02739572525024414,
+      "MutateDataType": 0.0013971328735351563,
+      "NeuronAliasDependencyInduction": 0.0002524852752685547,
+      "NeuronAliasDependencyReset": 0.046864986419677734,
+      "NeuronInstComb": 0.006575107574462891,
+      "NeuronLICM": 0.03473496437072754,
+      "NeuronLoopFusion": 0.008702993392944336,
+      "NeuronLoopInterchange": 0.0014033317565917969,
+      "NeuronSimplifier": 0.0068361759185791016,
+      "NeuronSimplifyPredicates": 0.0032160282135009766,
+      "NeuronValueNumbering": 0.002985239028930664,
+      "OptimizeAliasedCopyChain": 0.00058746337890625,
+      "OptimizeNKIKernels": 0.5256326198577881,
+      "PAGLayoutOpt": 0.07949376106262207,
+      "PComputeCutting": 0.005449533462524414,
+      "PGLayoutTilingPipeline": 0.5191597938537598,
+      "PGTiling": 0.1754302978515625,
+      "PadElimination": 0.0005822181701660156,
+      "ParAxesAnnotation": 0.0432279109954834,
+      "PartialLoopFusion": 0.0034580230712890625,
+      "PartialSimdFusion": 0.003893136978149414,
+      "PerfectLoopNest": 0.0017859935760498047,
+      "RecognizeOpIdiom": 0.003717660903930664,
+      "Recompute": 0.0006694793701171875,
+      "RelaxPredicates": 0.004067182540893555,
+      "Rematerialization": 0.002377748489379883,
+      "ReshapeWeights": 0.0010006427764892578,
+      "ResolveAccessConflict": 0.003880739212036133,
+      "ResolveComplicatePredicates": 0.0018019676208496094,
+      "RewriteReplicationMatmul": 0.0015823841094970703,
+      "RewriteWeights": 0.002522706985473633,
+      "SFKVectorizer": 0.11758708953857422,
+      "SimpleAllReduceTiling": 0.0028557777404785156,
+      "Simplifier": 0.0035266876220703125,
+      "SimplifyMacroPredicates": 0.005109071731567383,
+      "SimplifyNeuronTensor": 0.2750875949859619,
+      "SimplifySlice": 0.0010409355163574219,
+      "SimplifyTensor": 0.025161027908325195,
+      "SpillPSum": 0.04991626739501953,
+      "SplitAPUnionSets": 0.013710737228393555,
+      "SplitAccGrp": 0.001070261001586914,
+      "StaticProfiler": 0.005475759506225586,
+      "StaticTransposeLocalTensor": 0.0038781166076660156,
+      "SundaISel": 0.04041314125061035,
+      "TCTransform": 0.0008709430694580078,
+      "TensorInitialization": 0.004585742950439453,
+      "TensorOpSimplifier": 0.005696535110473633,
+      "TensorOpTransform": 0.021961212158203125,
+      "TileCCOps": 0.010190010070800781,
+      "TilingProfiler": 0.04765009880065918,
+      "TransformConvOp": 0.0030345916748046875,
+      "TritiumFusion": 0.07207250595092773,
+      "ValueNumbering": 0.002035379409790039,
+      "VectorizeDMA": 0.006017446517944336,
+      "VectorizeMatMult": 0.0027844905853271484,
+      "WeightCoalescing": 0.003088235855102539,
+      "ZeroSizeTensorElimination": 0.00019311904907226563
+    },
+    "tensorizer": {
+      "DMATilingProfiler::TotalInstructionsAfterTiling": 9240.0,
+      "StaticProfiler::AifUb": 74.75788116455078,
+      "StaticProfiler::ArithmeticIntensityTensorizer": 73.24122619628906,
+      "StaticProfiler::AverageDmaLength": 6860.4052734375,
+      "StaticProfiler::AverageFractalPeUtilization": 99.61843872070313,
+      "StaticProfiler::AveragePartitionUtilization": 98.97808837890625,
+      "StaticProfiler::AveragePeUtilization": 98.48223876953125,
+      "StaticProfiler::DDRTransferBytes": 318999072.0,
+      "StaticProfiler::InternalTransferBytes": 6136852.0,
+      "StaticProfiler::LoadExpanded": 40210.0,
+      "StaticProfiler::LocalizationEfficiency": 97.97124481201172,
+      "StaticProfiler::LocalizationEfficiencyIgnoreNonlocal": 100.67024993896484,
+      "StaticProfiler::StoreExpanded": 1195.0,
+      "StaticProfiler::TotalDMAExpanded": 41405.0,
+      "StaticProfiler::TotalDynamicInstancesCount": 9321.0,
+      "StaticProfiler::TotalDynamicInstancesWithMmPackedCount": 9225.0,
+      "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": 8128.0,
+      "TilingProfiler::NumPfTransposes": 4.0,
+      "TilingProfiler::NumPfTransposesForIo": 0.0,
+      "TilingProfiler::NumPfTransposesForLocal": 1.0,
+      "TilingProfiler::NumPfTransposesForNonlocal": 3.0,
+      "TilingProfiler::PfTransposeInstructions": 97.0,
+      "TilingProfiler::PfTransposeInstructionsForIo": 0.0,
+      "TilingProfiler::PfTransposeInstructionsForLocal": 1.0,
+      "TilingProfiler::PfTransposeInstructionsForNonlocal": 96.0,
+      "TilingProfiler::ReduceInstructionsAfterTiling": 4.0,
+      "TilingProfiler::SimdInstructionsAfterTiling": 166.0,
+      "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": 0.00025499999173916876,
+      "HoistCompute": 3.000000106112566e-06,
+      "IdentifyCrossPassTensors": 1.1000000085914508e-05,
+      "MemcastMotion": 1.1000000085914508e-05,
+      "PenguinizeFunctions": 1.2000000424450263e-05,
+      "PruneFunctions": 1.2999999853491317e-05,
+      "RemoveOptimizationBarriers": 1.2999999853491317e-05,
+      "ScatterMotion": 3.000000106112566e-06,
+      "TensorizerLegalizationPass": 1.5999999959603883e-05,
+      "VerifySupportedOps": 9.999999747378752e-06,
+      "algsimp": 8.299999899463728e-05,
+      "batchnorm_expander": 1.700000029813964e-05,
+      "boundary-marker-removal": 6.000000212225132e-06,
+      "call-inliner": 9.999999747378752e-06,
+      "canonicalize-boundary-marker": 4.999999873689376e-06,
+      "collective-stream-id-checker": 1.2000000424450263e-05,
+      "comparison-expander": 4.999999873689376e-06,
+      "computation-deduplicator": 2.300000051036477e-05,
+      "conditional-to-select": 7.999999979801942e-06,
+      "config-lowering": 6.800000119255856e-05,
+      "constant_folding": 9.000000318337698e-06,
+      "cse": 2.9000000722589903e-05,
+      "dce": 1.9999999949504854e-06,
+      "dynamic-slice-transpose": 7.000000096013537e-06,
+      "eliminate-redundant-compare": 6.000000212225132e-06,
+      "emit-offloaded-dropout": 2.2000000171829015e-05,
+      "flatten-call-graph": 9.000000318337698e-06,
+      "fuse-send-recv": 2.5999999706982635e-05,
+      "hilo::LegalizeAlias": 4.999999873689376e-06,
+      "hilo::NeuronInstCombine": 5.2999999752501026e-05,
+      "hilo::NeuronOpFusion": 9.999999747378752e-06,
+      "hilo::ReplaceTokenTypeWithU8Pass": 4.999999873689376e-06,
+      "hilo::ScheduleFusion": 9.999999974752427e-07,
+      "hilo::SixtyFourHack": 9.999999747378752e-06,
+      "hilo::VerifyAliasing": 1.9999999949504854e-06,
+      "hlo-mac-count": 4.8999998398358e-05,
+      "hlo-verifier": 0.00022499999613501132,
+      "legalize-ccops": 9.999999974752427e-07,
+      "legalize-compare": 3.999999989900971e-06,
+      "lower-argminmax-custom-call": 6.000000212225132e-06,
+      "map-inline": 1.2000000424450263e-05,
+      "metadata-naming": 2.9999999242136255e-05,
+      "mlir::detail::OpToOpPassAdaptor": 2.9999999242136255e-05,
+      "mlir::hlo::MhloToPyPenguin": 0.02389100007712841,
+      "mlir::mhlo::LowerComplexExtraPass": 9.000000136438757e-05,
+      "mlir::mhlo::LowerComplexPass": 0.0001140000022132881,
+      "native-to-custom-softmax": 1.4999999621068127e-05,
+      "native-to-custom-softmax-dx": 3.199999991920777e-05,
+      "operand_upcaster": 2.2000000171829015e-05,
+      "post-par-pipe-begin": 9.999999974752427e-07,
+      "post-par-pipe-end": 0.0,
+      "post-partition-simplification": 0.0006050000083632767,
+      "replace-minimum-constant": 1.1000000085914508e-05,
+      "reshape-mover": 3.999999989900971e-06,
+      "simplify-concat": 5.700000110664405e-05,
+      "simplify-while-loops": 1.9999999949504854e-06,
+      "transform-variadic-reduce": 1.2999999853491317e-05,
+      "tuple-simplifier": 4.999999873689376e-06,
+      "unpack-nested-aws-ntwsr": 3.000000106112566e-06,
+      "unroll-while-loop": 0.0
+    },
+    "hilo": {
+      "ArithmeticIntensity": 116.26366424560547,
+      "HloMacCount": 14025752576.0,
+      "Traffic": 241274912.0
+    }
+  },
+  "sg02": {
+    "compiletime": {
+      "CanonicalizeConv": 0.0,
+      "CanonicalizeForTensorizer": 1.2999999853491317e-05,
+      "Canonicalizer": 0.00031900001340545714,
+      "HoistCompute": 0.0,
+      "IdentifyCrossPassTensors": 1.2000000424450263e-05,
+      "MemcastMotion": 1.1000000085914508e-05,
+      "PenguinizeFunctions": 9.999999747378752e-06,
+      "PruneFunctions": 7.000000096013537e-06,
+      "RemoveOptimizationBarriers": 1.4999999621068127e-05,
+      "ScatterMotion": 9.999999974752427e-07,
+      "TensorizerLegalizationPass": 6.000000212225132e-06,
+      "VerifySupportedOps": 1.1000000085914508e-05,
+      "algsimp": 6.70000008540228e-05,
+      "batchnorm_expander": 1.2999999853491317e-05,
+      "boundary-marker-removal": 3.000000106112566e-06,
+      "call-inliner": 1.1000000085914508e-05,
+      "canonicalize-boundary-marker": 3.999999989900971e-06,
+      "collective-stream-id-checker": 4.999999873689376e-06,
+      "comparison-expander": 4.999999873689376e-06,
+      "computation-deduplicator": 2.2000000171829015e-05,
+      "conditional-to-select": 6.000000212225132e-06,
+      "config-lowering": 5.999999848427251e-05,
+      "constant_folding": 7.999999979801942e-06,
+      "cse": 1.2999999853491317e-05,
+      "dce": 9.999999974752427e-07,
+      "dynamic-slice-transpose": 3.999999989900971e-06,
+      "eliminate-redundant-compare": 3.000000106112566e-06,
+      "emit-offloaded-dropout": 2.099999983329326e-05,
+      "flatten-call-graph": 1.1000000085914508e-05,
+      "fuse-send-recv": 2.099999983329326e-05,
+      "hilo::LegalizeAlias": 1.9999999949504854e-06,
+      "hilo::NeuronInstCombine": 6.299999949987978e-05,
+      "hilo::NeuronOpFusion": 1.8999999156221747e-05,
+      "hilo::ReplaceTokenTypeWithU8Pass": 7.000000096013537e-06,
+      "hilo::ScheduleFusion": 1.9999999949504854e-06,
+      "hilo::SixtyFourHack": 3.7999998312443495e-05,
+      "hilo::VerifyAliasing": 9.999999974752427e-07,
+      "hlo-mac-count": 0.00019099999917671084,
+      "hlo-verifier": 0.0001740000006975606,
+      "legalize-ccops": 9.999999974752427e-07,
+      "legalize-compare": 3.000000106112566e-06,
+      "lower-argminmax-custom-call": 3.000000106112566e-06,
+      "map-inline": 1.2999999853491317e-05,
+      "metadata-naming": 1.700000029813964e-05,
+      "mlir::detail::OpToOpPassAdaptor": 3.300000025774352e-05,
+      "mlir::hlo::MhloToPyPenguin": 0.016152000054717064,
+      "mlir::mhlo::LowerComplexExtraPass": 0.0001230000052601099,
+      "mlir::mhlo::LowerComplexPass": 8.199999865610152e-05,
+      "native-to-custom-softmax": 1.1000000085914508e-05,
+      "native-to-custom-softmax-dx": 3.099999958067201e-05,
+      "operand_upcaster": 1.4000000192027073e-05,
+      "post-par-pipe-begin": 9.999999974752427e-07,
+      "post-par-pipe-end": 0.0,
+      "post-partition-simplification": 0.0005029999883845448,
+      "replace-minimum-constant": 9.999999747378752e-06,
+      "reshape-mover": 3.000000106112566e-06,
+      "simplify-concat": 4.999999873689376e-05,
+      "simplify-while-loops": 1.9999999949504854e-06,
+      "transform-variadic-reduce": 4.70000013592653e-05,
+      "tuple-simplifier": 6.000000212225132e-06,
+      "unpack-nested-aws-ntwsr": 3.999999989900971e-06,
+      "unroll-while-loop": 0.0
+    },
+    "hilo": {
+      "ArithmeticIntensity": 72.34154510498047,
+      "HloMacCount": 11341398016.0,
+      "Traffic": 313551424.0
+    }
+  }
+}

context_encoding_model/_tp0_bk0/neuron_config.json

@@ -0,0 +1,213 @@
+{
+  "_attn_implementation_autoset": false,
+  "_name_or_path": "mistralai/Mistral-7B-Instruct-v0.3",
+  "add_cross_attention": false,
+  "architectures": [
+    "MistralForCausalLM"
+  ],
+  "attention_dropout": 0.0,
+  "attribute_map": {},
+  "bad_words_ids": null,
+  "begin_suppress_tokens": null,
+  "bos_token_id": 1,
+  "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": 2,
+  "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": 14336,
+  "is_decoder": false,
+  "is_encoder_decoder": false,
+  "label2id": {
+    "LABEL_0": 0,
+    "LABEL_1": 1
+  },
+  "length_penalty": 1.0,
+  "max_length": 20,
+  "max_position_embeddings": 32768,
+  "metadata": null,
+  "min_length": 0,
+  "model_type": "mistral",
+  "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": "NeuronLlamaAttention",
+    "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": 4,
+    "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": 4,
+    "max_context_length": 2048,
+    "max_length": 2048,
+    "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": 2048,
+    "n_positions": 2048,
+    "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": 2048,
+    "pa_num_blocks": 4,
+    "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": 2048,
+    "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": 4,
+    "speculation_length": 0,
+    "start_rank_id": 0,
+    "target": null,
+    "tile_cc": false,
+    "tkg_batch_size": 4,
+    "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": 32,
+  "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-05,
+  "rope_theta": 1000000.0,
+  "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": "4.42.0.dev0",
+  "typical_p": 1.0,
+  "use_bfloat16": false,
+  "use_cache": true,
+  "vocab_size": 32768
+}

context_encoding_model/_tp0_bk1/command.txt

@@ -0,0 +1 @@
+neuronx-cc compile --framework=XLA model.MODULE_68c159ab1fef44a40212+6a9a7e72.hlo_module.pb --output model.MODULE_68c159ab1fef44a40212+6a9a7e72.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_68c159ab1fef44a40212+6a9a7e72.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=/models/mistral-7b-v0.3-instruct-neuronx/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": 99.64185333251953,
+      "StaticProfiler::AveragePartitionUtilization": 99.05009460449219,
+      "StaticProfiler::AveragePeUtilization": 98.57437896728516,
+      "StaticProfiler::LocalizationEfficiency": 97.34869384765625,
+      "StaticProfiler::LocalizationEfficiencyIgnoreNonlocal": 101.34577178955078,
+      "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": 0.023177146911621094,
+      "AffinePredicateResolution": 0.0010716915130615234,
+      "AliasDependencyElimination": 0.0001862049102783203,
+      "AliasDependencyInduction": 0.006262540817260742,
+      "AliasDependencyReset": 0.030141830444335938,
+      "BFComputeCutting": 0.0014030933380126953,
+      "BirCodeGenLoop": 0.14523863792419434,
+      "CCOpFusion": 0.015876293182373047,
+      "CanonicalizeConv": 6.399999983841553e-05,
+      "CanonicalizeDAGForPGTiling": 0.004258632659912109,
+      "CanonicalizeForTensorizer": 4.5000000682193786e-05,
+      "CanonicalizeIR": 0.0015943050384521484,
+      "Canonicalizer": 0.0009469999931752682,
+      "CoalesceCCOp": 0.006041288375854492,
+      "CommuteConcat": 0.0009081363677978516,
+      "DMALocalityOpt": 0.002227783203125,
+      "DMAProfiler": 0.006844043731689453,
+      "DMATilingProfiler": 0.004567146301269531,
+      "DataLocalityOpt": 0.1468503475189209,
+      "DataStreaming": 0.05744600296020508,
+      "DeConcat": 0.0009293556213378906,
+      "DeadCodeElimination": 0.0011928081512451172,
+      "DeadStoreElimination": 0.007216453552246094,
+      "DelinearIndices": 0.008558034896850586,
+      "Delinearization": 0.004823923110961914,
+      "DoNothing": 0.00029468536376953125,
+      "DramToDramTranspose": 0.03753399848937988,
+      "DumpGraphAndMetadata": 0.021843910217285156,
+      "EliminateDivs": 0.0021796226501464844,
+      "ExpandBatchNorm": 0.001444101333618164,
+      "ExpandISAMacro": 0.00778961181640625,
+      "FactorizeBlkDims": 0.00781702995300293,
+      "FactorizeThreadAxesInFreeDims": 0.002001047134399414,
+      "FlattenMacroLoop": 0.002486705780029297,
+      "GenericAccessSimplifier": 0.0008151531219482422,
+      "HoistCompute": 4.999999873689376e-06,
+      "IdentifyCrossPassTensors": 3.600000127335079e-05,
+      "InferInitValue": 0.03303098678588867,
+      "InferIntrinsicOnCC": 0.008953571319580078,
+      "InferNeuronTensor": 0.06123208999633789,
+      "InferNonlocalTensors": 0.015596389770507813,
+      "InferPSumTensor": 0.11387157440185547,
+      "InlineNativeKernels": 0.003874063491821289,
+      "InsertIOTransposes": 0.015367746353149414,
+      "InsertLocalTransposes": 0.00810861587524414,
+      "InsertOffloadedTransposes": 0.0030138492584228516,
+      "LICM": 0.002691507339477539,
+      "LateLegalizeInst": 0.007213115692138672,
+      "LateLegalizePostSplit": 0.0043141841888427734,
+      "LateLowerReshapeOp": 0.0011355876922607422,
+      "LateLowerTensorOp": 0.0013995170593261719,
+      "LateNeuronInstComb": 0.009235620498657227,
+      "LayoutPreprocessing": 0.03390383720397949,
+      "LayoutPreprocessingAndAnalysis": 0.11866283416748047,
+      "LayoutRequirementAnalysis": 0.009894132614135742,
+      "LegalizeCCOpLayout": 0.0019948482513427734,
+      "LegalizeOpLevelAlias": 0.0011715888977050781,
+      "LegalizePartitionReduce": 0.0018146038055419922,
+      "LegalizeSundaAccess": 0.036304473876953125,
+      "LegalizeSundaMacro": 0.015540599822998047,
+      "LegalizeType": 0.006517171859741211,
+      "LocalLayoutOpt": 0.023403167724609375,
+      "LoopFusion": 0.018198251724243164,
+      "LoopSplitting": 0.0003662109375,
+      "LowerBroadcast": 0.001718282699584961,
+      "LowerCCOpBlockAxis": 0.0071604251861572266,
+      "LowerComplexBroadcast": 0.0025353431701660156,
+      "LowerIntrinsics": 0.30090999603271484,
+      "LowerTensorOp": 0.010535478591918945,
+      "LowerTranspose": 0.0088653564453125,
+      "MacroGeneration": 0.02730083465576172,
+      "MaskPropagation": 0.0026755332946777344,
+      "MemcastMotion": 2.5000001187436283e-05,
+      "MemcpyElimination": 0.04188370704650879,
+      "MutateDataType": 0.0019674301147460938,
+      "NeuronAliasDependencyInduction": 0.00026416778564453125,
+      "NeuronAliasDependencyReset": 0.010982990264892578,
+      "NeuronInstComb": 0.004952669143676758,
+      "NeuronLICM": 0.0615081787109375,
+      "NeuronLoopFusion": 0.009150981903076172,
+      "NeuronLoopInterchange": 0.002344846725463867,
+      "NeuronSimplifier": 0.0074083805084228516,
+      "NeuronSimplifyPredicates": 0.009469747543334961,
+      "NeuronValueNumbering": 0.0031404495239257813,
+      "OptimizeAliasedCopyChain": 0.0016322135925292969,
+      "OptimizeNKIKernels": 0.6652572154998779,
+      "PAGLayoutOpt": 0.22568559646606445,
+      "PComputeCutting": 0.005068063735961914,
+      "PGLayoutTilingPipeline": 0.725257158279419,
+      "PGTiling": 0.16728448867797852,
+      "PadElimination": 0.0007500648498535156,
+      "ParAxesAnnotation": 0.05478692054748535,
+      "PartialLoopFusion": 0.006835222244262695,
+      "PartialSimdFusion": 0.007959842681884766,
+      "PenguinizeFunctions": 3.900000228895806e-05,
+      "PerfectLoopNest": 0.0017228126525878906,
+      "PruneFunctions": 3.300000025774352e-05,
+      "RecognizeOpIdiom": 0.0035974979400634766,
+      "Recompute": 0.00035452842712402344,
+      "RelaxPredicates": 0.005320310592651367,
+      "Rematerialization": 0.0019502639770507813,
+      "RemoveOptimizationBarriers": 3.9999998989515007e-05,
+      "ReshapeWeights": 0.0009279251098632813,
+      "ResolveAccessConflict": 0.0035495758056640625,
+      "ResolveComplicatePredicates": 0.0010292530059814453,
+      "RewriteReplicationMatmul": 0.0021431446075439453,
+      "RewriteWeights": 0.0024466514587402344,
+      "SFKVectorizer": 0.13248300552368164,
+      "ScatterMotion": 3.7000001611886546e-05,
+      "SimpleAllReduceTiling": 0.0030663013458251953,
+      "Simplifier": 0.0042324066162109375,
+      "SimplifyMacroPredicates": 0.0056989192962646484,
+      "SimplifyNeuronTensor": 0.2761991024017334,
+      "SimplifySlice": 0.0008549690246582031,
+      "SimplifyTensor": 0.004892826080322266,
+      "SpillPSum": 0.010103225708007813,
+      "SplitAPUnionSets": 0.013408899307250977,
+      "SplitAccGrp": 0.0011219978332519531,
+      "StaticProfiler": 0.0052869319915771484,
+      "StaticTransposeLocalTensor": 0.003991365432739258,
+      "SundaISel": 0.07178783416748047,
+      "TCTransform": 0.0011143684387207031,
+      "TensorInitialization": 0.03500223159790039,
+      "TensorOpSimplifier": 0.00494384765625,
+      "TensorOpTransform": 0.01940751075744629,
+      "TensorizerLegalizationPass": 4.3000000005122274e-05,
+      "TileCCOps": 0.007318258285522461,
+      "TilingProfiler": 0.007637739181518555,
+      "TransformConvOp": 0.002967357635498047,
+      "TritiumFusion": 0.06856966018676758,
+      "ValueNumbering": 0.001973867416381836,
+      "VectorizeDMA": 0.002585887908935547,
+      "VectorizeMatMult": 0.004343986511230469,
+      "VerifySupportedOps": 2.9999999242136255e-05,
+      "WeightCoalescing": 0.003629446029663086,
+      "ZeroSizeTensorElimination": 0.00018334388732910156,
+      "algsimp": 0.002331000054255128,
+      "batchnorm_expander": 4.5000000682193786e-05,
+      "boundary-marker-removal": 9.999999747378752e-06,
+      "call-inliner": 0.00041500001680105925,
+      "canonicalize-boundary-marker": 1.2999999853491317e-05,
+      "collective-stream-id-checker": 0.00011100000119768083,
+      "comparison-expander": 0.0005629999795928597,
+      "computation-deduplicator": 5.499999679159373e-05,
+      "conditional-to-select": 1.5999999959603883e-05,
+      "config-lowering": 0.00015900000289548188,
+      "constant-statistics": 0.00039400000241585076,
+      "constant_folding": 0.0002890000178012997,
+      "cse": 3.9999998989515007e-05,
+      "dce": 7.899999764049426e-05,
+      "dot_decomposer": 0.0016339999856427312,
+      "dynamic-slice-transpose": 1.799999881768599e-05,
+      "eliminate-redundant-compare": 0.00026000000070780516,
+      "emit-offloaded-dropout": 6.399999983841553e-05,
+      "flatten-call-graph": 0.0006720000528730452,
+      "fuse-send-recv": 6.800000119255856e-05,
+      "hilo::LegalizeAlias": 1.2000000424450263e-05,
+      "hilo::NeuronInstCombine": 0.00011300000187475234,
+      "hilo::NeuronOpFusion": 2.700000004551839e-05,
+      "hilo::ReplaceTokenTypeWithU8Pass": 2.099999983329326e-05,
+      "hilo::ScheduleFusion": 3.999999989900971e-06,
+      "hilo::SixtyFourHack": 6.0999998822808266e-05,
+      "hilo::VerifyAliasing": 4.999999873689376e-06,
+      "hlo-mac-count": 0.001120000029914081,
+      "hlo-verifier": 0.009619999676942825,
+      "instruction-histogram": 0.0010300000431016088,
+      "io-con-pipe-begin": 4.999999873689376e-06,
+      "io-con-pipe-end": 9.999999974752427e-07,
+      "io-layout-normalization": 0.0015699999639764428,
+      "io-statistics": 7.500000356230885e-05,
+      "legalize-ccops": 3.000000106112566e-06,
+      "legalize-compare": 1.1000000085914508e-05,
+      "lower-argminmax-custom-call": 9.000000318337698e-06,
+      "map-inline": 0.0009160000481642783,
+      "metadata-naming": 5.5999997130129486e-05,
+      "mlir::detail::OpToOpPassAdaptor": 8.299999899463728e-05,
+      "mlir::hlo::MhloToPyPenguin": 0.0715779960155487,
+      "mlir::mhlo::LowerComplexExtraPass": 0.00030700000934302807,
+      "mlir::mhlo::LowerComplexPass": 0.0004949999856762588,
+      "native-to-custom-softmax": 0.0007829999667592347,
+      "native-to-custom-softmax-dx": 0.0006629999843426049,
+      "operand_upcaster": 4.70000013592653e-05,
+      "opt-barrier-removal": 0.0005810000002384186,
+      "post-par-pipe-begin": 4.999999873689376e-06,
+      "post-par-pipe-end": 0.0,
+      "post-partition-simplification": 0.00171000009868294,
+      "pre-par-pipe-begin": 9.999999974752427e-07,
+      "pre-par-pipe-end": 0.0,
+      "pre-partition-simplification": 0.17589299380779266,
+      "replace-minimum-constant": 0.00033999999868683517,
+      "reshape-mover": 0.0001140000022132881,
+      "simplify-concat": 0.00015700000221841037,
+      "simplify-while-loops": 9.800000407267362e-05,
+      "transform-variadic-reduce": 6.199999916134402e-05,
+      "tuple-simplifier": 0.00028500001644715667,
+      "unpack-nested-aws-ntwsr": 0.0004239999980200082,
+      "unroll-while-loop": 1.700000029813964e-05,
+      "zero_sized_hlo_elimination": 0.0009629999985918403
+    },
+    "hilo": {
+      "ConstantSize": 533461.0,
+      "HloInputCount": 359.0,
+      "HloMacCount": 56438554624.0,
+      "HloOutputCount": 65.0,
+      "IfmapSize": 7785162752.0,
+      "OfmapSize": 536870912.0,
+      "OutputsReadFromCount": 0.0,
+      "PassthroughTensorsCount": 0.0,
+      "RedundantOutputCount": 0.0,
+      "Traffic": 766973312.0
+    },
+    "tensorizer": {
+      "DMATilingProfiler::TotalInstructionsAfterTiling": 9788.0,
+      "StaticProfiler::AifUb": 148.07814025878906,
+      "StaticProfiler::ArithmeticIntensityTensorizer": 144.15213012695313,
+      "StaticProfiler::AverageDmaLength": 6822.7958984375,
+      "StaticProfiler::DDRTransferBytes": 323193888.0,
+      "StaticProfiler::InternalTransferBytes": 12166164.0,
+      "StaticProfiler::LoadExpanded": 40466.0,
+      "StaticProfiler::StoreExpanded": 1835.0,
+      "StaticProfiler::TotalDMAExpanded": 42301.0,
+      "StaticProfiler::TotalDynamicInstancesCount": 10024.0,
+      "StaticProfiler::TotalDynamicInstancesWithMmPackedCount": 9928.0,
+      "StaticProfiler::TotalLNCComm": 0.0,
+      "StaticProfiler::TotalLNCCommTransfer": 0.0,
+      "TilingProfiler::BatchnormInstructionsAfterTiling": 0.0,
+      "TilingProfiler::DmaInstructionsAfterTiling": 0.0,
+      "TilingProfiler::GenericInstructionsAfterTiling": 4.0,
+      "TilingProfiler::MatMultInstructionsAfterTiling": 8576.0,
+      "TilingProfiler::NumPfTransposes": 4.0,
+      "TilingProfiler::NumPfTransposesForIo": 0.0,
+      "TilingProfiler::NumPfTransposesForLocal": 1.0,
+      "TilingProfiler::NumPfTransposesForNonlocal": 3.0,
+      "TilingProfiler::PfTransposeInstructions": 193.0,
+      "TilingProfiler::PfTransposeInstructionsForIo": 0.0,
+      "TilingProfiler::PfTransposeInstructionsForLocal": 1.0,
+      "TilingProfiler::PfTransposeInstructionsForNonlocal": 192.0,
+      "TilingProfiler::ReduceInstructionsAfterTiling": 4.0,
+      "TilingProfiler::SimdInstructionsAfterTiling": 166.0,
+      "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": 0.002076999982818961,
+      "call-inliner": 0.0003870000073220581,
+      "collective-stream-id-checker": 9.7999996796716e-05,
+      "comparison-expander": 0.0005489999894052744,
+      "constant-statistics": 0.00039400000241585076,
+      "constant_folding": 0.0002629999944474548,
+      "dce": 7.599999662488699e-05,
+      "dot_decomposer": 0.0016339999856427312,
+      "eliminate-redundant-compare": 0.0002500000118743628,
+      "flatten-call-graph": 0.0006440000142902136,
+      "hlo-mac-count": 0.0008500000112690032,
+      "hlo-verifier": 0.009083000011742115,
+      "instruction-histogram": 0.0010300000431016088,
+      "io-con-pipe-begin": 4.999999873689376e-06,
+      "io-con-pipe-end": 9.999999974752427e-07,
+      "io-layout-normalization": 0.0015699999639764428,
+      "io-statistics": 7.500000356230885e-05,
+      "map-inline": 0.0008820000221021473,
+      "native-to-custom-softmax": 0.0007539999787695706,
+      "native-to-custom-softmax-dx": 0.0005559999844990671,
+      "opt-barrier-removal": 0.0005810000002384186,
+      "pre-par-pipe-begin": 9.999999974752427e-07,
+      "pre-par-pipe-end": 0.0,
+      "pre-partition-simplification": 0.17589299380779266,
+      "replace-minimum-constant": 0.0003129999968223274,
+      "reshape-mover": 0.00010399999882793054,
+      "simplify-while-loops": 9.200000204145908e-05,
+      "tuple-simplifier": 0.0002699999895412475,
+      "unpack-nested-aws-ntwsr": 0.0004140000091865659,
+      "unroll-while-loop": 1.700000029813964e-05,
+      "zero_sized_hlo_elimination": 0.0009629999985918403
+    }
+  },
+  "cumsum": {
+    "compiletime": {
+      "CoalesceCCOp": 0.0002796649932861328,
+      "DMALocalityOpt": 0.0002567768096923828,
+      "DMAProfiler": 0.001027822494506836,
+      "DataStreaming": 0.0003445148468017578,
+      "DoNothing": 0.0001685619354248047,
+      "ExpandISAMacro": 0.0005648136138916016,
+      "FactorizeBlkDims": 0.0005543231964111328,
+      "InferPSumTensor": 0.0006105899810791016,
+      "LateLegalizeInst": 0.0005190372467041016,
+      "LateNeuronInstComb": 0.0005986690521240234,
+      "LegalizeSundaAccess": 0.0017101764678955078,
+      "LegalizeType": 0.0003490447998046875,
+      "LowerBroadcast": 0.00031828880310058594,
+      "LowerIntrinsics": 0.0002765655517578125,
+      "LowerTranspose": 0.0002956390380859375,
+      "NeuronInstComb": 0.0006248950958251953,
+      "NeuronLICM": 0.0005059242248535156,
+      "NeuronSimplifyPredicates": 0.003926515579223633,
+      "NeuronValueNumbering": 0.0004730224609375,
+      "SFKVectorizer": 0.0031404495239257813,
+      "SimpleAllReduceTiling": 0.0002665519714355469,
+      "SimplifyNeuronTensor": 0.000499725341796875,
+      "SpillPSum": 0.0006055831909179688,
+      "WeightCoalescing": 0.00028896331787109375
+    }
+  },
+  "sg00": {
+    "compiletime": {
+      "CanonicalizeConv": 2.8000000384054147e-05,
+      "CanonicalizeForTensorizer": 1.8999999156221747e-05,
+      "Canonicalizer": 0.0003549999964889139,
+      "HoistCompute": 1.9999999949504854e-06,
+      "IdentifyCrossPassTensors": 1.4000000192027073e-05,
+      "MemcastMotion": 1.4000000192027073e-05,
+      "PenguinizeFunctions": 1.700000029813964e-05,
+      "PruneFunctions": 1.4000000192027073e-05,
+      "RemoveOptimizationBarriers": 1.4999999621068127e-05,
+      "ScatterMotion": 9.000000318337698e-06,
+      "TensorizerLegalizationPass": 2.300000051036477e-05,
+      "VerifySupportedOps": 9.999999747378752e-06,
+      "algsimp": 9.40000027185306e-05,
+      "batchnorm_expander": 1.8000000636675395e-05,
+      "boundary-marker-removal": 3.000000106112566e-06,
+      "call-inliner": 9.000000318337698e-06,
+      "canonicalize-boundary-marker": 3.999999989900971e-06,
+      "collective-stream-id-checker": 1.9999999949504854e-06,
+      "comparison-expander": 4.999999873689376e-06,
+      "computation-deduplicator": 1.4999999621068127e-05,
+      "conditional-to-select": 4.999999873689376e-06,
+      "config-lowering": 5.400000009103678e-05,
+      "constant_folding": 7.999999979801942e-06,
+      "cse": 1.4999999621068127e-05,
+      "dce": 9.999999974752427e-07,
+      "dynamic-slice-transpose": 7.999999979801942e-06,
+      "eliminate-redundant-compare": 3.000000106112566e-06,
+      "emit-offloaded-dropout": 2.300000051036477e-05,
+      "flatten-call-graph": 9.000000318337698e-06,
+      "fuse-send-recv": 2.2000000171829015e-05,
+      "hilo::LegalizeAlias": 6.000000212225132e-06,
+      "hilo::NeuronInstCombine": 5.199999941396527e-05,
+      "hilo::NeuronOpFusion": 9.999999974752427e-07,
+      "hilo::ReplaceTokenTypeWithU8Pass": 7.999999979801942e-06,
+      "hilo::ScheduleFusion": 9.999999974752427e-07,
+      "hilo::SixtyFourHack": 1.2000000424450263e-05,
+      "hilo::VerifyAliasing": 1.9999999949504854e-06,
+      "hlo-mac-count": 3.7000001611886546e-05,
+      "hlo-verifier": 0.00017800000205170363,
+      "legalize-ccops": 9.999999974752427e-07,
+      "legalize-compare": 3.999999989900971e-06,
+      "lower-argminmax-custom-call": 3.000000106112566e-06,
+      "map-inline": 9.999999747378752e-06,
+      "metadata-naming": 1.5999999959603883e-05,
+      "mlir::detail::OpToOpPassAdaptor": 3.600000127335079e-05,
+      "mlir::hlo::MhloToPyPenguin": 0.031582001596689224,
+      "mlir::mhlo::LowerComplexExtraPass": 0.00010399999882793054,
+      "mlir::mhlo::LowerComplexPass": 0.00015100000018719584,
+      "native-to-custom-softmax": 9.999999747378752e-06,
+      "native-to-custom-softmax-dx": 4.400000034365803e-05,
+      "operand_upcaster": 1.4000000192027073e-05,
+      "post-par-pipe-begin": 9.999999974752427e-07,
+      "post-par-pipe-end": 0.0,
+      "post-partition-simplification": 0.0006259999936446548,
+      "replace-minimum-constant": 9.000000318337698e-06,
+      "reshape-mover": 3.999999989900971e-06,
+      "simplify-concat": 5.2999999752501026e-05,
+      "simplify-while-loops": 1.9999999949504854e-06,
+      "transform-variadic-reduce": 7.999999979801942e-06,
+      "tuple-simplifier": 4.999999873689376e-06,
+      "unpack-nested-aws-ntwsr": 3.000000106112566e-06,
+      "unroll-while-loop": 0.0
+    },
+    "hilo": {
+      "ArithmeticIntensity": 55.650630950927734,
+      "ConstantSize": 533461.0,
+      "HloInputCount": 359.0,
+      "HloMacCount": 5637144576.0,
+      "HloOutputCount": 65.0,
+      "IfmapSize": 7785162752.0,
+      "OfmapSize": 536870912.0,
+      "OutputsReadFromCount": 0.0,
+      "PassthroughTensorsCount": 0.0,
+      "RedundantOutputCount": 0.0,
+      "Traffic": 202590496.0
+    }
+  },
+  "sg0000": {
+    "compiletime": {
+      "AGOrderingAnalysisPass": 0.03828167915344238,
+      "AffinePredicateResolution": 0.0017518997192382813,
+      "AliasDependencyElimination": 0.00022554397583007813,
+      "AliasDependencyInduction": 0.014617204666137695,
+      "AliasDependencyReset": 0.1334824562072754,
+      "BFComputeCutting": 0.0033521652221679688,
+      "BirCodeGenLoop": 0.049553871154785156,
+      "CCOpFusion": 0.058977603912353516,
+      "CanonicalizeDAGForPGTiling": 0.005285024642944336,
+      "CanonicalizeIR": 0.003525257110595703,
+      "CoalesceCCOp": 0.0051691532135009766,
+      "CommuteConcat": 0.0010175704956054688,
+      "DMALocalityOpt": 0.0012638568878173828,
+      "DMAProfiler": 0.00444793701171875,
+      "DMATilingProfiler": 0.0043413639068603516,
+      "DataLocalityOpt": 0.21851897239685059,
+      "DataStreaming": 0.003998756408691406,
+      "DeConcat": 0.0009417533874511719,
+      "DeadCodeElimination": 0.001861572265625,
+      "DeadStoreElimination": 0.027139902114868164,
+      "DelinearIndices": 0.009176254272460938,
+      "Delinearization": 0.002630472183227539,
+      "DoNothing": 0.000125885009765625,
+      "DramToDramTranspose": 0.026975631713867188,
+      "DumpGraphAndMetadata": 0.012708663940429688,
+      "EliminateDivs": 0.02721691131591797,
+      "ExpandBatchNorm": 0.0034477710723876953,
+      "ExpandISAMacro": 0.0028188228607177734,
+      "FactorizeBlkDims": 0.020637035369873047,
+      "FactorizeThreadAxesInFreeDims": 0.0013651847839355469,
+      "FlattenMacroLoop": 0.004672050476074219,
+      "GenericAccessSimplifier": 0.0009248256683349609,
+      "InferInitValue": 0.03484296798706055,
+      "InferIntrinsicOnCC": 0.01741480827331543,
+      "InferNeuronTensor": 0.09466671943664551,
+      "InferNonlocalTensors": 0.16442584991455078,
+      "InferPSumTensor": 0.03455662727355957,
+      "InlineNativeKernels": 0.0014760494232177734,
+      "InsertIOTransposes": 0.04430508613586426,
+      "InsertLocalTransposes": 0.03803539276123047,
+      "InsertOffloadedTransposes": 0.004397869110107422,
+      "LICM": 0.0029926300048828125,
+      "LateLegalizeInst": 0.0474398136138916,
+      "LateLegalizePostSplit": 0.0027294158935546875,
+      "LateLowerReshapeOp": 0.002092599868774414,
+      "LateLowerTensorOp": 0.010335206985473633,
+      "LateNeuronInstComb": 0.07859206199645996,
+      "LayoutPreprocessing": 0.028697729110717773,
+      "LayoutPreprocessingAndAnalysis": 0.10470056533813477,
+      "LayoutRequirementAnalysis": 0.00815582275390625,
+      "LegalizeCCOpLayout": 0.003177165985107422,
+      "LegalizeOpLevelAlias": 0.0010378360748291016,
+      "LegalizePartitionReduce": 0.0013747215270996094,
+      "LegalizeSundaAccess": 0.04130220413208008,
+      "LegalizeSundaMacro": 0.010147333145141602,
+      "LegalizeType": 0.0038063526153564453,
+      "LocalLayoutOpt": 0.026745080947875977,
+      "LoopFusion": 0.006239652633666992,
+      "LoopSplitting": 0.00046253204345703125,
+      "LowerBroadcast": 0.003262758255004883,
+      "LowerCCOpBlockAxis": 0.044937849044799805,
+      "LowerComplexBroadcast": 0.0021724700927734375,
+      "LowerIntrinsics": 0.05569720268249512,
+      "LowerTensorOp": 0.016431331634521484,
+      "LowerTranspose": 0.04205179214477539,
+      "MacroGeneration": 0.18246054649353027,
+      "MaskPropagation": 0.008913278579711914,
+      "MemcpyElimination": 0.14683842658996582,
+      "MutateDataType": 0.0012350082397460938,
+      "NeuronAliasDependencyInduction": 0.00030422210693359375,
+      "NeuronAliasDependencyReset": 0.1192784309387207,
+      "NeuronInstComb": 0.013055801391601563,
+      "NeuronLICM": 0.008586406707763672,
+      "NeuronLoopFusion": 0.015465021133422852,
+      "NeuronLoopInterchange": 0.0020461082458496094,
+      "NeuronSimplifier": 0.05031108856201172,
+      "NeuronSimplifyPredicates": 0.005561351776123047,
+      "NeuronValueNumbering": 0.004122495651245117,
+      "OptimizeAliasedCopyChain": 0.0008661746978759766,
+      "OptimizeNKIKernels": 0.001760721206665039,
+      "PAGLayoutOpt": 0.49636244773864746,
+      "PComputeCutting": 0.008408069610595703,
+      "PGLayoutTilingPipeline": 1.6429040431976318,
+      "PGTiling": 0.6523551940917969,
+      "PadElimination": 0.0006270408630371094,
+      "ParAxesAnnotation": 0.32938289642333984,
+      "PartialLoopFusion": 0.016936302185058594,
+      "PartialSimdFusion": 0.012967824935913086,
+      "PerfectLoopNest": 0.002012968063354492,
+      "RecognizeOpIdiom": 0.0038352012634277344,
+      "Recompute": 0.0003256797790527344,
+      "RelaxPredicates": 0.0037698745727539063,
+      "Rematerialization": 0.003920078277587891,
+      "ReshapeWeights": 0.0009107589721679688,
+      "ResolveAccessConflict": 0.027127504348754883,
+      "ResolveComplicatePredicates": 0.0018274784088134766,
+      "RewriteReplicationMatmul": 0.002187490463256836,
+      "RewriteWeights": 0.006139278411865234,
+      "SFKVectorizer": 0.28095126152038574,
+      "SimpleAllReduceTiling": 0.0017786026000976563,
+      "Simplifier": 0.007991552352905273,
+      "SimplifyMacroPredicates": 0.044501304626464844,
+      "SimplifyNeuronTensor": 0.009800195693969727,
+      "SimplifySlice": 0.0012738704681396484,
+      "SimplifyTensor": 0.005684375762939453,
+      "SpillPSum": 0.013557672500610352,
+      "SplitAPUnionSets": 0.030807971954345703,
+      "SplitAccGrp": 0.0014066696166992188,
+      "StaticProfiler": 0.004358768463134766,
+      "StaticTransposeLocalTensor": 0.03508925437927246,
+      "SundaISel": 0.08105349540710449,
+      "TCTransform": 0.0011637210845947266,
+      "TensorInitialization": 0.012772321701049805,
+      "TensorOpSimplifier": 0.006896257400512695,
+      "TensorOpTransform": 0.07355618476867676,
+      "TileCCOps": 0.010430574417114258,
+      "TilingProfiler": 0.011465787887573242,
+      "TransformConvOp": 0.002571582794189453,
+      "TritiumFusion": 0.035936594009399414,
+      "ValueNumbering": 0.0023512840270996094,
+      "VectorizeDMA": 0.005614042282104492,
+      "VectorizeMatMult": 0.01862645149230957,
+      "WeightCoalescing": 0.0029726028442382813,
+      "ZeroSizeTensorElimination": 0.00018477439880371094
+    },
+    "tensorizer": {
+      "DMATilingProfiler::TotalInstructionsAfterTiling": 1880.0,
+      "StaticProfiler::AifUb": 56.6817626953125,
+      "StaticProfiler::ArithmeticIntensityTensorizer": 228.08059692382813,
+      "StaticProfiler::AverageDmaLength": 3633.1689453125,
+      "StaticProfiler::AverageFractalPeUtilization": 99.77600860595703,
+      "StaticProfiler::AveragePartitionUtilization": 95.7789077758789,
+      "StaticProfiler::AveragePeUtilization": 99.3130111694336,
+      "StaticProfiler::DDRTransferBytes": 53225736.0,
+      "StaticProfiler::InternalTransferBytes": 22481920.0,
+      "StaticProfiler::LoadExpanded": 9989.0,
+      "StaticProfiler::LocalizationEfficiency": 402.3879699707031,
+      "StaticProfiler::LocalizationEfficiencyIgnoreNonlocal": 466.7553405761719,
+      "StaticProfiler::StoreExpanded": 3713.0,
+      "StaticProfiler::TotalDMAExpanded": 13702.0,
+      "StaticProfiler::TotalDynamicInstancesCount": 1649.0,
+      "StaticProfiler::TotalDynamicInstancesWithMmPackedCount": 1645.0,
+      "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": 882.0,
+      "TilingProfiler::NumPfTransposes": 4.0,
+      "TilingProfiler::NumPfTransposesForIo": 0.0,
+      "TilingProfiler::NumPfTransposesForLocal": 3.0,
+      "TilingProfiler::NumPfTransposesForNonlocal": 1.0,
+      "TilingProfiler::PfTransposeInstructions": 136.0,
+      "TilingProfiler::PfTransposeInstructionsForIo": 0.0,
+      "TilingProfiler::PfTransposeInstructionsForLocal": 104.0,
+      "TilingProfiler::PfTransposeInstructionsForNonlocal": 32.0,
+      "TilingProfiler::ReduceInstructionsAfterTiling": 0.0,
+      "TilingProfiler::SimdInstructionsAfterTiling": 170.0,
+      "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": 0.1372241973876953,
+      "AffinePredicateResolution": 0.0012900829315185547,
+      "AliasDependencyElimination": 0.00016880035400390625,
+      "AliasDependencyInduction": 0.00756525993347168,
+      "AliasDependencyReset": 0.05401206016540527,
+      "BFComputeCutting": 0.0025975704193115234,
+      "BirCodeGenLoop": 0.06449413299560547,
+      "CCOpFusion": 0.05572986602783203,
+      "CanonicalizeDAGForPGTiling": 0.005880832672119141,
+      "CanonicalizeIR": 0.0017058849334716797,
+      "CoalesceCCOp": 0.007529258728027344,
+      "CommuteConcat": 0.0010385513305664063,
+      "DMALocalityOpt": 0.0021440982818603516,
+      "DMAProfiler": 0.006747007369995117,
+      "DMATilingProfiler": 0.0044765472412109375,
+      "DataLocalityOpt": 0.2653350830078125,
+      "DataStreaming": 0.044264793395996094,
+      "DeConcat": 0.0012521743774414063,
+      "DeadCodeElimination": 0.0011494159698486328,
+      "DeadStoreElimination": 0.030911922454833984,
+      "DelinearIndices": 0.008881092071533203,
+      "Delinearization": 0.0041217803955078125,
+      "DoNothing": 0.00018095970153808594,
+      "DramToDramTranspose": 0.0739753246307373,
+      "DumpGraphAndMetadata": 0.012173652648925781,
+      "EliminateDivs": 0.003404378890991211,
+      "ExpandBatchNorm": 0.0016205310821533203,
+      "ExpandISAMacro": 0.004185914993286133,
+      "FactorizeBlkDims": 0.016408443450927734,
+      "FactorizeThreadAxesInFreeDims": 0.002920866012573242,
+      "FlattenMacroLoop": 0.003094196319580078,
+      "GenericAccessSimplifier": 0.0010533332824707031,
+      "InferInitValue": 0.07179093360900879,
+      "InferIntrinsicOnCC": 0.011938810348510742,
+      "InferNeuronTensor": 0.08669257164001465,
+      "InferNonlocalTensors": 0.05255532264709473,
+      "InferPSumTensor": 0.028987407684326172,
+      "InlineNativeKernels": 0.0014684200286865234,
+      "InsertIOTransposes": 0.02220773696899414,
+      "InsertLocalTransposes": 0.009196281433105469,
+      "InsertOffloadedTransposes": 0.003507852554321289,
+      "LICM": 0.0032300949096679688,
+      "LateLegalizeInst": 0.006733417510986328,
+      "LateLegalizePostSplit": 0.003352642059326172,
+      "LateLowerReshapeOp": 0.0015027523040771484,
+      "LateLowerTensorOp": 0.005047321319580078,
+      "LateNeuronInstComb": 0.025189876556396484,
+      "LayoutPreprocessing": 0.07440996170043945,
+      "LayoutPreprocessingAndAnalysis": 0.11858534812927246,
+      "LayoutRequirementAnalysis": 0.009042739868164063,
+      "LegalizeCCOpLayout": 0.0016829967498779297,
+      "LegalizeOpLevelAlias": 0.0010685920715332031,
+      "LegalizePartitionReduce": 0.0022325515747070313,
+      "LegalizeSundaAccess": 0.03753972053527832,
+      "LegalizeSundaMacro": 0.009591341018676758,
+      "LegalizeType": 0.015005111694335938,
+      "LocalLayoutOpt": 0.04438447952270508,
+      "LoopFusion": 0.006495237350463867,
+      "LoopSplitting": 0.00040650367736816406,
+      "LowerBroadcast": 0.0017704963684082031,
+      "LowerCCOpBlockAxis": 0.027849197387695313,
+      "LowerComplexBroadcast": 0.0019054412841796875,
+      "LowerIntrinsics": 0.031024932861328125,
+      "LowerTensorOp": 0.010584592819213867,
+      "LowerTranspose": 0.011832475662231445,
+      "MacroGeneration": 0.09017419815063477,
+      "MaskPropagation": 0.0037479400634765625,
+      "MemcpyElimination": 0.12743902206420898,
+      "MutateDataType": 0.0015845298767089844,
+      "NeuronAliasDependencyInduction": 0.0002963542938232422,
+      "NeuronAliasDependencyReset": 0.011981010437011719,
+      "NeuronInstComb": 0.0240938663482666,
+      "NeuronLICM": 0.006990671157836914,
+      "NeuronLoopFusion": 0.03093886375427246,
+      "NeuronLoopInterchange": 0.0030279159545898438,
+      "NeuronSimplifier": 0.011322021484375,
+      "NeuronSimplifyPredicates": 0.0017275810241699219,
+      "NeuronValueNumbering": 0.0033507347106933594,
+      "OptimizeAliasedCopyChain": 0.0006761550903320313,
+      "OptimizeNKIKernels": 0.0017766952514648438,
+      "PAGLayoutOpt": 0.6475393772125244,
+      "PComputeCutting": 0.006641864776611328,
+      "PGLayoutTilingPipeline": 1.6072959899902344,
+      "PGTiling": 0.49286890029907227,
+      "PadElimination": 0.0004642009735107422,
+      "ParAxesAnnotation": 0.6226985454559326,
+      "PartialLoopFusion": 0.045456886291503906,
+      "PartialSimdFusion": 0.052849769592285156,
+      "PerfectLoopNest": 0.004052400588989258,
+      "RecognizeOpIdiom": 0.004006147384643555,
+      "Recompute": 0.0004963874816894531,
+      "RelaxPredicates": 0.0033147335052490234,
+      "Rematerialization": 0.002171039581298828,
+      "ReshapeWeights": 0.0008096694946289063,
+      "ResolveAccessConflict": 0.007338285446166992,
+      "ResolveComplicatePredicates": 0.00125885009765625,
+      "RewriteReplicationMatmul": 0.001424551010131836,
+      "RewriteWeights": 0.00428318977355957,
+      "SFKVectorizer": 0.2604203224182129,
+      "SimpleAllReduceTiling": 0.0015559196472167969,
+      "Simplifier": 0.004492521286010742,
+      "SimplifyMacroPredicates": 0.006475210189819336,
+      "SimplifyNeuronTensor": 0.012001514434814453,
+      "SimplifySlice": 0.0011138916015625,
+      "SimplifyTensor": 0.04663205146789551,
+      "SpillPSum": 0.017986297607421875,
+      "SplitAPUnionSets": 0.0433957576751709,
+      "SplitAccGrp": 0.0013687610626220703,
+      "StaticProfiler": 0.004006624221801758,
+      "StaticTransposeLocalTensor": 0.004962921142578125,
+      "SundaISel": 0.08249163627624512,
+      "TCTransform": 0.001107931137084961,
+      "TensorInitialization": 0.002826690673828125,
+      "TensorOpSimplifier": 0.005883216857910156,
+      "TensorOpTransform": 0.03051924705505371,
+      "TileCCOps": 0.005392313003540039,
+      "TilingProfiler": 0.05383157730102539,
+      "TransformConvOp": 0.0025300979614257813,
+      "TritiumFusion": 0.1005716323852539,
+      "ValueNumbering": 0.002640962600708008,
+      "VectorizeDMA": 0.0018579959869384766,
+      "VectorizeMatMult": 0.04578280448913574,
+      "WeightCoalescing": 0.0028417110443115234,
+      "ZeroSizeTensorElimination": 0.00020384788513183594
+    },
+    "tensorizer": {
+      "DMATilingProfiler::TotalInstructionsAfterTiling": 6724.0,
+      "StaticProfiler::AifUb": 223.01504516601563,
+      "StaticProfiler::ArithmeticIntensityTensorizer": 252.92869567871094,
+      "StaticProfiler::AverageDmaLength": 6766.392578125,
+      "StaticProfiler::AverageFractalPeUtilization": 100.0,
+      "StaticProfiler::AveragePartitionUtilization": 99.88066864013672,
+      "StaticProfiler::AveragePeUtilization": 100.0,
+      "StaticProfiler::DDRTransferBytes": 229515264.0,
+      "StaticProfiler::InternalTransferBytes": 20119552.0,
+      "StaticProfiler::LoadExpanded": 29313.0,
+      "StaticProfiler::LocalizationEfficiency": 113.41328430175781,
+      "StaticProfiler::LocalizationEfficiencyIgnoreNonlocal": 117.71570587158203,
+      "StaticProfiler::StoreExpanded": 2305.0,
+      "StaticProfiler::TotalDMAExpanded": 31618.0,
+      "StaticProfiler::TotalDynamicInstancesCount": 6708.0,
+      "StaticProfiler::TotalDynamicInstancesWithMmPackedCount": 6708.0,
+      "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": 5424.0,
+      "TilingProfiler::NumPfTransposes": 7.0,
+      "TilingProfiler::NumPfTransposesForIo": 3.0,
+      "TilingProfiler::NumPfTransposesForLocal": 2.0,
+      "TilingProfiler::NumPfTransposesForNonlocal": 2.0,
+      "TilingProfiler::PfTransposeInstructions": 268.0,
+      "TilingProfiler::PfTransposeInstructionsForIo": 68.0,
+      "TilingProfiler::PfTransposeInstructionsForLocal": 72.0,
+      "TilingProfiler::PfTransposeInstructionsForNonlocal": 128.0,
+      "TilingProfiler::ReduceInstructionsAfterTiling": 0.0,
+      "TilingProfiler::SimdInstructionsAfterTiling": 217.0,
+      "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": 0.023177146911621094,
+      "AffinePredicateResolution": 0.0010716915130615234,
+      "AliasDependencyElimination": 0.0001862049102783203,
+      "AliasDependencyInduction": 0.006262540817260742,
+      "AliasDependencyReset": 0.030141830444335938,
+      "BFComputeCutting": 0.0014030933380126953,
+      "BirCodeGenLoop": 0.14523863792419434,
+      "CCOpFusion": 0.015876293182373047,
+      "CanonicalizeDAGForPGTiling": 0.004258632659912109,
+      "CanonicalizeIR": 0.0015943050384521484,
+      "CoalesceCCOp": 0.005761623382568359,
+      "CommuteConcat": 0.0009081363677978516,
+      "DMALocalityOpt": 0.001971006393432617,
+      "DMAProfiler": 0.005816221237182617,
+      "DMATilingProfiler": 0.004567146301269531,
+      "DataLocalityOpt": 0.1468503475189209,
+      "DataStreaming": 0.05710148811340332,
+      "DeConcat": 0.0009293556213378906,
+      "DeadCodeElimination": 0.0011928081512451172,
+      "DeadStoreElimination": 0.007216453552246094,
+      "DelinearIndices": 0.008558034896850586,
+      "Delinearization": 0.004823923110961914,
+      "DoNothing": 0.00012612342834472656,
+      "DramToDramTranspose": 0.03753399848937988,
+      "DumpGraphAndMetadata": 0.021843910217285156,
+      "EliminateDivs": 0.0021796226501464844,
+      "ExpandBatchNorm": 0.001444101333618164,
+      "ExpandISAMacro": 0.0072247982025146484,
+      "FactorizeBlkDims": 0.007262706756591797,
+      "FactorizeThreadAxesInFreeDims": 0.002001047134399414,
+      "FlattenMacroLoop": 0.002486705780029297,
+      "GenericAccessSimplifier": 0.0008151531219482422,
+      "InferInitValue": 0.03303098678588867,
+      "InferIntrinsicOnCC": 0.008953571319580078,
+      "InferNeuronTensor": 0.06123208999633789,
+      "InferNonlocalTensors": 0.015596389770507813,
+      "InferPSumTensor": 0.11326098442077637,
+      "InlineNativeKernels": 0.003874063491821289,
+      "InsertIOTransposes": 0.015367746353149414,
+      "InsertLocalTransposes": 0.00810861587524414,
+      "InsertOffloadedTransposes": 0.0030138492584228516,
+      "LICM": 0.002691507339477539,
+      "LateLegalizeInst": 0.00669407844543457,
+      "LateLegalizePostSplit": 0.0043141841888427734,
+      "LateLowerReshapeOp": 0.0011355876922607422,
+      "LateLowerTensorOp": 0.0013995170593261719,
+      "LateNeuronInstComb": 0.008636951446533203,
+      "LayoutPreprocessing": 0.03390383720397949,
+      "LayoutPreprocessingAndAnalysis": 0.11866283416748047,
+      "LayoutRequirementAnalysis": 0.009894132614135742,
+      "LegalizeCCOpLayout": 0.0019948482513427734,
+      "LegalizeOpLevelAlias": 0.0011715888977050781,
+      "LegalizePartitionReduce": 0.0018146038055419922,
+      "LegalizeSundaAccess": 0.03459429740905762,
+      "LegalizeSundaMacro": 0.015540599822998047,
+      "LegalizeType": 0.0061681270599365234,
+      "LocalLayoutOpt": 0.023403167724609375,
+      "LoopFusion": 0.018198251724243164,
+      "LoopSplitting": 0.0003662109375,
+      "LowerBroadcast": 0.001399993896484375,
+      "LowerCCOpBlockAxis": 0.0071604251861572266,
+      "LowerComplexBroadcast": 0.0025353431701660156,
+      "LowerIntrinsics": 0.30063343048095703,
+      "LowerTensorOp": 0.010535478591918945,
+      "LowerTranspose": 0.008569717407226563,
+      "MacroGeneration": 0.02730083465576172,
+      "MaskPropagation": 0.0026755332946777344,
+      "MemcpyElimination": 0.04188370704650879,
+      "MutateDataType": 0.0019674301147460938,
+      "NeuronAliasDependencyInduction": 0.00026416778564453125,
+      "NeuronAliasDependencyReset": 0.010982990264892578,
+      "NeuronInstComb": 0.0043277740478515625,
+      "NeuronLICM": 0.061002254486083984,
+      "NeuronLoopFusion": 0.009150981903076172,
+      "NeuronLoopInterchange": 0.002344846725463867,
+      "NeuronSimplifier": 0.0074083805084228516,
+      "NeuronSimplifyPredicates": 0.005543231964111328,
+      "NeuronValueNumbering": 0.0026674270629882813,
+      "OptimizeAliasedCopyChain": 0.0016322135925292969,
+      "OptimizeNKIKernels": 0.6652572154998779,
+      "PAGLayoutOpt": 0.22568559646606445,
+      "PComputeCutting": 0.005068063735961914,
+      "PGLayoutTilingPipeline": 0.725257158279419,
+      "PGTiling": 0.16728448867797852,
+      "PadElimination": 0.0007500648498535156,
+      "ParAxesAnnotation": 0.05478692054748535,
+      "PartialLoopFusion": 0.006835222244262695,
+      "PartialSimdFusion": 0.007959842681884766,
+      "PerfectLoopNest": 0.0017228126525878906,
+      "RecognizeOpIdiom": 0.0035974979400634766,
+      "Recompute": 0.00035452842712402344,
+      "RelaxPredicates": 0.005320310592651367,
+      "Rematerialization": 0.0019502639770507813,
+      "ReshapeWeights": 0.0009279251098632813,
+      "ResolveAccessConflict": 0.0035495758056640625,
+      "ResolveComplicatePredicates": 0.0010292530059814453,
+      "RewriteReplicationMatmul": 0.0021431446075439453,
+      "RewriteWeights": 0.0024466514587402344,
+      "SFKVectorizer": 0.12934255599975586,
+      "SimpleAllReduceTiling": 0.0027997493743896484,
+      "Simplifier": 0.0042324066162109375,
+      "SimplifyMacroPredicates": 0.0056989192962646484,
+      "SimplifyNeuronTensor": 0.2756993770599365,
+      "SimplifySlice": 0.0008549690246582031,
+      "SimplifyTensor": 0.004892826080322266,
+      "SpillPSum": 0.009497642517089844,
+      "SplitAPUnionSets": 0.013408899307250977,
+      "SplitAccGrp": 0.0011219978332519531,
+      "StaticProfiler": 0.0052869319915771484,
+      "StaticTransposeLocalTensor": 0.003991365432739258,
+      "SundaISel": 0.07178783416748047,
+      "TCTransform": 0.0011143684387207031,
+      "TensorInitialization": 0.03500223159790039,
+      "TensorOpSimplifier": 0.00494384765625,
+      "TensorOpTransform": 0.01940751075744629,
+      "TileCCOps": 0.007318258285522461,
+      "TilingProfiler": 0.007637739181518555,
+      "TransformConvOp": 0.002967357635498047,
+      "TritiumFusion": 0.06856966018676758,
+      "ValueNumbering": 0.001973867416381836,
+      "VectorizeDMA": 0.002585887908935547,
+      "VectorizeMatMult": 0.004343986511230469,
+      "WeightCoalescing": 0.003340482711791992,
+      "ZeroSizeTensorElimination": 0.00018334388732910156
+    },
+    "tensorizer": {
+      "DMATilingProfiler::TotalInstructionsAfterTiling": 9788.0,
+      "StaticProfiler::AifUb": 148.07814025878906,
+      "StaticProfiler::ArithmeticIntensityTensorizer": 144.15213012695313,
+      "StaticProfiler::AverageDmaLength": 6822.7958984375,
+      "StaticProfiler::AverageFractalPeUtilization": 99.64185333251953,
+      "StaticProfiler::AveragePartitionUtilization": 99.05009460449219,
+      "StaticProfiler::AveragePeUtilization": 98.57437896728516,
+      "StaticProfiler::DDRTransferBytes": 323193888.0,
+      "StaticProfiler::InternalTransferBytes": 12166164.0,
+      "StaticProfiler::LoadExpanded": 40466.0,
+      "StaticProfiler::LocalizationEfficiency": 97.34869384765625,
+      "StaticProfiler::LocalizationEfficiencyIgnoreNonlocal": 101.34577178955078,
+      "StaticProfiler::StoreExpanded": 1835.0,
+      "StaticProfiler::TotalDMAExpanded": 42301.0,
+      "StaticProfiler::TotalDynamicInstancesCount": 10024.0,
+      "StaticProfiler::TotalDynamicInstancesWithMmPackedCount": 9928.0,
+      "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": 8576.0,
+      "TilingProfiler::NumPfTransposes": 4.0,
+      "TilingProfiler::NumPfTransposesForIo": 0.0,
+      "TilingProfiler::NumPfTransposesForLocal": 1.0,
+      "TilingProfiler::NumPfTransposesForNonlocal": 3.0,
+      "TilingProfiler::PfTransposeInstructions": 193.0,
+      "TilingProfiler::PfTransposeInstructionsForIo": 0.0,
+      "TilingProfiler::PfTransposeInstructionsForLocal": 1.0,
+      "TilingProfiler::PfTransposeInstructionsForNonlocal": 192.0,
+      "TilingProfiler::ReduceInstructionsAfterTiling": 4.0,
+      "TilingProfiler::SimdInstructionsAfterTiling": 166.0,
+      "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": 1.5999999959603883e-05,
+      "CanonicalizeForTensorizer": 1.2000000424450263e-05,
+      "Canonicalizer": 0.00024399999529123306,
+      "HoistCompute": 1.9999999949504854e-06,
+      "IdentifyCrossPassTensors": 9.999999747378752e-06,
+      "MemcastMotion": 1.1000000085914508e-05,
+      "PenguinizeFunctions": 1.2999999853491317e-05,
+      "PruneFunctions": 1.2999999853491317e-05,
+      "RemoveOptimizationBarriers": 1.1000000085914508e-05,
+      "ScatterMotion": 2.700000004551839e-05,
+      "TensorizerLegalizationPass": 1.4000000192027073e-05,
+      "VerifySupportedOps": 9.000000318337698e-06,
+      "algsimp": 9.200000204145908e-05,
+      "batchnorm_expander": 1.4000000192027073e-05,
+      "boundary-marker-removal": 3.999999989900971e-06,
+      "call-inliner": 9.000000318337698e-06,
+      "canonicalize-boundary-marker": 4.999999873689376e-06,
+      "collective-stream-id-checker": 7.000000096013537e-06,
+      "comparison-expander": 3.999999989900971e-06,
+      "computation-deduplicator": 1.9999999494757503e-05,
+      "conditional-to-select": 4.999999873689376e-06,
+      "config-lowering": 4.5000000682193786e-05,
+      "constant_folding": 9.000000318337698e-06,
+      "cse": 1.2000000424450263e-05,
+      "dce": 9.999999974752427e-07,
+      "dynamic-slice-transpose": 4.999999873689376e-06,
+      "eliminate-redundant-compare": 3.999999989900971e-06,
+      "emit-offloaded-dropout": 2.099999983329326e-05,
+      "flatten-call-graph": 7.999999979801942e-06,
+      "fuse-send-recv": 2.700000004551839e-05,
+      "hilo::LegalizeAlias": 3.999999989900971e-06,
+      "hilo::NeuronInstCombine": 5.6000000768108293e-05,
+      "hilo::NeuronOpFusion": 1.5999999959603883e-05,
+      "hilo::ReplaceTokenTypeWithU8Pass": 6.000000212225132e-06,
+      "hilo::ScheduleFusion": 9.999999974752427e-07,
+      "hilo::SixtyFourHack": 9.999999747378752e-06,
+      "hilo::VerifyAliasing": 1.9999999949504854e-06,
+      "hlo-mac-count": 3.600000127335079e-05,
+      "hlo-verifier": 0.00019299999985378236,
+      "legalize-ccops": 9.999999974752427e-07,
+      "legalize-compare": 3.999999989900971e-06,
+      "lower-argminmax-custom-call": 3.000000106112566e-06,
+      "map-inline": 1.2000000424450263e-05,
+      "metadata-naming": 2.4000000848900527e-05,
+      "mlir::detail::OpToOpPassAdaptor": 2.9000000722589903e-05,
+      "mlir::hlo::MhloToPyPenguin": 0.023865999653935432,
+      "mlir::mhlo::LowerComplexExtraPass": 0.0001049999991664663,
+      "mlir::mhlo::LowerComplexPass": 0.00016999999934341758,
+      "native-to-custom-softmax": 9.999999747378752e-06,
+      "native-to-custom-softmax-dx": 3.300000025774352e-05,
+      "operand_upcaster": 1.9999999494757503e-05,
+      "post-par-pipe-begin": 1.9999999949504854e-06,
+      "post-par-pipe-end": 0.0,
+      "post-partition-simplification": 0.000577000027988106,
+      "replace-minimum-constant": 7.999999979801942e-06,
+      "reshape-mover": 3.000000106112566e-06,
+      "simplify-concat": 5.400000009103678e-05,
+      "simplify-while-loops": 1.9999999949504854e-06,
+      "transform-variadic-reduce": 9.000000318337698e-06,
+      "tuple-simplifier": 4.999999873689376e-06,
+      "unpack-nested-aws-ntwsr": 3.000000106112566e-06,
+      "unroll-while-loop": 0.0
+    },
+    "hilo": {
+      "ArithmeticIntensity": 227.59315490722656,
+      "HloMacCount": 28185722880.0,
+      "Traffic": 247685152.0
+    }
+  },
+  "sg02": {
+    "compiletime": {
+      "CanonicalizeConv": 1.9999999494757503e-05,
+      "CanonicalizeForTensorizer": 1.4000000192027073e-05,
+      "Canonicalizer": 0.0003480000013951212,
+      "HoistCompute": 9.999999974752427e-07,
+      "IdentifyCrossPassTensors": 1.2000000424450263e-05,
+      "MemcastMotion": 0.0,
+      "PenguinizeFunctions": 9.000000318337698e-06,
+      "PruneFunctions": 6.000000212225132e-06,
+      "RemoveOptimizationBarriers": 1.4000000192027073e-05,
+      "ScatterMotion": 9.999999974752427e-07,
+      "TensorizerLegalizationPass": 6.000000212225132e-06,
+      "VerifySupportedOps": 1.1000000085914508e-05,
+      "algsimp": 6.800000119255856e-05,
+      "batchnorm_expander": 1.2999999853491317e-05,
+      "boundary-marker-removal": 3.000000106112566e-06,
+      "call-inliner": 9.999999747378752e-06,
+      "canonicalize-boundary-marker": 3.999999989900971e-06,
+      "collective-stream-id-checker": 3.999999989900971e-06,
+      "comparison-expander": 4.999999873689376e-06,
+      "computation-deduplicator": 1.9999999494757503e-05,
+      "conditional-to-select": 6.000000212225132e-06,
+      "config-lowering": 5.999999848427251e-05,
+      "constant_folding": 9.000000318337698e-06,
+      "cse": 1.2999999853491317e-05,
+      "dce": 9.999999974752427e-07,
+      "dynamic-slice-transpose": 4.999999873689376e-06,
+      "eliminate-redundant-compare": 3.000000106112566e-06,
+      "emit-offloaded-dropout": 1.9999999494757503e-05,
+      "flatten-call-graph": 1.1000000085914508e-05,
+      "fuse-send-recv": 1.8999999156221747e-05,
+      "hilo::LegalizeAlias": 1.9999999949504854e-06,
+      "hilo::NeuronInstCombine": 4.999999873689376e-06,
+      "hilo::NeuronOpFusion": 9.999999747378752e-06,
+      "hilo::ReplaceTokenTypeWithU8Pass": 7.000000096013537e-06,
+      "hilo::ScheduleFusion": 1.9999999949504854e-06,
+      "hilo::SixtyFourHack": 3.899999865097925e-05,
+      "hilo::VerifyAliasing": 9.999999974752427e-07,
+      "hlo-mac-count": 0.00019700000120792538,
+      "hlo-verifier": 0.00016599999798927456,
+      "legalize-ccops": 9.999999974752427e-07,
+      "legalize-compare": 3.000000106112566e-06,
+      "lower-argminmax-custom-call": 3.000000106112566e-06,
+      "map-inline": 1.2000000424450263e-05,
+      "metadata-naming": 1.5999999959603883e-05,
+      "mlir::detail::OpToOpPassAdaptor": 1.8000000636675395e-05,
+      "mlir::hlo::MhloToPyPenguin": 0.016130000352859497,
+      "mlir::mhlo::LowerComplexExtraPass": 9.7999996796716e-05,
+      "mlir::mhlo::LowerComplexPass": 0.0001740000006975606,
+      "native-to-custom-softmax": 9.000000318337698e-06,
+      "native-to-custom-softmax-dx": 2.9999999242136255e-05,
+      "operand_upcaster": 1.2999999853491317e-05,
+      "post-par-pipe-begin": 1.9999999949504854e-06,
+      "post-par-pipe-end": 0.0,
+      "post-partition-simplification": 0.0005070000188425183,
+      "replace-minimum-constant": 9.999999747378752e-06,
+      "reshape-mover": 3.000000106112566e-06,
+      "simplify-concat": 4.999999873689376e-05,
+      "simplify-while-loops": 1.9999999949504854e-06,
+      "transform-variadic-reduce": 4.5000000682193786e-05,
+      "tuple-simplifier": 4.999999873689376e-06,
+      "unpack-nested-aws-ntwsr": 3.999999989900971e-06,
+      "unroll-while-loop": 0.0
+    },
+    "hilo": {
+      "ArithmeticIntensity": 142.82192993164063,
+      "HloMacCount": 22615687168.0,
+      "Traffic": 316697664.0
+    }
+  }
+}

context_encoding_model/_tp0_bk1/neuron_config.json

@@ -0,0 +1,213 @@
+{
+  "_attn_implementation_autoset": false,
+  "_name_or_path": "mistralai/Mistral-7B-Instruct-v0.3",
+  "add_cross_attention": false,
+  "architectures": [
+    "MistralForCausalLM"
+  ],
+  "attention_dropout": 0.0,
+  "attribute_map": {},
+  "bad_words_ids": null,
+  "begin_suppress_tokens": null,
+  "bos_token_id": 1,
+  "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": 2,
+  "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": 14336,
+  "is_decoder": false,
+  "is_encoder_decoder": false,
+  "label2id": {
+    "LABEL_0": 0,
+    "LABEL_1": 1
+  },
+  "length_penalty": 1.0,
+  "max_length": 20,
+  "max_position_embeddings": 32768,
+  "metadata": null,
+  "min_length": 0,
+  "model_type": "mistral",
+  "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": "NeuronLlamaAttention",
+    "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": 4,
+    "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": 4,
+    "max_context_length": 2048,
+    "max_length": 2048,
+    "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": 2048,
+    "n_positions": 2048,
+    "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": 2048,
+    "pa_num_blocks": 4,
+    "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": 2048,
+    "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": 4,
+    "speculation_length": 0,
+    "start_rank_id": 0,
+    "target": null,
+    "tile_cc": false,
+    "tkg_batch_size": 4,
+    "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": 32,
+  "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-05,
+  "rope_theta": 1000000.0,
+  "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": "4.42.0.dev0",
+  "typical_p": 1.0,
+  "use_bfloat16": false,
+  "use_cache": true,
+  "vocab_size": 32768
+}

context_encoding_model/_tp0_bk2/command.txt

@@ -0,0 +1 @@
+neuronx-cc compile --framework=XLA model.MODULE_78e5291800ea5b96a03b+442879bd.hlo_module.pb --output model.MODULE_78e5291800ea5b96a03b+442879bd.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_78e5291800ea5b96a03b+442879bd.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=/models/mistral-7b-v0.3-instruct-neuronx/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": 99.68099975585938,
+      "StaticProfiler::AveragePartitionUtilization": 99.27783203125,
+      "StaticProfiler::AveragePeUtilization": 98.72872924804688,
+      "StaticProfiler::LocalizationEfficiency": 96.15084838867188,
+      "StaticProfiler::LocalizationEfficiencyIgnoreNonlocal": 102.69681549072266,
+      "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": 0.018404245376586914,
+      "AffinePredicateResolution": 0.0012030601501464844,
+      "AliasDependencyElimination": 0.00032448768615722656,
+      "AliasDependencyInduction": 0.006628751754760742,
+      "AliasDependencyReset": 0.04145312309265137,
+      "BFComputeCutting": 0.001703023910522461,
+      "BirCodeGenLoop": 0.21541404724121094,
+      "CCOpFusion": 0.017800331115722656,
+      "CanonicalizeConv": 1.5999999959603883e-05,
+      "CanonicalizeDAGForPGTiling": 0.004300355911254883,
+      "CanonicalizeForTensorizer": 5.0000002374872565e-05,
+      "CanonicalizeIR": 0.0014739036560058594,
+      "Canonicalizer": 0.000981000019237399,
+      "CoalesceCCOp": 0.006689786911010742,
+      "CommuteConcat": 0.0008399486541748047,
+      "DMALocalityOpt": 0.0020704269409179688,
+      "DMAProfiler": 0.0075225830078125,
+      "DMATilingProfiler": 0.04760599136352539,
+      "DataLocalityOpt": 0.12700510025024414,
+      "DataStreaming": 0.006787538528442383,
+      "DeConcat": 0.00067138671875,
+      "DeadCodeElimination": 0.0010039806365966797,
+      "DeadStoreElimination": 0.006083250045776367,
+      "DelinearIndices": 0.004990339279174805,
+      "Delinearization": 0.0024585723876953125,
+      "DoNothing": 0.0002791881561279297,
+      "DramToDramTranspose": 0.01742863655090332,
+      "DumpGraphAndMetadata": 0.022494077682495117,
+      "EliminateDivs": 0.0031578540802001953,
+      "ExpandBatchNorm": 0.0016427040100097656,
+      "ExpandISAMacro": 0.005059242248535156,
+      "FactorizeBlkDims": 0.016604900360107422,
+      "FactorizeThreadAxesInFreeDims": 0.0011897087097167969,
+      "FlattenMacroLoop": 0.0030732154846191406,
+      "GenericAccessSimplifier": 0.0007944107055664063,
+      "HoistCompute": 1.4000000192027073e-05,
+      "IdentifyCrossPassTensors": 3.899999865097925e-05,
+      "InferInitValue": 0.08164691925048828,
+      "InferIntrinsicOnCC": 0.00891733169555664,
+      "InferNeuronTensor": 0.021415233612060547,
+      "InferNonlocalTensors": 0.020433664321899414,
+      "InferPSumTensor": 0.06488466262817383,
+      "InlineNativeKernels": 0.002740621566772461,
+      "InsertIOTransposes": 0.01563739776611328,
+      "InsertLocalTransposes": 0.004204511642456055,
+      "InsertOffloadedTransposes": 0.0030889511108398438,
+      "LICM": 0.0025815963745117188,
+      "LateLegalizeInst": 0.007286548614501953,
+      "LateLegalizePostSplit": 0.004240512847900391,
+      "LateLowerReshapeOp": 0.0011327266693115234,
+      "LateLowerTensorOp": 0.0026912689208984375,
+      "LateNeuronInstComb": 0.060045719146728516,
+      "LayoutPreprocessing": 0.0245664119720459,
+      "LayoutPreprocessingAndAnalysis": 0.07019805908203125,
+      "LayoutRequirementAnalysis": 0.004818916320800781,
+      "LegalizeCCOpLayout": 0.0020072460174560547,
+      "LegalizeOpLevelAlias": 0.0011057853698730469,
+      "LegalizePartitionReduce": 0.0015609264373779297,
+      "LegalizeSundaAccess": 0.0679481029510498,
+      "LegalizeSundaMacro": 0.018359661102294922,
+      "LegalizeType": 0.006083011627197266,
+      "LocalLayoutOpt": 0.014153003692626953,
+      "LoopFusion": 0.005875825881958008,
+      "LoopSplitting": 0.0004029273986816406,
+      "LowerBroadcast": 0.0018463134765625,
+      "LowerCCOpBlockAxis": 0.003938198089599609,
+      "LowerComplexBroadcast": 0.004299163818359375,
+      "LowerIntrinsics": 0.26036787033081055,
+      "LowerTensorOp": 0.010293006896972656,
+      "LowerTranspose": 0.009520769119262695,
+      "MacroGeneration": 0.03194475173950195,
+      "MaskPropagation": 0.0027740001678466797,
+      "MemcastMotion": 1.1000000085914508e-05,
+      "MemcpyElimination": 0.02656841278076172,
+      "MutateDataType": 0.0012118816375732422,
+      "NeuronAliasDependencyInduction": 0.0003993511199951172,
+      "NeuronAliasDependencyReset": 0.017134428024291992,
+      "NeuronInstComb": 0.005475759506225586,
+      "NeuronLICM": 0.013751506805419922,
+      "NeuronLoopFusion": 0.011570215225219727,
+      "NeuronLoopInterchange": 0.002625703811645508,
+      "NeuronSimplifier": 0.012082099914550781,
+      "NeuronSimplifyPredicates": 0.006382942199707031,
+      "NeuronValueNumbering": 0.005650520324707031,
+      "OptimizeAliasedCopyChain": 0.0005443096160888672,
+      "OptimizeNKIKernels": 0.6168310642242432,
+      "PAGLayoutOpt": 0.10135579109191895,
+      "PComputeCutting": 0.004828453063964844,
+      "PGLayoutTilingPipeline": 0.6252057552337646,
+      "PGTiling": 0.1626896858215332,
+      "PadElimination": 0.00044655799865722656,
+      "ParAxesAnnotation": 0.07457947731018066,
+      "PartialLoopFusion": 0.047167062759399414,
+      "PartialSimdFusion": 0.006536245346069336,
+      "PenguinizeFunctions": 4.3000000005122274e-05,
+      "PerfectLoopNest": 0.003509998321533203,
+      "PruneFunctions": 1.8999999156221747e-05,
+      "RecognizeOpIdiom": 0.0035545825958251953,
+      "Recompute": 0.00031566619873046875,
+      "RelaxPredicates": 0.004423379898071289,
+      "Rematerialization": 0.001989126205444336,
+      "RemoveOptimizationBarriers": 4.400000034365803e-05,
+      "ReshapeWeights": 0.0011234283447265625,
+      "ResolveAccessConflict": 0.003659963607788086,
+      "ResolveComplicatePredicates": 0.0010039806365966797,
+      "RewriteReplicationMatmul": 0.0014824867248535156,
+      "RewriteWeights": 0.0052568912506103516,
+      "SFKVectorizer": 0.18593907356262207,
+      "ScatterMotion": 7.000000096013537e-06,
+      "SimpleAllReduceTiling": 0.0033435821533203125,
+      "Simplifier": 0.003216266632080078,
+      "SimplifyMacroPredicates": 0.005851030349731445,
+      "SimplifyNeuronTensor": 0.2244417667388916,
+      "SimplifySlice": 0.0011301040649414063,
+      "SimplifyTensor": 0.009292364120483398,
+      "SpillPSum": 0.01206660270690918,
+      "SplitAPUnionSets": 0.03329944610595703,
+      "SplitAccGrp": 0.0012478828430175781,
+      "StaticProfiler": 0.005914926528930664,
+      "StaticTransposeLocalTensor": 0.0035593509674072266,
+      "SundaISel": 0.09800505638122559,
+      "TCTransform": 0.0008752346038818359,
+      "TensorInitialization": 0.004550933837890625,
+      "TensorOpSimplifier": 0.004429817199707031,
+      "TensorOpTransform": 0.035314083099365234,
+      "TensorizerLegalizationPass": 4.70000013592653e-05,
+      "TileCCOps": 0.006109952926635742,
+      "TilingProfiler": 0.0074079036712646484,
+      "TransformConvOp": 0.0027616024017333984,
+      "TritiumFusion": 0.07009291648864746,
+      "ValueNumbering": 0.0019168853759765625,
+      "VectorizeDMA": 0.002183198928833008,
+      "VectorizeMatMult": 0.0026412010192871094,
+      "VerifySupportedOps": 3.9999998989515007e-05,
+      "WeightCoalescing": 0.003337383270263672,
+      "ZeroSizeTensorElimination": 0.00018453598022460938,
+      "algsimp": 0.002238999819383025,
+      "batchnorm_expander": 4.099999932805076e-05,
+      "boundary-marker-removal": 1.1000000085914508e-05,
+      "call-inliner": 0.00045200000749900937,
+      "canonicalize-boundary-marker": 1.4999999621068127e-05,
+      "collective-stream-id-checker": 7.100000220816582e-05,
+      "comparison-expander": 0.0004180000105407089,
+      "computation-deduplicator": 6.399999983841553e-05,
+      "conditional-to-select": 1.5999999959603883e-05,
+      "config-lowering": 0.00016199999663513154,
+      "constant-statistics": 0.0004290000069886446,
+      "constant_folding": 0.0003000000142492354,
+      "cse": 3.9999998989515007e-05,
+      "dce": 8.099999831756577e-05,
+      "dot_decomposer": 0.0011439999798312783,
+      "dynamic-slice-transpose": 1.3999999282532372e-05,
+      "eliminate-redundant-compare": 0.0002579999854788184,
+      "emit-offloaded-dropout": 6.70000008540228e-05,
+      "flatten-call-graph": 0.0006840000278316438,
+      "fuse-send-recv": 6.70000008540228e-05,
+      "hilo::LegalizeAlias": 1.3999999282532372e-05,
+      "hilo::NeuronInstCombine": 0.00015199999324977398,
+      "hilo::NeuronOpFusion": 2.700000004551839e-05,
+      "hilo::ReplaceTokenTypeWithU8Pass": 4.3000000005122274e-05,
+      "hilo::ScheduleFusion": 2.9000000722589903e-05,
+      "hilo::SixtyFourHack": 8.199999865610152e-05,
+      "hilo::VerifyAliasing": 6.000000212225132e-06,
+      "hlo-mac-count": 0.0011340000201016665,
+      "hlo-verifier": 0.006370999850332737,
+      "instruction-histogram": 0.001025000005029142,
+      "io-con-pipe-begin": 3.999999989900971e-06,
+      "io-con-pipe-end": 9.999999974752427e-07,
+      "io-layout-normalization": 0.00108299998100847,
+      "io-statistics": 9.600000339560211e-05,
+      "legalize-ccops": 3.999999989900971e-06,
+      "legalize-compare": 1.1999999514955562e-05,
+      "lower-argminmax-custom-call": 9.000000318337698e-06,
+      "map-inline": 0.00063699996098876,
+      "metadata-naming": 4.8000001697801054e-05,
+      "mlir::detail::OpToOpPassAdaptor": 7.599999662488699e-05,
+      "mlir::hlo::MhloToPyPenguin": 0.0729999989271164,
+      "mlir::mhlo::LowerComplexExtraPass": 0.0003330000035930425,
+      "mlir::mhlo::LowerComplexPass": 0.00041099998634308577,
+      "native-to-custom-softmax": 0.0005629999795928597,
+      "native-to-custom-softmax-dx": 0.000522000016644597,
+      "operand_upcaster": 7.500000356230885e-05,
+      "opt-barrier-removal": 0.0004529999860096723,
+      "post-par-pipe-begin": 4.999999873689376e-06,
+      "post-par-pipe-end": 0.0,
+      "post-partition-simplification": 0.001686000032350421,
+      "pre-par-pipe-begin": 9.999999974752427e-07,
+      "pre-par-pipe-end": 0.0,
+      "pre-partition-simplification": 0.11061400175094604,
+      "replace-minimum-constant": 0.00034699999378062785,
+      "reshape-mover": 0.00011300000187475234,
+      "simplify-concat": 0.000155999994603917,
+      "simplify-while-loops": 9.800000407267362e-05,
+      "transform-variadic-reduce": 7.700000423938036e-05,
+      "tuple-simplifier": 0.00028600002406165004,
+      "unpack-nested-aws-ntwsr": 0.000311999989207834,
+      "unroll-while-loop": 1.8000000636675395e-05,
+      "zero_sized_hlo_elimination": 0.000653000024612993
+    },
+    "hilo": {
+      "ConstantSize": 1057749.0,
+      "HloInputCount": 359.0,
+      "HloMacCount": 113883742208.0,
+      "HloOutputCount": 65.0,
+      "IfmapSize": 7785164800.0,
+      "OfmapSize": 536870912.0,
+      "OutputsReadFromCount": 0.0,
+      "PassthroughTensorsCount": 0.0,
+      "RedundantOutputCount": 0.0,
+      "Traffic": 795959680.0
+    },
+    "tensorizer": {
+      "DMATilingProfiler::TotalInstructionsAfterTiling": 11124.0,
+      "StaticProfiler::AifUb": 291.8241882324219,
+      "StaticProfiler::ArithmeticIntensityTensorizer": 280.5914306640625,
+      "StaticProfiler::AverageDmaLength": 1148.8248291015625,
+      "StaticProfiler::DDRTransferBytes": 331583520.0,
+      "StaticProfiler::InternalTransferBytes": 24224788.0,
+      "StaticProfiler::LoadExpanded": 293906.0,
+      "StaticProfiler::StoreExpanded": 3627.0,
+      "StaticProfiler::TotalDMAExpanded": 297533.0,
+      "StaticProfiler::TotalDynamicInstancesCount": 13170.0,
+      "StaticProfiler::TotalDynamicInstancesWithMmPackedCount": 13074.0,
+      "StaticProfiler::TotalLNCComm": 0.0,
+      "StaticProfiler::TotalLNCCommTransfer": 0.0,
+      "TilingProfiler::BatchnormInstructionsAfterTiling": 0.0,
+      "TilingProfiler::DmaInstructionsAfterTiling": 0.0,
+      "TilingProfiler::GenericInstructionsAfterTiling": 4.0,
+      "TilingProfiler::MatMultInstructionsAfterTiling": 9472.0,
+      "TilingProfiler::NumPfTransposes": 4.0,
+      "TilingProfiler::NumPfTransposesForIo": 0.0,
+      "TilingProfiler::NumPfTransposesForLocal": 1.0,
+      "TilingProfiler::NumPfTransposesForNonlocal": 3.0,
+      "TilingProfiler::PfTransposeInstructions": 385.0,
+      "TilingProfiler::PfTransposeInstructionsForIo": 0.0,
+      "TilingProfiler::PfTransposeInstructionsForLocal": 1.0,
+      "TilingProfiler::PfTransposeInstructionsForNonlocal": 384.0,
+      "TilingProfiler::ReduceInstructionsAfterTiling": 4.0,
+      "TilingProfiler::SimdInstructionsAfterTiling": 166.0,
+      "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": 0.0020379999186843634,
+      "call-inliner": 0.0004239999980200082,
+      "collective-stream-id-checker": 5.8000001445179805e-05,
+      "comparison-expander": 0.0004030000127386302,
+      "constant-statistics": 0.0004290000069886446,
+      "constant_folding": 0.0002739999908953905,
+      "dce": 7.79999973019585e-05,
+      "dot_decomposer": 0.0011439999798312783,
+      "eliminate-redundant-compare": 0.0002479999966453761,
+      "flatten-call-graph": 0.0006559999892488122,
+      "hlo-mac-count": 0.0008549999911338091,
+      "hlo-verifier": 0.005834000185132027,
+      "instruction-histogram": 0.001025000005029142,
+      "io-con-pipe-begin": 3.999999989900971e-06,
+      "io-con-pipe-end": 9.999999974752427e-07,
+      "io-layout-normalization": 0.00108299998100847,
+      "io-statistics": 9.600000339560211e-05,
+      "map-inline": 0.0006009999779053032,
+      "native-to-custom-softmax": 0.0005319999763742089,
+      "native-to-custom-softmax-dx": 0.0003640000068116933,
+      "opt-barrier-removal": 0.0004529999860096723,
+      "pre-par-pipe-begin": 9.999999974752427e-07,
+      "pre-par-pipe-end": 0.0,
+      "pre-partition-simplification": 0.11061400175094604,
+      "replace-minimum-constant": 0.0003220000071451068,
+      "reshape-mover": 0.00010199999815085903,
+      "simplify-while-loops": 9.200000204145908e-05,
+      "tuple-simplifier": 0.00027200000477023423,
+      "unpack-nested-aws-ntwsr": 0.0003009999927598983,
+      "unroll-while-loop": 1.8000000636675395e-05,
+      "zero_sized_hlo_elimination": 0.000653000024612993
+    }
+  },
+  "cumsum": {
+    "compiletime": {
+      "CoalesceCCOp": 0.0002663135528564453,
+      "DMALocalityOpt": 0.00022459030151367188,
+      "DMAProfiler": 0.0010578632354736328,
+      "DataStreaming": 0.0003883838653564453,
+      "DoNothing": 0.00015854835510253906,
+      "ExpandISAMacro": 0.0006091594696044922,
+      "FactorizeBlkDims": 0.00047659873962402344,
+      "InferPSumTensor": 0.0010039806365966797,
+      "LateLegalizeInst": 0.0005145072937011719,
+      "LateNeuronInstComb": 0.0005576610565185547,
+      "LegalizeSundaAccess": 0.0016772747039794922,
+      "LegalizeType": 0.00028705596923828125,
+      "LowerBroadcast": 0.0002655982971191406,
+      "LowerIntrinsics": 0.0002892017364501953,
+      "LowerTranspose": 0.0002830028533935547,
+      "NeuronInstComb": 0.0005538463592529297,
+      "NeuronLICM": 0.00041985511779785156,
+      "NeuronSimplifyPredicates": 0.003001689910888672,
+      "NeuronValueNumbering": 0.0004851818084716797,
+      "SFKVectorizer": 0.003008604049682617,
+      "SimpleAllReduceTiling": 0.00024580955505371094,
+      "SimplifyNeuronTensor": 0.0004801750183105469,
+      "SpillPSum": 0.0005950927734375,
+      "WeightCoalescing": 0.00028634071350097656
+    }
+  },
+  "sg00": {
+    "compiletime": {
+      "CanonicalizeConv": 7.999999979801942e-06,
+      "CanonicalizeForTensorizer": 1.4999999621068127e-05,
+      "Canonicalizer": 0.00033999999868683517,
+      "HoistCompute": 3.000000106112566e-06,
+      "IdentifyCrossPassTensors": 1.2000000424450263e-05,
+      "MemcastMotion": 3.999999989900971e-06,
+      "PenguinizeFunctions": 1.4999999621068127e-05,
+      "PruneFunctions": 4.999999873689376e-06,
+      "RemoveOptimizationBarriers": 1.4000000192027073e-05,
+      "ScatterMotion": 0.0,
+      "TensorizerLegalizationPass": 2.2000000171829015e-05,
+      "VerifySupportedOps": 1.1000000085914508e-05,
+      "algsimp": 6.800000119255856e-05,
+      "batchnorm_expander": 1.4000000192027073e-05,
+      "boundary-marker-removal": 3.000000106112566e-06,
+      "call-inliner": 7.999999979801942e-06,
+      "canonicalize-boundary-marker": 4.999999873689376e-06,
+      "collective-stream-id-checker": 1.9999999949504854e-06,
+      "comparison-expander": 3.999999989900971e-06,
+      "computation-deduplicator": 1.4000000192027073e-05,
+      "conditional-to-select": 4.999999873689376e-06,
+      "config-lowering": 5.2999999752501026e-05,
+      "constant_folding": 7.999999979801942e-06,
+      "cse": 1.2999999853491317e-05,
+      "dce": 9.999999974752427e-07,
+      "dynamic-slice-transpose": 4.999999873689376e-06,
+      "eliminate-redundant-compare": 3.000000106112566e-06,
+      "emit-offloaded-dropout": 2.2000000171829015e-05,
+      "flatten-call-graph": 9.000000318337698e-06,
+      "fuse-send-recv": 1.9999999494757503e-05,
+      "hilo::LegalizeAlias": 4.999999873689376e-06,
+      "hilo::NeuronInstCombine": 8.900000102585182e-05,
+      "hilo::NeuronOpFusion": 1.8000000636675395e-05,
+      "hilo::ReplaceTokenTypeWithU8Pass": 1.4999999621068127e-05,
+      "hilo::ScheduleFusion": 6.000000212225132e-06,
+      "hilo::SixtyFourHack": 1.1000000085914508e-05,
+      "hilo::VerifyAliasing": 1.9999999949504854e-06,
+      "hlo-mac-count": 3.7999998312443495e-05,
+      "hlo-verifier": 0.00014899999951012433,
+      "legalize-ccops": 1.9999999949504854e-06,
+      "legalize-compare": 3.999999989900971e-06,
+      "lower-argminmax-custom-call": 3.000000106112566e-06,
+      "map-inline": 1.1000000085914508e-05,
+      "metadata-naming": 1.4999999621068127e-05,
+      "mlir::detail::OpToOpPassAdaptor": 2.099999983329326e-05,
+      "mlir::hlo::MhloToPyPenguin": 0.03171300143003464,
+      "mlir::mhlo::LowerComplexExtraPass": 0.00011700000322889537,
+      "mlir::mhlo::LowerComplexPass": 0.00014899999951012433,
+      "native-to-custom-softmax": 1.1000000085914508e-05,
+      "native-to-custom-softmax-dx": 9.100000170292333e-05,
+      "operand_upcaster": 2.4000000848900527e-05,
+      "post-par-pipe-begin": 9.999999974752427e-07,
+      "post-par-pipe-end": 0.0,
+      "post-partition-simplification": 0.0005540000274777412,
+      "replace-minimum-constant": 7.999999979801942e-06,
+      "reshape-mover": 3.999999989900971e-06,
+      "simplify-concat": 5.199999941396527e-05,
+      "simplify-while-loops": 1.9999999949504854e-06,
+      "transform-variadic-reduce": 9.000000318337698e-06,
+      "tuple-simplifier": 3.999999989900971e-06,
+      "unpack-nested-aws-ntwsr": 3.000000106112566e-06,
+      "unroll-while-loop": 0.0
+    },
+    "hilo": {
+      "ArithmeticIntensity": 111.23420715332031,
+      "ConstantSize": 1057749.0,
+      "HloInputCount": 359.0,
+      "HloMacCount": 11811160064.0,
+      "HloOutputCount": 65.0,
+      "IfmapSize": 7785164800.0,
+      "OfmapSize": 536870912.0,
+      "OutputsReadFromCount": 0.0,
+      "PassthroughTensorsCount": 0.0,
+      "RedundantOutputCount": 0.0,
+      "Traffic": 212365600.0
+    }
+  },
+  "sg0000": {
+    "compiletime": {
+      "AGOrderingAnalysisPass": 0.027149438858032227,
+      "AffinePredicateResolution": 0.0014154911041259766,
+      "AliasDependencyElimination": 0.00019693374633789063,
+      "AliasDependencyInduction": 0.008726119995117188,
+      "AliasDependencyReset": 0.11011934280395508,
+      "BFComputeCutting": 0.002534627914428711,
+      "BirCodeGenLoop": 0.04940390586853027,
+      "CCOpFusion": 0.03955817222595215,
+      "CanonicalizeDAGForPGTiling": 0.0033364295959472656,
+      "CanonicalizeIR": 0.0017902851104736328,
+      "CoalesceCCOp": 0.004707813262939453,
+      "CommuteConcat": 0.0011563301086425781,
+      "DMALocalityOpt": 0.0012772083282470703,
+      "DMAProfiler": 0.00436711311340332,
+      "DMATilingProfiler": 0.0046231746673583984,
+      "DataLocalityOpt": 0.12483954429626465,
+      "DataStreaming": 0.004174470901489258,
+      "DeConcat": 0.001886606216430664,
+      "DeadCodeElimination": 0.0011522769927978516,
+      "DeadStoreElimination": 0.040479183197021484,
+      "DelinearIndices": 0.006760358810424805,
+      "Delinearization": 0.003210306167602539,
+      "DoNothing": 0.00013256072998046875,
+      "DramToDramTranspose": 0.024677276611328125,
+      "DumpGraphAndMetadata": 0.012624502182006836,
+      "EliminateDivs": 0.0033044815063476563,
+      "ExpandBatchNorm": 0.0016093254089355469,
+      "ExpandISAMacro": 0.002683877944946289,
+      "FactorizeBlkDims": 0.02043914794921875,
+      "FactorizeThreadAxesInFreeDims": 0.0014965534210205078,
+      "FlattenMacroLoop": 0.005756378173828125,
+      "GenericAccessSimplifier": 0.0009303092956542969,
+      "InferInitValue": 0.057647705078125,
+      "InferIntrinsicOnCC": 0.009967327117919922,
+      "InferNeuronTensor": 0.037697553634643555,
+      "InferNonlocalTensors": 0.19877171516418457,
+      "InferPSumTensor": 0.05426788330078125,
+      "InlineNativeKernels": 0.0014045238494873047,
+      "InsertIOTransposes": 0.012149333953857422,
+      "InsertLocalTransposes": 0.006654024124145508,
+      "InsertOffloadedTransposes": 0.042548418045043945,
+      "LICM": 0.004395484924316406,
+      "LateLegalizeInst": 0.006590127944946289,
+      "LateLegalizePostSplit": 0.0027794837951660156,
+      "LateLowerReshapeOp": 0.021605253219604492,
+      "LateLowerTensorOp": 0.02600717544555664,
+      "LateNeuronInstComb": 0.03907942771911621,
+      "LayoutPreprocessing": 0.07296419143676758,
+      "LayoutPreprocessingAndAnalysis": 0.09938645362854004,
+      "LayoutRequirementAnalysis": 0.013081550598144531,
+      "LegalizeCCOpLayout": 0.002216339111328125,
+      "LegalizeOpLevelAlias": 0.001260519027709961,
+      "LegalizePartitionReduce": 0.0013256072998046875,
+      "LegalizeSundaAccess": 0.057663917541503906,
+      "LegalizeSundaMacro": 0.007729530334472656,
+      "LegalizeType": 0.003155231475830078,
+      "LocalLayoutOpt": 0.05669593811035156,
+      "LoopFusion": 0.007777690887451172,
+      "LoopSplitting": 0.0003819465637207031,
+      "LowerBroadcast": 0.0032148361206054688,
+      "LowerCCOpBlockAxis": 0.004834890365600586,
+      "LowerComplexBroadcast": 0.003167390823364258,
+      "LowerIntrinsics": 0.025065183639526367,
+      "LowerTensorOp": 0.011690616607666016,
+      "LowerTranspose": 0.026025772094726563,
+      "MacroGeneration": 0.19495820999145508,
+      "MaskPropagation": 0.00906229019165039,
+      "MemcpyElimination": 0.16968393325805664,
+      "MutateDataType": 0.0013375282287597656,
+      "NeuronAliasDependencyInduction": 0.00027489662170410156,
+      "NeuronAliasDependencyReset": 0.049456119537353516,
+      "NeuronInstComb": 0.013614177703857422,
+      "NeuronLICM": 0.007915735244750977,
+      "NeuronLoopFusion": 0.020767688751220703,
+      "NeuronLoopInterchange": 0.0029578208923339844,
+      "NeuronSimplifier": 0.015317201614379883,
+      "NeuronSimplifyPredicates": 0.005019426345825195,
+      "NeuronValueNumbering": 0.0039823055267333984,
+      "OptimizeAliasedCopyChain": 0.0007677078247070313,
+      "OptimizeNKIKernels": 0.0017440319061279297,
+      "PAGLayoutOpt": 0.6516931056976318,
+      "PComputeCutting": 0.008367061614990234,
+      "PGLayoutTilingPipeline": 1.5081679821014404,
+      "PGTiling": 0.33455348014831543,
+      "PadElimination": 0.0004961490631103516,
+      "ParAxesAnnotation": 0.39901041984558105,
+      "PartialLoopFusion": 0.024018049240112305,
+      "PartialSimdFusion": 0.012691020965576172,
+      "PerfectLoopNest": 0.001743316650390625,
+      "RecognizeOpIdiom": 0.004468679428100586,
+      "Recompute": 0.0003910064697265625,
+      "RelaxPredicates": 0.0034019947052001953,
+      "Rematerialization": 0.006110429763793945,
+      "ReshapeWeights": 0.0015015602111816406,
+      "ResolveAccessConflict": 0.0035288333892822266,
+      "ResolveComplicatePredicates": 0.00148773193359375,
+      "RewriteReplicationMatmul": 0.0015423297882080078,
+      "RewriteWeights": 0.0054814815521240234,
+      "SFKVectorizer": 0.27728796005249023,
+      "SimpleAllReduceTiling": 0.0017549991607666016,
+      "Simplifier": 0.00467371940612793,
+      "SimplifyMacroPredicates": 0.01883554458618164,
+      "SimplifyNeuronTensor": 0.03848075866699219,
+      "SimplifySlice": 0.0010480880737304688,
+      "SimplifyTensor": 0.0058362483978271484,
+      "SpillPSum": 0.017147064208984375,
+      "SplitAPUnionSets": 0.02936863899230957,
+      "SplitAccGrp": 0.0024695396423339844,
+      "StaticProfiler": 0.004050016403198242,
+      "StaticTransposeLocalTensor": 0.004538059234619141,
+      "SundaISel": 0.10911178588867188,
+      "TCTransform": 0.0014197826385498047,
+      "TensorInitialization": 0.009113788604736328,
+      "TensorOpSimplifier": 0.00628209114074707,
+      "TensorOpTransform": 0.027051925659179688,
+      "TileCCOps": 0.025897979736328125,
+      "TilingProfiler": 0.010443687438964844,
+      "TransformConvOp": 0.0024106502532958984,
+      "TritiumFusion": 0.15155267715454102,
+      "ValueNumbering": 0.0042994022369384766,
+      "VectorizeDMA": 0.006690025329589844,
+      "VectorizeMatMult": 0.06819844245910645,
+      "WeightCoalescing": 0.003052949905395508,
+      "ZeroSizeTensorElimination": 0.00017309188842773438
+    },
+    "tensorizer": {
+      "DMATilingProfiler::TotalInstructionsAfterTiling": 2978.0,
+      "StaticProfiler::AifUb": 120.88131713867188,
+      "StaticProfiler::ArithmeticIntensityTensorizer": 409.4889831542969,
+      "StaticProfiler::AverageDmaLength": 1928.296630859375,
+      "StaticProfiler::AverageFractalPeUtilization": 99.95079803466797,
+      "StaticProfiler::AveragePartitionUtilization": 99.86166381835938,
+      "StaticProfiler::AveragePeUtilization": 99.79874420166016,
+      "StaticProfiler::DDRTransferBytes": 64557320.0,
+      "StaticProfiler::InternalTransferBytes": 44957696.0,
+      "StaticProfiler::LoadExpanded": 23043.0,
+      "StaticProfiler::LocalizationEfficiency": 338.7529296875,
+      "StaticProfiler::LocalizationEfficiencyIgnoreNonlocal": 438.4559631347656,
+      "StaticProfiler::StoreExpanded": 5505.0,
+      "StaticProfiler::TotalDMAExpanded": 28548.0,
+      "StaticProfiler::TotalDynamicInstancesCount": 3406.0,
+      "StaticProfiler::TotalDynamicInstancesWithMmPackedCount": 3403.0,
+      "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": 1796.0,
+      "TilingProfiler::NumPfTransposes": 5.0,
+      "TilingProfiler::NumPfTransposesForIo": 1.0,
+      "TilingProfiler::NumPfTransposesForLocal": 3.0,
+      "TilingProfiler::NumPfTransposesForNonlocal": 1.0,
+      "TilingProfiler::PfTransposeInstructions": 528.0,
+      "TilingProfiler::PfTransposeInstructionsForIo": 128.0,
+      "TilingProfiler::PfTransposeInstructionsForLocal": 336.0,
+      "TilingProfiler::PfTransposeInstructionsForNonlocal": 64.0,
+      "TilingProfiler::ReduceInstructionsAfterTiling": 0.0,
+      "TilingProfiler::SimdInstructionsAfterTiling": 240.0,
+      "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": 0.11623430252075195,
+      "AffinePredicateResolution": 0.0013043880462646484,
+      "AliasDependencyElimination": 0.0002994537353515625,
+      "AliasDependencyInduction": 0.008661746978759766,
+      "AliasDependencyReset": 0.10846662521362305,
+      "BFComputeCutting": 0.002373218536376953,
+      "BirCodeGenLoop": 0.03558611869812012,
+      "CCOpFusion": 0.07521343231201172,
+      "CanonicalizeDAGForPGTiling": 0.003203153610229492,
+      "CanonicalizeIR": 0.001676797866821289,
+      "CoalesceCCOp": 0.00453639030456543,
+      "CommuteConcat": 0.0011301040649414063,
+      "DMALocalityOpt": 0.001081228256225586,
+      "DMAProfiler": 0.004008054733276367,
+      "DMATilingProfiler": 0.005870342254638672,
+      "DataLocalityOpt": 0.28061389923095703,
+      "DataStreaming": 0.003554821014404297,
+      "DeConcat": 0.0023250579833984375,
+      "DeadCodeElimination": 0.0011475086212158203,
+      "DeadStoreElimination": 0.03395438194274902,
+      "DelinearIndices": 0.008394002914428711,
+      "Delinearization": 0.0034415721893310547,
+      "DoNothing": 0.00012373924255371094,
+      "DramToDramTranspose": 0.08242917060852051,
+      "DumpGraphAndMetadata": 0.013036966323852539,
+      "EliminateDivs": 0.003282785415649414,
+      "ExpandBatchNorm": 0.001613616943359375,
+      "ExpandISAMacro": 0.0023279190063476563,
+      "FactorizeBlkDims": 0.013802051544189453,
+      "FactorizeThreadAxesInFreeDims": 0.002902507781982422,
+      "FlattenMacroLoop": 0.003271341323852539,
+      "GenericAccessSimplifier": 0.0010917186737060547,
+      "InferInitValue": 0.07010388374328613,
+      "InferIntrinsicOnCC": 0.009655475616455078,
+      "InferNeuronTensor": 0.13204073905944824,
+      "InferNonlocalTensors": 0.07003545761108398,
+      "InferPSumTensor": 0.02702808380126953,
+      "InlineNativeKernels": 0.0015347003936767578,
+      "InsertIOTransposes": 0.06269145011901855,
+      "InsertLocalTransposes": 0.03624296188354492,
+      "InsertOffloadedTransposes": 0.0031883716583251953,
+      "LICM": 0.03521370887756348,
+      "LateLegalizeInst": 0.004965066909790039,
+      "LateLegalizePostSplit": 0.002466440200805664,
+      "LateLowerReshapeOp": 0.0013051033020019531,
+      "LateLowerTensorOp": 0.005318164825439453,
+      "LateNeuronInstComb": 0.052803993225097656,
+      "LayoutPreprocessing": 0.09067535400390625,
+      "LayoutPreprocessingAndAnalysis": 0.11781954765319824,
+      "LayoutRequirementAnalysis": 0.007211446762084961,
+      "LegalizeCCOpLayout": 0.0017924308776855469,
+      "LegalizeOpLevelAlias": 0.0011508464813232422,
+      "LegalizePartitionReduce": 0.0017359256744384766,
+      "LegalizeSundaAccess": 0.015620946884155273,
+      "LegalizeSundaMacro": 0.018268585205078125,
+      "LegalizeType": 0.0046215057373046875,
+      "LocalLayoutOpt": 0.043848276138305664,
+      "LoopFusion": 0.0060999393463134766,
+      "LoopSplitting": 0.0003905296325683594,
+      "LowerBroadcast": 0.0017726421356201172,
+      "LowerCCOpBlockAxis": 0.004726886749267578,
+      "LowerComplexBroadcast": 0.0022859573364257813,
+      "LowerIntrinsics": 0.03918600082397461,
+      "LowerTensorOp": 0.010795831680297852,
+      "LowerTranspose": 0.05237984657287598,
+      "MacroGeneration": 0.09652948379516602,
+      "MaskPropagation": 0.0031325817108154297,
+      "MemcpyElimination": 0.17755985260009766,
+      "MutateDataType": 0.0014410018920898438,
+      "NeuronAliasDependencyInduction": 0.0003838539123535156,
+      "NeuronAliasDependencyReset": 0.013648271560668945,
+      "NeuronInstComb": 0.013708114624023438,
+      "NeuronLICM": 0.007151603698730469,
+      "NeuronLoopFusion": 0.025348663330078125,
+      "NeuronLoopInterchange": 0.001642465591430664,
+      "NeuronSimplifier": 0.02402472496032715,
+      "NeuronSimplifyPredicates": 0.001932382583618164,
+      "NeuronValueNumbering": 0.0032088756561279297,
+      "OptimizeAliasedCopyChain": 0.0006577968597412109,
+      "OptimizeNKIKernels": 0.0014371871948242188,
+      "PAGLayoutOpt": 0.4579346179962158,
+      "PComputeCutting": 0.01232600212097168,
+      "PGLayoutTilingPipeline": 1.3759241104125977,
+      "PGTiling": 0.3235757350921631,
+      "PadElimination": 0.00040984153747558594,
+      "ParAxesAnnotation": 0.3829355239868164,
+      "PartialLoopFusion": 0.013192892074584961,
+      "PartialSimdFusion": 0.04269862174987793,
+      "PerfectLoopNest": 0.00379180908203125,
+      "RecognizeOpIdiom": 0.008028745651245117,
+      "Recompute": 0.00033402442932128906,
+      "RelaxPredicates": 0.002958059310913086,
+      "Rematerialization": 0.002005338668823242,
+      "ReshapeWeights": 0.0014390945434570313,
+      "ResolveAccessConflict": 0.004492282867431641,
+      "ResolveComplicatePredicates": 0.0012967586517333984,
+      "RewriteReplicationMatmul": 0.002475738525390625,
+      "RewriteWeights": 0.005257844924926758,
+      "SFKVectorizer": 0.19685626029968262,
+      "SimpleAllReduceTiling": 0.0014569759368896484,
+      "Simplifier": 0.004656791687011719,
+      "SimplifyMacroPredicates": 0.006787300109863281,
+      "SimplifyNeuronTensor": 0.006094932556152344,
+      "SimplifySlice": 0.0012688636779785156,
+      "SimplifyTensor": 0.0069119930267333984,
+      "SpillPSum": 0.020178794860839844,
+      "SplitAPUnionSets": 0.016407012939453125,
+      "SplitAccGrp": 0.0013232231140136719,
+      "StaticProfiler": 0.003663778305053711,
+      "StaticTransposeLocalTensor": 0.004960536956787109,
+      "SundaISel": 0.04571676254272461,
+      "TCTransform": 0.001123189926147461,
+      "TensorInitialization": 0.002666473388671875,
+      "TensorOpSimplifier": 0.006299495697021484,
+      "TensorOpTransform": 0.031577110290527344,
+      "TileCCOps": 0.0058362483978271484,
+      "TilingProfiler": 0.05374789237976074,
+      "TransformConvOp": 0.0022611618041992188,
+      "TritiumFusion": 0.11932778358459473,
+      "ValueNumbering": 0.0037584304809570313,
+      "VectorizeDMA": 0.0016014575958251953,
+      "VectorizeMatMult": 0.007041454315185547,
+      "WeightCoalescing": 0.0025255680084228516,
+      "ZeroSizeTensorElimination": 0.0001938343048095703
+    },
+    "tensorizer": {
+      "DMATilingProfiler::TotalInstructionsAfterTiling": 8708.0,
+      "StaticProfiler::AifUb": 437.5911865234375,
+      "StaticProfiler::ArithmeticIntensityTensorizer": 488.4932556152344,
+      "StaticProfiler::AverageDmaLength": 845.4234619140625,
+      "StaticProfiler::AverageFractalPeUtilization": 100.0,
+      "StaticProfiler::AveragePartitionUtilization": 99.8521957397461,
+      "StaticProfiler::AveragePeUtilization": 100.0,
+      "StaticProfiler::DDRTransferBytes": 240992256.0,
+      "StaticProfiler::InternalTransferBytes": 43515904.0,
+      "StaticProfiler::LoadExpanded": 275585.0,
+      "StaticProfiler::LocalizationEfficiency": 111.63233184814453,
+      "StaticProfiler::LocalizationEfficiencyIgnoreNonlocal": 119.98538208007813,
+      "StaticProfiler::StoreExpanded": 5121.0,
+      "StaticProfiler::TotalDMAExpanded": 280706.0,
+      "StaticProfiler::TotalDynamicInstancesCount": 10829.0,
+      "StaticProfiler::TotalDynamicInstancesWithMmPackedCount": 10829.0,
+      "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": 6784.0,
+      "TilingProfiler::NumPfTransposes": 7.0,
+      "TilingProfiler::NumPfTransposesForIo": 3.0,
+      "TilingProfiler::NumPfTransposesForLocal": 2.0,
+      "TilingProfiler::NumPfTransposesForNonlocal": 2.0,
+      "TilingProfiler::PfTransposeInstructions": 664.0,
+      "TilingProfiler::PfTransposeInstructionsForIo": 136.0,
+      "TilingProfiler::PfTransposeInstructionsForLocal": 272.0,
+      "TilingProfiler::PfTransposeInstructionsForNonlocal": 256.0,
+      "TilingProfiler::ReduceInstructionsAfterTiling": 0.0,
+      "TilingProfiler::SimdInstructionsAfterTiling": 279.0,
+      "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": 0.018404245376586914,
+      "AffinePredicateResolution": 0.0012030601501464844,
+      "AliasDependencyElimination": 0.00032448768615722656,
+      "AliasDependencyInduction": 0.006628751754760742,
+      "AliasDependencyReset": 0.04145312309265137,
+      "BFComputeCutting": 0.001703023910522461,
+      "BirCodeGenLoop": 0.21541404724121094,
+      "CCOpFusion": 0.017800331115722656,
+      "CanonicalizeDAGForPGTiling": 0.004300355911254883,
+      "CanonicalizeIR": 0.0014739036560058594,
+      "CoalesceCCOp": 0.006423473358154297,
+      "CommuteConcat": 0.0008399486541748047,
+      "DMALocalityOpt": 0.0018458366394042969,
+      "DMAProfiler": 0.006464719772338867,
+      "DMATilingProfiler": 0.04760599136352539,
+      "DataLocalityOpt": 0.12700510025024414,
+      "DataStreaming": 0.0063991546630859375,
+      "DeConcat": 0.00067138671875,
+      "DeadCodeElimination": 0.0010039806365966797,
+      "DeadStoreElimination": 0.006083250045776367,
+      "DelinearIndices": 0.004990339279174805,
+      "Delinearization": 0.0024585723876953125,
+      "DoNothing": 0.00012063980102539063,
+      "DramToDramTranspose": 0.01742863655090332,
+      "DumpGraphAndMetadata": 0.022494077682495117,
+      "EliminateDivs": 0.0031578540802001953,
+      "ExpandBatchNorm": 0.0016427040100097656,
+      "ExpandISAMacro": 0.004450082778930664,
+      "FactorizeBlkDims": 0.0161283016204834,
+      "FactorizeThreadAxesInFreeDims": 0.0011897087097167969,
+      "FlattenMacroLoop": 0.0030732154846191406,
+      "GenericAccessSimplifier": 0.0007944107055664063,
+      "InferInitValue": 0.08164691925048828,
+      "InferIntrinsicOnCC": 0.00891733169555664,
+      "InferNeuronTensor": 0.021415233612060547,
+      "InferNonlocalTensors": 0.020433664321899414,
+      "InferPSumTensor": 0.06388068199157715,
+      "InlineNativeKernels": 0.002740621566772461,
+      "InsertIOTransposes": 0.01563739776611328,
+      "InsertLocalTransposes": 0.004204511642456055,
+      "InsertOffloadedTransposes": 0.0030889511108398438,
+      "LICM": 0.0025815963745117188,
+      "LateLegalizeInst": 0.006772041320800781,
+      "LateLegalizePostSplit": 0.004240512847900391,
+      "LateLowerReshapeOp": 0.0011327266693115234,
+      "LateLowerTensorOp": 0.0026912689208984375,
+      "LateNeuronInstComb": 0.05948805809020996,
+      "LayoutPreprocessing": 0.0245664119720459,
+      "LayoutPreprocessingAndAnalysis": 0.07019805908203125,
+      "LayoutRequirementAnalysis": 0.004818916320800781,
+      "LegalizeCCOpLayout": 0.0020072460174560547,
+      "LegalizeOpLevelAlias": 0.0011057853698730469,
+      "LegalizePartitionReduce": 0.0015609264373779297,
+      "LegalizeSundaAccess": 0.06627082824707031,
+      "LegalizeSundaMacro": 0.018359661102294922,
+      "LegalizeType": 0.005795955657958984,
+      "LocalLayoutOpt": 0.014153003692626953,
+      "LoopFusion": 0.005875825881958008,
+      "LoopSplitting": 0.0004029273986816406,
+      "LowerBroadcast": 0.0015807151794433594,
+      "LowerCCOpBlockAxis": 0.003938198089599609,
+      "LowerComplexBroadcast": 0.004299163818359375,
+      "LowerIntrinsics": 0.26007866859436035,
+      "LowerTensorOp": 0.010293006896972656,
+      "LowerTranspose": 0.00923776626586914,
+      "MacroGeneration": 0.03194475173950195,
+      "MaskPropagation": 0.0027740001678466797,
+      "MemcpyElimination": 0.02656841278076172,
+      "MutateDataType": 0.0012118816375732422,
+      "NeuronAliasDependencyInduction": 0.0003993511199951172,
+      "NeuronAliasDependencyReset": 0.017134428024291992,
+      "NeuronInstComb": 0.004921913146972656,
+      "NeuronLICM": 0.01333165168762207,
+      "NeuronLoopFusion": 0.011570215225219727,
+      "NeuronLoopInterchange": 0.002625703811645508,
+      "NeuronSimplifier": 0.012082099914550781,
+      "NeuronSimplifyPredicates": 0.0033812522888183594,
+      "NeuronValueNumbering": 0.0051653385162353516,
+      "OptimizeAliasedCopyChain": 0.0005443096160888672,
+      "OptimizeNKIKernels": 0.6168310642242432,
+      "PAGLayoutOpt": 0.10135579109191895,
+      "PComputeCutting": 0.004828453063964844,
+      "PGLayoutTilingPipeline": 0.6252057552337646,
+      "PGTiling": 0.1626896858215332,
+      "PadElimination": 0.00044655799865722656,
+      "ParAxesAnnotation": 0.07457947731018066,
+      "PartialLoopFusion": 0.047167062759399414,
+      "PartialSimdFusion": 0.006536245346069336,
+      "PerfectLoopNest": 0.003509998321533203,
+      "RecognizeOpIdiom": 0.0035545825958251953,
+      "Recompute": 0.00031566619873046875,
+      "RelaxPredicates": 0.004423379898071289,
+      "Rematerialization": 0.001989126205444336,
+      "ReshapeWeights": 0.0011234283447265625,
+      "ResolveAccessConflict": 0.003659963607788086,
+      "ResolveComplicatePredicates": 0.0010039806365966797,
+      "RewriteReplicationMatmul": 0.0014824867248535156,
+      "RewriteWeights": 0.0052568912506103516,
+      "SFKVectorizer": 0.18293046951293945,
+      "SimpleAllReduceTiling": 0.0030977725982666016,
+      "Simplifier": 0.003216266632080078,
+      "SimplifyMacroPredicates": 0.005851030349731445,
+      "SimplifyNeuronTensor": 0.22396159172058105,
+      "SimplifySlice": 0.0011301040649414063,
+      "SimplifyTensor": 0.009292364120483398,
+      "SpillPSum": 0.01147150993347168,
+      "SplitAPUnionSets": 0.03329944610595703,
+      "SplitAccGrp": 0.0012478828430175781,
+      "StaticProfiler": 0.005914926528930664,
+      "StaticTransposeLocalTensor": 0.0035593509674072266,
+      "SundaISel": 0.09800505638122559,
+      "TCTransform": 0.0008752346038818359,
+      "TensorInitialization": 0.004550933837890625,
+      "TensorOpSimplifier": 0.004429817199707031,
+      "TensorOpTransform": 0.035314083099365234,
+      "TileCCOps": 0.006109952926635742,
+      "TilingProfiler": 0.0074079036712646484,
+      "TransformConvOp": 0.0027616024017333984,
+      "TritiumFusion": 0.07009291648864746,
+      "ValueNumbering": 0.0019168853759765625,
+      "VectorizeDMA": 0.002183198928833008,
+      "VectorizeMatMult": 0.0026412010192871094,
+      "WeightCoalescing": 0.0030510425567626953,
+      "ZeroSizeTensorElimination": 0.00018453598022460938
+    },
+    "tensorizer": {
+      "DMATilingProfiler::TotalInstructionsAfterTiling": 11124.0,
+      "StaticProfiler::AifUb": 291.8241882324219,
+      "StaticProfiler::ArithmeticIntensityTensorizer": 280.5914306640625,
+      "StaticProfiler::AverageDmaLength": 1148.8248291015625,
+      "StaticProfiler::AverageFractalPeUtilization": 99.68099975585938,
+      "StaticProfiler::AveragePartitionUtilization": 99.27783203125,
+      "StaticProfiler::AveragePeUtilization": 98.72872924804688,
+      "StaticProfiler::DDRTransferBytes": 331583520.0,
+      "StaticProfiler::InternalTransferBytes": 24224788.0,
+      "StaticProfiler::LoadExpanded": 293906.0,
+      "StaticProfiler::LocalizationEfficiency": 96.15084838867188,
+      "StaticProfiler::LocalizationEfficiencyIgnoreNonlocal": 102.69681549072266,
+      "StaticProfiler::StoreExpanded": 3627.0,
+      "StaticProfiler::TotalDMAExpanded": 297533.0,
+      "StaticProfiler::TotalDynamicInstancesCount": 13170.0,
+      "StaticProfiler::TotalDynamicInstancesWithMmPackedCount": 13074.0,
+      "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": 9472.0,
+      "TilingProfiler::NumPfTransposes": 4.0,
+      "TilingProfiler::NumPfTransposesForIo": 0.0,
+      "TilingProfiler::NumPfTransposesForLocal": 1.0,
+      "TilingProfiler::NumPfTransposesForNonlocal": 3.0,
+      "TilingProfiler::PfTransposeInstructions": 385.0,
+      "TilingProfiler::PfTransposeInstructionsForIo": 0.0,
+      "TilingProfiler::PfTransposeInstructionsForLocal": 1.0,
+      "TilingProfiler::PfTransposeInstructionsForNonlocal": 384.0,
+      "TilingProfiler::ReduceInstructionsAfterTiling": 4.0,
+      "TilingProfiler::SimdInstructionsAfterTiling": 166.0,
+      "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": 7.000000096013537e-06,
+      "CanonicalizeForTensorizer": 1.4000000192027073e-05,
+      "Canonicalizer": 0.0002880000101868063,
+      "HoistCompute": 1.1000000085914508e-05,
+      "IdentifyCrossPassTensors": 1.1000000085914508e-05,
+      "MemcastMotion": 7.000000096013537e-06,
+      "PenguinizeFunctions": 1.2999999853491317e-05,
+      "PruneFunctions": 4.999999873689376e-06,
+      "RemoveOptimizationBarriers": 1.2000000424450263e-05,
+      "ScatterMotion": 6.000000212225132e-06,
+      "TensorizerLegalizationPass": 1.4999999621068127e-05,
+      "VerifySupportedOps": 1.1000000085914508e-05,
+      "algsimp": 6.600000051548705e-05,
+      "batchnorm_expander": 1.4000000192027073e-05,
+      "boundary-marker-removal": 3.999999989900971e-06,
+      "call-inliner": 9.000000318337698e-06,
+      "canonicalize-boundary-marker": 4.999999873689376e-06,
+      "collective-stream-id-checker": 6.000000212225132e-06,
+      "comparison-expander": 4.999999873689376e-06,
+      "computation-deduplicator": 2.499999936844688e-05,
+      "conditional-to-select": 4.999999873689376e-06,
+      "config-lowering": 4.8999998398358e-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": 2.2000000171829015e-05,
+      "flatten-call-graph": 7.999999979801942e-06,
+      "fuse-send-recv": 2.5999999706982635e-05,
+      "hilo::LegalizeAlias": 4.999999873689376e-06,
+      "hilo::NeuronInstCombine": 5.0999999075429514e-05,
+      "hilo::NeuronOpFusion": 4.999999873689376e-06,
+      "hilo::ReplaceTokenTypeWithU8Pass": 1.5999999959603883e-05,
+      "hilo::ScheduleFusion": 6.000000212225132e-06,
+      "hilo::SixtyFourHack": 9.999999747378752e-06,
+      "hilo::VerifyAliasing": 1.9999999949504854e-06,
+      "hlo-mac-count": 3.5000000934815034e-05,
+      "hlo-verifier": 0.00019299999985378236,
+      "legalize-ccops": 9.999999974752427e-07,
+      "legalize-compare": 4.999999873689376e-06,
+      "lower-argminmax-custom-call": 3.000000106112566e-06,
+      "map-inline": 1.2000000424450263e-05,
+      "metadata-naming": 1.8000000636675395e-05,
+      "mlir::detail::OpToOpPassAdaptor": 2.300000051036477e-05,
+      "mlir::hlo::MhloToPyPenguin": 0.02390900067985058,
+      "mlir::mhlo::LowerComplexExtraPass": 0.0001140000022132881,
+      "mlir::mhlo::LowerComplexPass": 0.0002209999947808683,
+      "native-to-custom-softmax": 9.999999747378752e-06,
+      "native-to-custom-softmax-dx": 2.9999999242136255e-05,
+      "operand_upcaster": 3.300000025774352e-05,
+      "post-par-pipe-begin": 1.9999999949504854e-06,
+      "post-par-pipe-end": 0.0,
+      "post-partition-simplification": 0.0005729999975301325,
+      "replace-minimum-constant": 7.000000096013537e-06,
+      "reshape-mover": 3.999999989900971e-06,
+      "simplify-concat": 5.400000009103678e-05,
+      "simplify-while-loops": 1.9999999949504854e-06,
+      "transform-variadic-reduce": 9.000000318337698e-06,
+      "tuple-simplifier": 4.999999873689376e-06,
+      "unpack-nested-aws-ntwsr": 3.000000106112566e-06,
+      "unroll-while-loop": 0.0
+    },
+    "hilo": {
+      "ArithmeticIntensity": 436.7417907714844,
+      "HloMacCount": 56908316672.0,
+      "Traffic": 260603936.0
+    }
+  },
+  "sg02": {
+    "compiletime": {
+      "CanonicalizeConv": 9.999999974752427e-07,
+      "CanonicalizeForTensorizer": 2.099999983329326e-05,
+      "Canonicalizer": 0.0003530000103637576,
+      "HoistCompute": 0.0,
+      "IdentifyCrossPassTensors": 1.5999999959603883e-05,
+      "MemcastMotion": 0.0,
+      "PenguinizeFunctions": 1.4999999621068127e-05,
+      "PruneFunctions": 9.000000318337698e-06,
+      "RemoveOptimizationBarriers": 1.8000000636675395e-05,
+      "ScatterMotion": 9.999999974752427e-07,
+      "TensorizerLegalizationPass": 9.999999747378752e-06,
+      "VerifySupportedOps": 1.8000000636675395e-05,
+      "algsimp": 6.70000008540228e-05,
+      "batchnorm_expander": 1.2999999853491317e-05,
+      "boundary-marker-removal": 3.999999989900971e-06,
+      "call-inliner": 1.1000000085914508e-05,
+      "canonicalize-boundary-marker": 4.999999873689376e-06,
+      "collective-stream-id-checker": 4.999999873689376e-06,
+      "comparison-expander": 6.000000212225132e-06,
+      "computation-deduplicator": 2.499999936844688e-05,
+      "conditional-to-select": 6.000000212225132e-06,
+      "config-lowering": 5.999999848427251e-05,
+      "constant_folding": 9.000000318337698e-06,
+      "cse": 1.4000000192027073e-05,
+      "dce": 9.999999974752427e-07,
+      "dynamic-slice-transpose": 4.999999873689376e-06,
+      "eliminate-redundant-compare": 3.000000106112566e-06,
+      "emit-offloaded-dropout": 2.300000051036477e-05,
+      "flatten-call-graph": 1.1000000085914508e-05,
+      "fuse-send-recv": 2.099999983329326e-05,
+      "hilo::LegalizeAlias": 3.999999989900971e-06,
+      "hilo::NeuronInstCombine": 1.2000000424450263e-05,
+      "hilo::NeuronOpFusion": 3.999999989900971e-06,
+      "hilo::ReplaceTokenTypeWithU8Pass": 1.2000000424450263e-05,
+      "hilo::ScheduleFusion": 1.700000029813964e-05,
+      "hilo::SixtyFourHack": 6.0999998822808266e-05,
+      "hilo::VerifyAliasing": 1.9999999949504854e-06,
+      "hlo-mac-count": 0.00020599999697878957,
+      "hlo-verifier": 0.00019500000053085387,
+      "legalize-ccops": 9.999999974752427e-07,
+      "legalize-compare": 3.000000106112566e-06,
+      "lower-argminmax-custom-call": 3.000000106112566e-06,
+      "map-inline": 1.2999999853491317e-05,
+      "metadata-naming": 1.4999999621068127e-05,
+      "mlir::detail::OpToOpPassAdaptor": 3.199999991920777e-05,
+      "mlir::hlo::MhloToPyPenguin": 0.017378000542521477,
+      "mlir::mhlo::LowerComplexExtraPass": 0.00010199999815085903,
+      "mlir::mhlo::LowerComplexPass": 4.099999932805076e-05,
+      "native-to-custom-softmax": 9.999999747378752e-06,
+      "native-to-custom-softmax-dx": 3.7000001611886546e-05,
+      "operand_upcaster": 1.8000000636675395e-05,
+      "post-par-pipe-begin": 1.9999999949504854e-06,
+      "post-par-pipe-end": 0.0,
+      "post-partition-simplification": 0.0005590000073425472,
+      "replace-minimum-constant": 9.999999747378752e-06,
+      "reshape-mover": 3.000000106112566e-06,
+      "simplify-concat": 4.999999873689376e-05,
+      "simplify-while-loops": 1.9999999949504854e-06,
+      "transform-variadic-reduce": 5.900000178371556e-05,
+      "tuple-simplifier": 4.999999873689376e-06,
+      "unpack-nested-aws-ntwsr": 4.999999873689376e-06,
+      "unroll-while-loop": 0.0
+    },
+    "hilo": {
+      "ArithmeticIntensity": 279.6634216308594,
+      "HloMacCount": 45164265472.0,
+      "Traffic": 322990144.0
+    }
+  }
+}

context_encoding_model/_tp0_bk2/neuron_config.json

@@ -0,0 +1,213 @@
+{
+  "_attn_implementation_autoset": false,
+  "_name_or_path": "mistralai/Mistral-7B-Instruct-v0.3",
+  "add_cross_attention": false,
+  "architectures": [
+    "MistralForCausalLM"
+  ],
+  "attention_dropout": 0.0,
+  "attribute_map": {},
+  "bad_words_ids": null,
+  "begin_suppress_tokens": null,
+  "bos_token_id": 1,
+  "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": 2,
+  "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": 14336,
+  "is_decoder": false,
+  "is_encoder_decoder": false,
+  "label2id": {
+    "LABEL_0": 0,
+    "LABEL_1": 1
+  },
+  "length_penalty": 1.0,
+  "max_length": 20,
+  "max_position_embeddings": 32768,
+  "metadata": null,
+  "min_length": 0,
+  "model_type": "mistral",
+  "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": "NeuronLlamaAttention",
+    "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": 4,
+    "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": 4,
+    "max_context_length": 2048,
+    "max_length": 2048,
+    "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": 2048,
+    "n_positions": 2048,
+    "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": 2048,
+    "pa_num_blocks": 4,
+    "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": 2048,
+    "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": 4,
+    "speculation_length": 0,
+    "start_rank_id": 0,
+    "target": null,
+    "tile_cc": false,
+    "tkg_batch_size": 4,
+    "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": 32,
+  "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-05,
+  "rope_theta": 1000000.0,
+  "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": "4.42.0.dev0",
+  "typical_p": 1.0,
+  "use_bfloat16": false,
+  "use_cache": true,
+  "vocab_size": 32768
+}

context_encoding_model/_tp0_bk3/command.txt

@@ -0,0 +1 @@
+neuronx-cc compile --framework=XLA model.MODULE_2e1f11fbf72d40b46e64+5ae2bfda.hlo_module.pb --output model.MODULE_2e1f11fbf72d40b46e64+5ae2bfda.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

generation_config.json

@@ -0,0 +1,6 @@
+{
+  "_from_model_config": true,
+  "bos_token_id": 1,
+  "eos_token_id": 2,
+  "transformers_version": "4.42.0.dev0"
+}

neuron_config.json

@@ -0,0 +1,211 @@
+{
+  "_attn_implementation_autoset": false,
+  "_name_or_path": "mistralai/Mistral-7B-Instruct-v0.3",
+  "add_cross_attention": false,
+  "architectures": [
+    "MistralForCausalLM"
+  ],
+  "attention_dropout": 0.0,
+  "attribute_map": {},
+  "bad_words_ids": null,
+  "begin_suppress_tokens": null,
+  "bos_token_id": 1,
+  "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": 2,
+  "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": 14336,
+  "is_decoder": false,
+  "is_encoder_decoder": false,
+  "label2id": {
+    "LABEL_0": 0,
+    "LABEL_1": 1
+  },
+  "length_penalty": 1.0,
+  "max_length": 20,
+  "max_position_embeddings": 32768,
+  "metadata": null,
+  "min_length": 0,
+  "model_type": "mistral",
+  "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": "NeuronLlamaAttention",
+    "attn_kernel_enabled": null,
+    "attn_tkg_builtin_kernel_enabled": false,
+    "attn_tkg_nki_kernel_enabled": false,
+    "batch_size": 4,
+    "bucket_n_active_tokens": false,
+    "buckets": [
+      2048
+    ],
+    "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": 4,
+    "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": 4,
+    "max_context_length": 2048,
+    "max_length": 2048,
+    "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": 2048,
+    "n_positions": 2048,
+    "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": 2048,
+    "pa_num_blocks": 4,
+    "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": 2048,
+    "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": 4,
+    "speculation_length": 0,
+    "start_rank_id": 0,
+    "target": null,
+    "tile_cc": false,
+    "tkg_batch_size": 4,
+    "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": 32,
+  "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-05,
+  "rope_theta": 1000000.0,
+  "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": "4.42.0.dev0",
+  "typical_p": 1.0,
+  "use_bfloat16": false,
+  "use_cache": true,
+  "vocab_size": 32768
+}

params.json

@@ -0,0 +1,11 @@
+{
+    "dim": 4096,
+    "n_layers": 32,
+    "head_dim": 128,
+    "hidden_dim": 14336,
+    "n_heads": 32,
+    "n_kv_heads": 8,
+    "norm_eps": 1e-05,
+    "vocab_size": 32768,
+    "rope_theta": 1000000.0
+}

special_tokens_map.json

@@ -0,0 +1,23 @@
+{
+  "bos_token": {
+    "content": "<s>",
+    "lstrip": false,
+    "normalized": false,
+    "rstrip": false,
+    "single_word": false
+  },
+  "eos_token": {
+    "content": "</s>",
+    "lstrip": false,
+    "normalized": false,
+    "rstrip": false,
+    "single_word": false
+  },
+  "unk_token": {
+    "content": "<unk>",
+    "lstrip": false,
+    "normalized": false,
+    "rstrip": false,
+    "single_word": false
+  }
+}

token_generation_model/_tp0_bk0/command.txt

@@ -0,0 +1 @@
+neuronx-cc compile --framework=XLA model.MODULE_67d3774d5bacfe6ba851+72d461cc.hlo_module.pb --output model.MODULE_67d3774d5bacfe6ba851+72d461cc.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_67d3774d5bacfe6ba851+72d461cc.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=/models/mistral-7b-v0.3-instruct-neuronx/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": 99.3995361328125,
+      "StaticProfiler::AveragePartitionUtilization": 98.55674743652344,
+      "StaticProfiler::AveragePeUtilization": 97.5706558227539,
+      "StaticProfiler::LocalizationEfficiency": 116.26432037353516,
+      "StaticProfiler::LocalizationEfficiencyIgnoreNonlocal": 116.41236877441406,
+      "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": 1.7505581378936768,
+      "AffinePredicateResolution": 0.046631813049316406,
+      "AliasDependencyElimination": 0.002095460891723633,
+      "AliasDependencyInduction": 0.340775728225708,
+      "AliasDependencyReset": 0.35799717903137207,
+      "BFComputeCutting": 0.0918271541595459,
+      "BirCodeGenLoop": 1.9524388313293457,
+      "CCOpFusion": 0.6108810901641846,
+      "CanonicalizeConv": 3.999999989900971e-06,
+      "CanonicalizeDAGForPGTiling": 0.16879653930664063,
+      "CanonicalizeForTensorizer": 0.00046999999904073775,
+      "CanonicalizeIR": 0.061083316802978516,
+      "Canonicalizer": 0.006062000058591366,
+      "CoalesceCCOp": 0.1506364345550537,
+      "CommuteConcat": 0.030129432678222656,
+      "DMALocalityOpt": 0.028980255126953125,
+      "DMAProfiler": 0.07041788101196289,
+      "DMATilingProfiler": 0.06604599952697754,
+      "DataLocalityOpt": 2.1674563884735107,
+      "DataStreaming": 0.10460638999938965,
+      "DeConcat": 0.016613245010375977,
+      "DeadCodeElimination": 0.03136706352233887,
+      "DeadStoreElimination": 1.2213473320007324,
+      "DelinearIndices": 0.28975749015808105,
+      "Delinearization": 0.18038702011108398,
+      "DoNothing": 0.00029587745666503906,
+      "DramToDramTranspose": 1.0997114181518555,
+      "DumpGraphAndMetadata": 0.15712499618530273,
+      "EliminateDivs": 0.14302921295166016,
+      "ExpandBatchNorm": 0.05438518524169922,
+      "ExpandISAMacro": 0.06612992286682129,
+      "FactorizeBlkDims": 0.26380443572998047,
+      "FactorizeThreadAxesInFreeDims": 0.03626608848571777,
+      "FlattenMacroLoop": 0.0791018009185791,
+      "GenericAccessSimplifier": 0.028739452362060547,
+      "HoistCompute": 4.999999873689376e-05,
+      "IdentifyCrossPassTensors": 0.00021100000594742596,
+      "InferInitValue": 1.245818853378296,
+      "InferIntrinsicOnCC": 0.3838510513305664,
+      "InferNeuronTensor": 1.782684326171875,
+      "InferNonlocalTensors": 4.216822624206543,
+      "InferPSumTensor": 0.8002936840057373,
+      "InlineNativeKernels": 0.04331803321838379,
+      "InsertIOTransposes": 0.9620308876037598,
+      "InsertLocalTransposes": 0.9075572490692139,
+      "InsertOffloadedTransposes": 0.06905245780944824,
+      "LICM": 0.09245109558105469,
+      "LateLegalizeInst": 0.1371767520904541,
+      "LateLegalizePostSplit": 0.07115364074707031,
+      "LateLowerReshapeOp": 0.037413835525512695,
+      "LateLowerTensorOp": 0.2697789669036865,
+      "LateNeuronInstComb": 0.36163973808288574,
+      "LayoutPreprocessing": 0.8670649528503418,
+      "LayoutPreprocessingAndAnalysis": 1.1549065113067627,
+      "LayoutRequirementAnalysis": 0.27439332008361816,
+      "LegalizeCCOpLayout": 0.0667877197265625,
+      "LegalizeOpLevelAlias": 0.02463841438293457,
+      "LegalizePartitionReduce": 0.031178951263427734,
+      "LegalizeSundaAccess": 1.0370941162109375,
+      "LegalizeSundaMacro": 0.30040812492370605,
+      "LegalizeType": 0.15511703491210938,
+      "LocalLayoutOpt": 0.5859987735748291,
+      "LoopFusion": 0.31392574310302734,
+      "LoopSplitting": 0.03730463981628418,
+      "LowerBroadcast": 0.043883323669433594,
+      "LowerCCOpBlockAxis": 0.21113038063049316,
+      "LowerComplexBroadcast": 0.13543224334716797,
+      "LowerIntrinsics": 0.5533030033111572,
+      "LowerTensorOp": 0.4427661895751953,
+      "LowerTranspose": 0.445873498916626,
+      "MacroGeneration": 2.473928213119507,
+      "MaskPropagation": 0.11900115013122559,
+      "MemcastMotion": 0.00013499999477062374,
+      "MemcpyElimination": 3.895864963531494,
+      "MutateDataType": 0.0398097038269043,
+      "NeuronAliasDependencyInduction": 0.018038034439086914,
+      "NeuronAliasDependencyReset": 0.03026747703552246,
+      "NeuronInstComb": 0.2801692485809326,
+      "NeuronLICM": 0.2402355670928955,
+      "NeuronLoopFusion": 0.578554630279541,
+      "NeuronLoopInterchange": 0.04063010215759277,
+      "NeuronSimplifier": 0.338550329208374,
+      "NeuronSimplifyPredicates": 0.05692338943481445,
+      "NeuronValueNumbering": 0.08049511909484863,
+      "OptimizeAliasedCopyChain": 0.010602712631225586,
+      "OptimizeNKIKernels": 0.35573911666870117,
+      "PAGLayoutOpt": 41.70356750488281,
+      "PComputeCutting": 0.31862449645996094,
+      "PGLayoutTilingPipeline": 54.89517593383789,
+      "PGTiling": 4.955427169799805,
+      "PadElimination": 0.008247613906860352,
+      "ParAxesAnnotation": 40.78300476074219,
+      "PartialLoopFusion": 0.27280497550964355,
+      "PartialSimdFusion": 0.23333501815795898,
+      "PenguinizeFunctions": 0.00019500000053085387,
+      "PerfectLoopNest": 0.054427146911621094,
+      "PruneFunctions": 0.0002859999949578196,
+      "RecognizeOpIdiom": 0.14151334762573242,
+      "Recompute": 0.006689548492431641,
+      "RelaxPredicates": 0.11504864692687988,
+      "Rematerialization": 0.1454627513885498,
+      "RemoveOptimizationBarriers": 0.0001289999927394092,
+      "ReshapeWeights": 0.02218008041381836,
+      "ResolveAccessConflict": 0.2085890769958496,
+      "ResolveComplicatePredicates": 0.04809713363647461,
+      "RewriteReplicationMatmul": 0.04198598861694336,
+      "RewriteWeights": 0.06269216537475586,
+      "SFKVectorizer": 4.823036193847656,
+      "ScatterMotion": 0.004162999801337719,
+      "SimpleAllReduceTiling": 0.050234317779541016,
+      "Simplifier": 0.09822559356689453,
+      "SimplifyMacroPredicates": 0.17580604553222656,
+      "SimplifyNeuronTensor": 0.3997023105621338,
+      "SimplifySlice": 0.030458927154541016,
+      "SimplifyTensor": 0.19034838676452637,
+      "SpillPSum": 0.750556230545044,
+      "SplitAPUnionSets": 0.21661949157714844,
+      "SplitAccGrp": 0.034674882888793945,
+      "StaticProfiler": 0.1068570613861084,
+      "StaticTransposeLocalTensor": 0.23836159706115723,
+      "SundaISel": 1.3462426662445068,
+      "TCTransform": 0.03340411186218262,
+      "TensorInitialization": 0.08121585845947266,
+      "TensorOpSimplifier": 0.1906282901763916,
+      "TensorOpTransform": 0.768507719039917,
+      "TensorizerLegalizationPass": 0.00017600000137463212,
+      "TileCCOps": 0.1881403923034668,
+      "TilingProfiler": 0.4166131019592285,
+      "TransformConvOp": 0.04472208023071289,
+      "TritiumFusion": 1.3044919967651367,
+      "ValueNumbering": 0.0843496322631836,
+      "VectorizeDMA": 0.1174325942993164,
+      "VectorizeMatMult": 0.018784761428833008,
+      "VerifySupportedOps": 0.00020500000391621143,
+      "WeightCoalescing": 0.041903018951416016,
+      "ZeroSizeTensorElimination": 0.0003848075866699219,
+      "algsimp": 0.0015930000226944685,
+      "batchnorm_expander": 0.0005460000247694552,
+      "boundary-marker-removal": 0.0002610000083222985,
+      "call-inliner": 0.0002479999966453761,
+      "canonicalize-boundary-marker": 0.0002809999859891832,
+      "collective-stream-id-checker": 4.8000001697801054e-05,
+      "comparison-expander": 0.0002800000074785203,
+      "computation-deduplicator": 0.0003530000103637576,
+      "conditional-to-select": 8.399999933317304e-05,
+      "config-lowering": 0.00025499999173916876,
+      "constant_folding": 0.00017600000137463212,
+      "cse": 0.00037900000461377203,
+      "dce": 4.5000000682193786e-05,
+      "dynamic-slice-transpose": 0.0001610000035725534,
+      "eliminate-redundant-compare": 0.0001610000035725534,
+      "emit-offloaded-dropout": 0.0002629999944474548,
+      "flatten-call-graph": 0.00025599999935366213,
+      "fuse-send-recv": 0.0010339999571442604,
+      "hilo::LegalizeAlias": 0.0022499999031424522,
+      "hilo::NeuronInstCombine": 0.0009289999725297093,
+      "hilo::NeuronOpFusion": 0.00037399999564513564,
+      "hilo::ReplaceTokenTypeWithU8Pass": 0.0003389999910723418,
+      "hilo::ScheduleFusion": 3.9999998989515007e-05,
+      "hilo::SixtyFourHack": 0.00022699999681208283,
+      "hilo::VerifyAliasing": 0.00013600000238511711,
+      "hlo-mac-count": 0.0006539999740198255,
+      "hlo-verifier": 0.004976999945938587,
+      "io-con-pipe-begin": 3.000000106112566e-06,
+      "io-con-pipe-end": 9.999999974752427e-07,
+      "io-layout-normalization": 0.0007140000234358013,
+      "legalize-ccops": 2.2000000171829015e-05,
+      "legalize-compare": 0.00023799999326001853,
+      "lower-argminmax-custom-call": 0.00014899999951012433,
+      "map-inline": 0.0005290000117383897,
+      "metadata-naming": 0.0007159999804571271,
+      "mlir::detail::OpToOpPassAdaptor": 0.00019799999427050352,
+      "mlir::hlo::MhloToPyPenguin": 0.04695200175046921,
+      "mlir::mhlo::LowerComplexExtraPass": 0.002549000084400177,
+      "mlir::mhlo::LowerComplexPass": 0.0030499999411404133,
+      "native-to-custom-softmax": 0.0002849999873433262,
+      "native-to-custom-softmax-dx": 0.00030399998649954796,
+      "operand_upcaster": 0.0005130000063218176,
+      "post-par-pipe-begin": 9.999999974752427e-07,
+      "post-par-pipe-end": 0.0,
+      "post-partition-simplification": 0.05249100178480148,
+      "pre-hlo-begin": 1.9999999949504854e-06,
+      "pre-hlo-end": 9.999999974752427e-07,
+      "replace-minimum-constant": 0.0001740000006975606,
+      "reshape-mover": 7.500000356230885e-05,
+      "simplify-concat": 0.0017800000496208668,
+      "simplify-while-loops": 5.700000110664405e-05,
+      "transform-variadic-reduce": 0.00047900000936351717,
+      "tuple-simplifier": 0.00017699999443721026,
+      "unpack-nested-aws-ntwsr": 0.00023999999393709004,
+      "unroll-while-loop": 1.1000000085914508e-05
+    },
+    "hilo": {
+      "HloMacCount": 14294450176.0,
+      "Traffic": 7787338240.0
+    },
+    "tensorizer": {
+      "DMATilingProfiler::TotalInstructionsAfterTiling": 239380,
+      "StaticProfiler::AifUb": 3.780517101287842,
+      "StaticProfiler::ArithmeticIntensityTensorizer": 4.395392417907715,
+      "StaticProfiler::AverageDmaLength": 7173.484375,
+      "StaticProfiler::DDRTransferBytes": 7159761404,
+      "StaticProfiler::InternalTransferBytes": 29110320,
+      "StaticProfiler::LoadExpanded": 931327,
+      "StaticProfiler::StoreExpanded": 2344,
+      "StaticProfiler::TotalDMAExpanded": 933671,
+      "StaticProfiler::TotalDynamicInstancesCount": 249148,
+      "StaticProfiler::TotalDynamicInstancesWithMmPackedCount": 244610,
+      "StaticProfiler::TotalLNCComm": 0,
+      "StaticProfiler::TotalLNCCommTransfer": 0,
+      "TilingProfiler::BatchnormInstructionsAfterTiling": 0,
+      "TilingProfiler::DmaInstructionsAfterTiling": 0,
+      "TilingProfiler::GenericInstructionsAfterTiling": 329,
+      "TilingProfiler::MatMultInstructionsAfterTiling": 219136,
+      "TilingProfiler::NumPfTransposes": 394,
+      "TilingProfiler::NumPfTransposesForIo": 33,
+      "TilingProfiler::NumPfTransposesForLocal": 226,
+      "TilingProfiler::NumPfTransposesForNonlocal": 135,
+      "TilingProfiler::PfTransposeInstructions": 5649,
+      "TilingProfiler::PfTransposeInstructionsForIo": 513,
+      "TilingProfiler::PfTransposeInstructionsForLocal": 900,
+      "TilingProfiler::PfTransposeInstructionsForNonlocal": 4236,
+      "TilingProfiler::ReduceInstructionsAfterTiling": 67,
+      "TilingProfiler::SimdInstructionsAfterTiling": 6366,
+      "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": 3.999999989900971e-06,
+      "CanonicalizeForTensorizer": 0.00046999999904073775,
+      "Canonicalizer": 0.006062000058591366,
+      "HoistCompute": 4.999999873689376e-05,
+      "IdentifyCrossPassTensors": 0.00021100000594742596,
+      "MemcastMotion": 0.00013499999477062374,
+      "PenguinizeFunctions": 0.00019500000053085387,
+      "PruneFunctions": 0.0002859999949578196,
+      "RemoveOptimizationBarriers": 0.0001289999927394092,
+      "ScatterMotion": 0.004162999801337719,
+      "TensorizerLegalizationPass": 0.00017600000137463212,
+      "VerifySupportedOps": 0.00020500000391621143,
+      "algsimp": 0.0015930000226944685,
+      "batchnorm_expander": 0.0005460000247694552,
+      "boundary-marker-removal": 0.0002610000083222985,
+      "call-inliner": 0.0002479999966453761,
+      "canonicalize-boundary-marker": 0.0002809999859891832,
+      "collective-stream-id-checker": 4.8000001697801054e-05,
+      "comparison-expander": 0.0002800000074785203,
+      "computation-deduplicator": 0.0003530000103637576,
+      "conditional-to-select": 8.399999933317304e-05,
+      "config-lowering": 0.00025499999173916876,
+      "constant_folding": 0.00017600000137463212,
+      "cse": 0.00037900000461377203,
+      "dce": 4.5000000682193786e-05,
+      "dynamic-slice-transpose": 0.0001610000035725534,
+      "eliminate-redundant-compare": 0.0001610000035725534,
+      "emit-offloaded-dropout": 0.0002629999944474548,
+      "flatten-call-graph": 0.00025599999935366213,
+      "fuse-send-recv": 0.0010339999571442604,
+      "hilo::LegalizeAlias": 0.0022499999031424522,
+      "hilo::NeuronInstCombine": 0.0009289999725297093,
+      "hilo::NeuronOpFusion": 0.00037399999564513564,
+      "hilo::ReplaceTokenTypeWithU8Pass": 0.0003389999910723418,
+      "hilo::ScheduleFusion": 3.9999998989515007e-05,
+      "hilo::SixtyFourHack": 0.00022699999681208283,
+      "hilo::VerifyAliasing": 0.00013600000238511711,
+      "hlo-mac-count": 0.0006539999740198255,
+      "hlo-verifier": 0.004976999945938587,
+      "io-con-pipe-begin": 3.000000106112566e-06,
+      "io-con-pipe-end": 9.999999974752427e-07,
+      "io-layout-normalization": 0.0007140000234358013,
+      "legalize-ccops": 2.2000000171829015e-05,
+      "legalize-compare": 0.00023799999326001853,
+      "lower-argminmax-custom-call": 0.00014899999951012433,
+      "map-inline": 0.0005290000117383897,
+      "metadata-naming": 0.0007159999804571271,
+      "mlir::detail::OpToOpPassAdaptor": 0.00019799999427050352,
+      "mlir::hlo::MhloToPyPenguin": 0.04695200175046921,
+      "mlir::mhlo::LowerComplexExtraPass": 0.002549000084400177,
+      "mlir::mhlo::LowerComplexPass": 0.0030499999411404133,
+      "native-to-custom-softmax": 0.0002849999873433262,
+      "native-to-custom-softmax-dx": 0.00030399998649954796,
+      "operand_upcaster": 0.0005130000063218176,
+      "post-par-pipe-begin": 9.999999974752427e-07,
+      "post-par-pipe-end": 0.0,
+      "post-partition-simplification": 0.05249100178480148,
+      "pre-hlo-begin": 1.9999999949504854e-06,
+      "pre-hlo-end": 9.999999974752427e-07,
+      "replace-minimum-constant": 0.0001740000006975606,
+      "reshape-mover": 7.500000356230885e-05,
+      "simplify-concat": 0.0017800000496208668,
+      "simplify-while-loops": 5.700000110664405e-05,
+      "transform-variadic-reduce": 0.00047900000936351717,
+      "tuple-simplifier": 0.00017699999443721026,
+      "unpack-nested-aws-ntwsr": 0.00023999999393709004,
+      "unroll-while-loop": 1.1000000085914508e-05
+    }
+  },
+  "cumsum": {
+    "compiletime": {
+      "CoalesceCCOp": 0.0002486705780029297,
+      "DMALocalityOpt": 0.00020575523376464844,
+      "DMAProfiler": 0.0008375644683837891,
+      "DataStreaming": 0.0002942085266113281,
+      "DoNothing": 0.000156402587890625,
+      "ExpandISAMacro": 0.0005459785461425781,
+      "FactorizeBlkDims": 0.00045418739318847656,
+      "InferPSumTensor": 0.00048661231994628906,
+      "LateLegalizeInst": 0.00043702125549316406,
+      "LateNeuronInstComb": 0.0004830360412597656,
+      "LegalizeSundaAccess": 0.0015192031860351563,
+      "LegalizeType": 0.0002753734588623047,
+      "LowerBroadcast": 0.00025272369384765625,
+      "LowerIntrinsics": 0.000255584716796875,
+      "LowerTranspose": 0.0002532005310058594,
+      "NeuronInstComb": 0.0005066394805908203,
+      "NeuronLICM": 0.00046443939208984375,
+      "NeuronSimplifyPredicates": 0.002785921096801758,
+      "NeuronValueNumbering": 0.00043272972106933594,
+      "SFKVectorizer": 0.002558469772338867,
+      "SimpleAllReduceTiling": 0.00026297569274902344,
+      "SimplifyNeuronTensor": 0.00043082237243652344,
+      "SpillPSum": 0.0005290508270263672,
+      "WeightCoalescing": 0.00025010108947753906
+    }
+  },
+  "sg00": {
+    "hilo": {
+      "ArithmeticIntensity": 3.671203136444092,
+      "HloMacCount": 14294450176.0,
+      "Traffic": 7787338240.0
+    }
+  },
+  "sg0000": {
+    "compiletime": {
+      "AGOrderingAnalysisPass": 1.7505581378936768,
+      "AffinePredicateResolution": 0.046631813049316406,
+      "AliasDependencyElimination": 0.002095460891723633,
+      "AliasDependencyInduction": 0.340775728225708,
+      "AliasDependencyReset": 0.35799717903137207,
+      "BFComputeCutting": 0.0918271541595459,
+      "BirCodeGenLoop": 1.9524388313293457,
+      "CCOpFusion": 0.6108810901641846,
+      "CanonicalizeDAGForPGTiling": 0.16879653930664063,
+      "CanonicalizeIR": 0.061083316802978516,
+      "CoalesceCCOp": 0.15038776397705078,
+      "CommuteConcat": 0.030129432678222656,
+      "DMALocalityOpt": 0.028774499893188477,
+      "DMAProfiler": 0.0695803165435791,
+      "DMATilingProfiler": 0.06604599952697754,
+      "DataLocalityOpt": 2.1674563884735107,
+      "DataStreaming": 0.10431218147277832,
+      "DeConcat": 0.016613245010375977,
+      "DeadCodeElimination": 0.03136706352233887,
+      "DeadStoreElimination": 1.2213473320007324,
+      "DelinearIndices": 0.28975749015808105,
+      "Delinearization": 0.18038702011108398,
+      "DoNothing": 0.00013947486877441406,
+      "DramToDramTranspose": 1.0997114181518555,
+      "DumpGraphAndMetadata": 0.15712499618530273,
+      "EliminateDivs": 0.14302921295166016,
+      "ExpandBatchNorm": 0.05438518524169922,
+      "ExpandISAMacro": 0.06558394432067871,
+      "FactorizeBlkDims": 0.263350248336792,
+      "FactorizeThreadAxesInFreeDims": 0.03626608848571777,
+      "FlattenMacroLoop": 0.0791018009185791,
+      "GenericAccessSimplifier": 0.028739452362060547,
+      "InferInitValue": 1.245818853378296,
+      "InferIntrinsicOnCC": 0.3838510513305664,
+      "InferNeuronTensor": 1.782684326171875,
+      "InferNonlocalTensors": 4.216822624206543,
+      "InferPSumTensor": 0.799807071685791,
+      "InlineNativeKernels": 0.04331803321838379,
+      "InsertIOTransposes": 0.9620308876037598,
+      "InsertLocalTransposes": 0.9075572490692139,
+      "InsertOffloadedTransposes": 0.06905245780944824,
+      "LICM": 0.09245109558105469,
+      "LateLegalizeInst": 0.13673973083496094,
+      "LateLegalizePostSplit": 0.07115364074707031,
+      "LateLowerReshapeOp": 0.037413835525512695,
+      "LateLowerTensorOp": 0.2697789669036865,
+      "LateNeuronInstComb": 0.361156702041626,
+      "LayoutPreprocessing": 0.8670649528503418,
+      "LayoutPreprocessingAndAnalysis": 1.1549065113067627,
+      "LayoutRequirementAnalysis": 0.27439332008361816,
+      "LegalizeCCOpLayout": 0.0667877197265625,
+      "LegalizeOpLevelAlias": 0.02463841438293457,
+      "LegalizePartitionReduce": 0.031178951263427734,
+      "LegalizeSundaAccess": 1.0355749130249023,
+      "LegalizeSundaMacro": 0.30040812492370605,
+      "LegalizeType": 0.15484166145324707,
+      "LocalLayoutOpt": 0.5859987735748291,
+      "LoopFusion": 0.31392574310302734,
+      "LoopSplitting": 0.03730463981628418,
+      "LowerBroadcast": 0.04363059997558594,
+      "LowerCCOpBlockAxis": 0.21113038063049316,
+      "LowerComplexBroadcast": 0.13543224334716797,
+      "LowerIntrinsics": 0.5530474185943604,
+      "LowerTensorOp": 0.4427661895751953,
+      "LowerTranspose": 0.4456202983856201,
+      "MacroGeneration": 2.473928213119507,
+      "MaskPropagation": 0.11900115013122559,
+      "MemcpyElimination": 3.895864963531494,
+      "MutateDataType": 0.0398097038269043,
+      "NeuronAliasDependencyInduction": 0.018038034439086914,
+      "NeuronAliasDependencyReset": 0.03026747703552246,
+      "NeuronInstComb": 0.2796626091003418,
+      "NeuronLICM": 0.23977112770080566,
+      "NeuronLoopFusion": 0.578554630279541,
+      "NeuronLoopInterchange": 0.04063010215759277,
+      "NeuronSimplifier": 0.338550329208374,
+      "NeuronSimplifyPredicates": 0.054137468338012695,
+      "NeuronValueNumbering": 0.0800623893737793,
+      "OptimizeAliasedCopyChain": 0.010602712631225586,
+      "OptimizeNKIKernels": 0.35573911666870117,
+      "PAGLayoutOpt": 41.70356750488281,
+      "PComputeCutting": 0.31862449645996094,
+      "PGLayoutTilingPipeline": 54.89517593383789,
+      "PGTiling": 4.955427169799805,
+      "PadElimination": 0.008247613906860352,
+      "ParAxesAnnotation": 40.78300476074219,
+      "PartialLoopFusion": 0.27280497550964355,
+      "PartialSimdFusion": 0.23333501815795898,
+      "PerfectLoopNest": 0.054427146911621094,
+      "RecognizeOpIdiom": 0.14151334762573242,
+      "Recompute": 0.006689548492431641,
+      "RelaxPredicates": 0.11504864692687988,
+      "Rematerialization": 0.1454627513885498,
+      "ReshapeWeights": 0.02218008041381836,
+      "ResolveAccessConflict": 0.2085890769958496,
+      "ResolveComplicatePredicates": 0.04809713363647461,
+      "RewriteReplicationMatmul": 0.04198598861694336,
+      "RewriteWeights": 0.06269216537475586,
+      "SFKVectorizer": 4.820477485656738,
+      "SimpleAllReduceTiling": 0.04997134208679199,
+      "Simplifier": 0.09822559356689453,
+      "SimplifyMacroPredicates": 0.17580604553222656,
+      "SimplifyNeuronTensor": 0.39927148818969727,
+      "SimplifySlice": 0.030458927154541016,
+      "SimplifyTensor": 0.19034838676452637,
+      "SpillPSum": 0.7500271797180176,
+      "SplitAPUnionSets": 0.21661949157714844,
+      "SplitAccGrp": 0.034674882888793945,
+      "StaticProfiler": 0.1068570613861084,
+      "StaticTransposeLocalTensor": 0.23836159706115723,
+      "SundaISel": 1.3462426662445068,
+      "TCTransform": 0.03340411186218262,
+      "TensorInitialization": 0.08121585845947266,
+      "TensorOpSimplifier": 0.1906282901763916,
+      "TensorOpTransform": 0.768507719039917,
+      "TileCCOps": 0.1881403923034668,
+      "TilingProfiler": 0.4166131019592285,
+      "TransformConvOp": 0.04472208023071289,
+      "TritiumFusion": 1.3044919967651367,
+      "ValueNumbering": 0.0843496322631836,
+      "VectorizeDMA": 0.1174325942993164,
+      "VectorizeMatMult": 0.018784761428833008,
+      "WeightCoalescing": 0.04165291786193848,
+      "ZeroSizeTensorElimination": 0.0003848075866699219
+    },
+    "tensorizer": {
+      "DMATilingProfiler::TotalInstructionsAfterTiling": 239380,
+      "StaticProfiler::AifUb": 3.780517101287842,
+      "StaticProfiler::ArithmeticIntensityTensorizer": 4.395392417907715,
+      "StaticProfiler::AverageDmaLength": 7173.484375,
+      "StaticProfiler::AverageFractalPeUtilization": 99.3995361328125,
+      "StaticProfiler::AveragePartitionUtilization": 98.55674743652344,
+      "StaticProfiler::AveragePeUtilization": 97.5706558227539,
+      "StaticProfiler::DDRTransferBytes": 7159761404,
+      "StaticProfiler::InternalTransferBytes": 29110320,
+      "StaticProfiler::LoadExpanded": 931327,
+      "StaticProfiler::LocalizationEfficiency": 116.26432037353516,
+      "StaticProfiler::LocalizationEfficiencyIgnoreNonlocal": 116.41236877441406,
+      "StaticProfiler::StoreExpanded": 2344,
+      "StaticProfiler::TotalDMAExpanded": 933671,
+      "StaticProfiler::TotalDynamicInstancesCount": 249148,
+      "StaticProfiler::TotalDynamicInstancesWithMmPackedCount": 244610,
+      "StaticProfiler::TotalLNCComm": 0,
+      "StaticProfiler::TotalLNCCommTransfer": 0,
+      "TilingProfiler::AveragePartitionUtilizationAfterTiling": 0,
+      "TilingProfiler::AveragePeUtilizationAfterTiling": 0,
+      "TilingProfiler::BatchnormInstructionsAfterTiling": 0,
+      "TilingProfiler::DmaInstructionsAfterTiling": 0,
+      "TilingProfiler::GenericInstructionsAfterTiling": 329,
+      "TilingProfiler::MatMultInstructionsAfterTiling": 219136,
+      "TilingProfiler::NumPfTransposes": 394,
+      "TilingProfiler::NumPfTransposesForIo": 33,
+      "TilingProfiler::NumPfTransposesForLocal": 226,
+      "TilingProfiler::NumPfTransposesForNonlocal": 135,
+      "TilingProfiler::PfTransposeInstructions": 5649,
+      "TilingProfiler::PfTransposeInstructionsForIo": 513,
+      "TilingProfiler::PfTransposeInstructionsForLocal": 900,
+      "TilingProfiler::PfTransposeInstructionsForNonlocal": 4236,
+      "TilingProfiler::ReduceInstructionsAfterTiling": 67,
+      "TilingProfiler::SimdInstructionsAfterTiling": 6366,
+      "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/neuron_config.json

@@ -0,0 +1,213 @@
+{
+  "_attn_implementation_autoset": false,
+  "_name_or_path": "mistralai/Mistral-7B-Instruct-v0.3",
+  "add_cross_attention": false,
+  "architectures": [
+    "MistralForCausalLM"
+  ],
+  "attention_dropout": 0.0,
+  "attribute_map": {},
+  "bad_words_ids": null,
+  "begin_suppress_tokens": null,
+  "bos_token_id": 1,
+  "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": 2,
+  "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": 14336,
+  "is_decoder": false,
+  "is_encoder_decoder": false,
+  "label2id": {
+    "LABEL_0": 0,
+    "LABEL_1": 1
+  },
+  "length_penalty": 1.0,
+  "max_length": 20,
+  "max_position_embeddings": 32768,
+  "metadata": null,
+  "min_length": 0,
+  "model_type": "mistral",
+  "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": "NeuronLlamaAttention",
+    "attn_kernel_enabled": null,
+    "attn_tkg_builtin_kernel_enabled": false,
+    "attn_tkg_nki_kernel_enabled": false,
+    "batch_size": 4,
+    "bucket_n_active_tokens": false,
+    "buckets": [
+      128
+    ],
+    "cast_type": "config",
+    "cc_pipeline_tiling_factor": 1,
+    "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": false,
+    "is_prefix_caching": false,
+    "k_cache_transposed": false,
+    "kv_cache_batch_size": 4,
+    "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": 4,
+    "max_context_length": 2048,
+    "max_length": 2048,
+    "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": 1,
+    "n_positions": 2048,
+    "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": 2048,
+    "pa_num_blocks": 4,
+    "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": 2048,
+    "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": 4,
+    "speculation_length": 0,
+    "start_rank_id": 0,
+    "target": null,
+    "tile_cc": false,
+    "tkg_batch_size": 4,
+    "token_generation_buckets": [
+      128
+    ],
+    "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": 32,
+  "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-05,
+  "rope_theta": 1000000.0,
+  "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": "4.42.0.dev0",
+  "typical_p": 1.0,
+  "use_bfloat16": false,
+  "use_cache": true,
+  "vocab_size": 32768
+}

token_generation_model/_tp0_bk1/command.txt

@@ -0,0 +1 @@
+neuronx-cc compile --framework=XLA model.MODULE_92bbfea7801df2fea75e+4948da29.hlo_module.pb --output model.MODULE_92bbfea7801df2fea75e+4948da29.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 --verbose=35

token_generation_model/_tp0_bk1/compile_flags.MODULE_92bbfea7801df2fea75e+4948da29.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=/models/mistral-7b-v0.3-instruct-neuronx/token_generation_model/_tp0_bk1/log-neuron-cc.txt"]

token_generation_model/_tp0_bk1/global_metric_store.json

@@ -0,0 +1,540 @@
+{
+  "Average": {
+    "tensorizer": {
+      "StaticProfiler::AverageFractalPeUtilization": 99.41757202148438,
+      "StaticProfiler::AveragePartitionUtilization": 98.54876708984375,
+      "StaticProfiler::AveragePeUtilization": 97.57201385498047,
+      "StaticProfiler::LocalizationEfficiency": 115.72197723388672,
+      "StaticProfiler::LocalizationEfficiencyIgnoreNonlocal": 115.86865234375,
+      "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": 2.5377185344696045,
+      "AffinePredicateResolution": 0.09365105628967285,
+      "AliasDependencyElimination": 0.0029387474060058594,
+      "AliasDependencyInduction": 0.9149439334869385,
+      "AliasDependencyReset": 0.9402039051055908,
+      "BFComputeCutting": 0.08754968643188477,
+      "BirCodeGenLoop": 1.745743751525879,
+      "CCOpFusion": 0.6600353717803955,
+      "CanonicalizeConv": 7.100000220816582e-05,
+      "CanonicalizeDAGForPGTiling": 0.18796682357788086,
+      "CanonicalizeForTensorizer": 0.0007590000168420374,
+      "CanonicalizeIR": 0.11113405227661133,
+      "Canonicalizer": 0.007284999825060368,
+      "CoalesceCCOp": 0.17181134223937988,
+      "CommuteConcat": 0.03446626663208008,
+      "DMALocalityOpt": 0.0312197208404541,
+      "DMAProfiler": 0.07886791229248047,
+      "DMATilingProfiler": 0.07289671897888184,
+      "DataLocalityOpt": 2.36210560798645,
+      "DataStreaming": 0.11742329597473145,
+      "DeConcat": 0.018580198287963867,
+      "DeadCodeElimination": 0.034860849380493164,
+      "DeadStoreElimination": 1.3254823684692383,
+      "DelinearIndices": 0.3446221351623535,
+      "Delinearization": 0.1506659984588623,
+      "DoNothing": 0.0002808570861816406,
+      "DramToDramTranspose": 1.202136516571045,
+      "DumpGraphAndMetadata": 0.16636061668395996,
+      "EliminateDivs": 0.2120811939239502,
+      "ExpandBatchNorm": 0.08646512031555176,
+      "ExpandISAMacro": 0.07142186164855957,
+      "FactorizeBlkDims": 0.2816791534423828,
+      "FactorizeThreadAxesInFreeDims": 0.04656529426574707,
+      "FlattenMacroLoop": 0.09647870063781738,
+      "GenericAccessSimplifier": 0.03339672088623047,
+      "HoistCompute": 0.00010299999848939478,
+      "IdentifyCrossPassTensors": 0.00017699999443721026,
+      "InferInitValue": 1.0866944789886475,
+      "InferIntrinsicOnCC": 0.36792969703674316,
+      "InferNeuronTensor": 1.8716816902160645,
+      "InferNonlocalTensors": 4.868690013885498,
+      "InferPSumTensor": 1.1845901012420654,
+      "InlineNativeKernels": 0.046529531478881836,
+      "InsertIOTransposes": 0.8822572231292725,
+      "InsertLocalTransposes": 0.9775772094726563,
+      "InsertOffloadedTransposes": 0.07813811302185059,
+      "LICM": 0.09977531433105469,
+      "LateLegalizeInst": 0.15209054946899414,
+      "LateLegalizePostSplit": 0.07948708534240723,
+      "LateLowerReshapeOp": 0.044101715087890625,
+      "LateLowerTensorOp": 0.9023852348327637,
+      "LateNeuronInstComb": 0.3997476100921631,
+      "LayoutPreprocessing": 1.0312676429748535,
+      "LayoutPreprocessingAndAnalysis": 1.378429651260376,
+      "LayoutRequirementAnalysis": 0.3332996368408203,
+      "LegalizeCCOpLayout": 0.11418366432189941,
+      "LegalizeOpLevelAlias": 0.030529260635375977,
+      "LegalizePartitionReduce": 0.03310394287109375,
+      "LegalizeSundaAccess": 1.0588877201080322,
+      "LegalizeSundaMacro": 0.32631635665893555,
+      "LegalizeType": 0.15178585052490234,
+      "LocalLayoutOpt": 0.736790657043457,
+      "LoopFusion": 0.3599417209625244,
+      "LoopSplitting": 0.03815197944641113,
+      "LowerBroadcast": 0.05054831504821777,
+      "LowerCCOpBlockAxis": 0.24752306938171387,
+      "LowerComplexBroadcast": 0.15025925636291504,
+      "LowerIntrinsics": 0.5705869197845459,
+      "LowerTensorOp": 0.6496686935424805,
+      "LowerTranspose": 0.47455263137817383,
+      "MacroGeneration": 2.619112730026245,
+      "MaskPropagation": 0.13018035888671875,
+      "MemcastMotion": 0.00015799999528098851,
+      "MemcpyElimination": 9.560661315917969,
+      "MutateDataType": 0.04511690139770508,
+      "NeuronAliasDependencyInduction": 0.01844191551208496,
+      "NeuronAliasDependencyReset": 0.032405853271484375,
+      "NeuronInstComb": 0.31444406509399414,
+      "NeuronLICM": 0.2590310573577881,
+      "NeuronLoopFusion": 0.7092890739440918,
+      "NeuronLoopInterchange": 0.0486292839050293,
+      "NeuronSimplifier": 0.3838067054748535,
+      "NeuronSimplifyPredicates": 0.06155133247375488,
+      "NeuronValueNumbering": 0.09124898910522461,
+      "OptimizeAliasedCopyChain": 0.01573491096496582,
+      "OptimizeNKIKernels": 0.37357568740844727,
+      "PAGLayoutOpt": 43.78666687011719,
+      "PComputeCutting": 0.2951951026916504,
+      "PGLayoutTilingPipeline": 58.915504455566406,
+      "PGTiling": 5.879387855529785,
+      "PadElimination": 0.015177726745605469,
+      "ParAxesAnnotation": 42.794742584228516,
+      "PartialLoopFusion": 0.3791334629058838,
+      "PartialSimdFusion": 0.28896236419677734,
+      "PenguinizeFunctions": 0.00035099999513477087,
+      "PerfectLoopNest": 0.06396722793579102,
+      "PruneFunctions": 0.0005990000208839774,
+      "RecognizeOpIdiom": 0.14191770553588867,
+      "Recompute": 0.008079290390014648,
+      "RelaxPredicates": 0.12253475189208984,
+      "Rematerialization": 0.15925836563110352,
+      "RemoveOptimizationBarriers": 0.0001049999991664663,
+      "ReshapeWeights": 0.02360057830810547,
+      "ResolveAccessConflict": 0.23374605178833008,
+      "ResolveComplicatePredicates": 0.09412980079650879,
+      "RewriteReplicationMatmul": 0.05128669738769531,
+      "RewriteWeights": 0.069183349609375,
+      "SFKVectorizer": 5.670078754425049,
+      "ScatterMotion": 0.007300000172108412,
+      "SimpleAllReduceTiling": 0.05962991714477539,
+      "Simplifier": 0.11846542358398438,
+      "SimplifyMacroPredicates": 0.20037221908569336,
+      "SimplifyNeuronTensor": 0.4303016662597656,
+      "SimplifySlice": 0.03480982780456543,
+      "SimplifyTensor": 0.21754741668701172,
+      "SpillPSum": 0.5786852836608887,
+      "SplitAPUnionSets": 0.2246379852294922,
+      "SplitAccGrp": 0.041144371032714844,
+      "StaticProfiler": 0.11391091346740723,
+      "StaticTransposeLocalTensor": 0.2581653594970703,
+      "SundaISel": 1.714808702468872,
+      "TCTransform": 0.03960013389587402,
+      "TensorInitialization": 0.08710885047912598,
+      "TensorOpSimplifier": 0.5631639957427979,
+      "TensorOpTransform": 1.7516539096832275,
+      "TensorizerLegalizationPass": 0.00033099998836405575,
+      "TileCCOps": 0.2776148319244385,
+      "TilingProfiler": 0.4766700267791748,
+      "TransformConvOp": 0.05919003486633301,
+      "TritiumFusion": 1.33201265335083,
+      "ValueNumbering": 0.09897065162658691,
+      "VectorizeDMA": 0.1225745677947998,
+      "VectorizeMatMult": 0.0306241512298584,
+      "VerifySupportedOps": 0.0002410000015515834,
+      "WeightCoalescing": 0.047278642654418945,
+      "ZeroSizeTensorElimination": 0.0004246234893798828,
+      "algsimp": 0.003000000026077032,
+      "batchnorm_expander": 0.0011419999646022916,
+      "boundary-marker-removal": 0.0003319999959785491,
+      "call-inliner": 0.0004710000066552311,
+      "canonicalize-boundary-marker": 0.00036800000816583633,
+      "collective-stream-id-checker": 8.399999933317304e-05,
+      "comparison-expander": 0.00036199999158270657,
+      "computation-deduplicator": 0.000446999998530373,
+      "conditional-to-select": 0.00016700000560376793,
+      "config-lowering": 0.0003150000120513141,
+      "constant_folding": 0.0003420000139158219,
+      "cse": 0.0009110000100918114,
+      "dce": 0.00011300000187475234,
+      "dynamic-slice-transpose": 0.00029799999902024865,
+      "eliminate-redundant-compare": 0.00032900000223889947,
+      "emit-offloaded-dropout": 0.0004870000120718032,
+      "flatten-call-graph": 0.00035600000410340726,
+      "fuse-send-recv": 0.002119000069797039,
+      "hilo::LegalizeAlias": 0.007114000152796507,
+      "hilo::NeuronInstCombine": 0.002025000052526593,
+      "hilo::NeuronOpFusion": 1.4999999621068127e-05,
+      "hilo::ReplaceTokenTypeWithU8Pass": 0.0002280000044265762,
+      "hilo::ScheduleFusion": 7.100000220816582e-05,
+      "hilo::SixtyFourHack": 0.0005000000237487257,
+      "hilo::VerifyAliasing": 0.000295000005280599,
+      "hlo-mac-count": 0.0010900000343099236,
+      "hlo-verifier": 0.008117999881505966,
+      "io-con-pipe-begin": 1.1000000085914508e-05,
+      "io-con-pipe-end": 9.999999974752427e-07,
+      "io-layout-normalization": 0.0011190000222995877,
+      "legalize-ccops": 4.70000013592653e-05,
+      "legalize-compare": 0.00031999999191612005,
+      "lower-argminmax-custom-call": 0.0002849999873433262,
+      "map-inline": 0.0006060000159777701,
+      "metadata-naming": 0.001366000040434301,
+      "mlir::detail::OpToOpPassAdaptor": 0.00028899998869746923,
+      "mlir::hlo::MhloToPyPenguin": 0.06228100135922432,
+      "mlir::mhlo::LowerComplexExtraPass": 0.0029380000196397305,
+      "mlir::mhlo::LowerComplexPass": 0.004118999931961298,
+      "native-to-custom-softmax": 0.000371000001905486,
+      "native-to-custom-softmax-dx": 0.0003800000122282654,
+      "operand_upcaster": 0.0007779999868944287,
+      "post-par-pipe-begin": 1.9999999949504854e-06,
+      "post-par-pipe-end": 0.0,
+      "post-partition-simplification": 0.09592299908399582,
+      "pre-hlo-begin": 3.999999989900971e-06,
+      "pre-hlo-end": 9.999999974752427e-07,
+      "replace-minimum-constant": 0.0003220000071451068,
+      "reshape-mover": 0.00015500000154133886,
+      "simplify-concat": 0.0029800001066178083,
+      "simplify-while-loops": 0.00015199999324977398,
+      "transform-variadic-reduce": 0.000577000027988106,
+      "tuple-simplifier": 0.00037200000951997936,
+      "unpack-nested-aws-ntwsr": 0.000307999987853691,
+      "unroll-while-loop": 2.700000004551839e-05
+    },
+    "hilo": {
+      "HloMacCount": 14361559040.0,
+      "Traffic": 7787340288.0
+    },
+    "tensorizer": {
+      "DMATilingProfiler::TotalInstructionsAfterTiling": 241479,
+      "StaticProfiler::AifUb": 4.066871166229248,
+      "StaticProfiler::ArithmeticIntensityTensorizer": 4.706264019012451,
+      "StaticProfiler::AverageDmaLength": 6434.17578125,
+      "StaticProfiler::DDRTransferBytes": 7193317884,
+      "StaticProfiler::InternalTransferBytes": 47988784,
+      "StaticProfiler::LoadExpanded": 1054211,
+      "StaticProfiler::StoreExpanded": 2218,
+      "StaticProfiler::TotalDMAExpanded": 1056429,
+      "StaticProfiler::TotalDynamicInstancesCount": 251892,
+      "StaticProfiler::TotalDynamicInstancesWithMmPackedCount": 247287,
+      "StaticProfiler::TotalLNCComm": 0,
+      "StaticProfiler::TotalLNCCommTransfer": 0,
+      "TilingProfiler::BatchnormInstructionsAfterTiling": 0,
+      "TilingProfiler::DmaInstructionsAfterTiling": 0,
+      "TilingProfiler::GenericInstructionsAfterTiling": 281,
+      "TilingProfiler::MatMultInstructionsAfterTiling": 220160,
+      "TilingProfiler::NumPfTransposes": 394,
+      "TilingProfiler::NumPfTransposesForIo": 33,
+      "TilingProfiler::NumPfTransposesForLocal": 226,
+      "TilingProfiler::NumPfTransposesForNonlocal": 135,
+      "TilingProfiler::PfTransposeInstructions": 6226,
+      "TilingProfiler::PfTransposeInstructionsForIo": 1026,
+      "TilingProfiler::PfTransposeInstructionsForLocal": 964,
+      "TilingProfiler::PfTransposeInstructionsForNonlocal": 4236,
+      "TilingProfiler::ReduceInstructionsAfterTiling": 67,
+      "TilingProfiler::SimdInstructionsAfterTiling": 6399,
+      "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": 7.100000220816582e-05,
+      "CanonicalizeForTensorizer": 0.0007590000168420374,
+      "Canonicalizer": 0.007284999825060368,
+      "HoistCompute": 0.00010299999848939478,
+      "IdentifyCrossPassTensors": 0.00017699999443721026,
+      "MemcastMotion": 0.00015799999528098851,
+      "PenguinizeFunctions": 0.00035099999513477087,
+      "PruneFunctions": 0.0005990000208839774,
+      "RemoveOptimizationBarriers": 0.0001049999991664663,
+      "ScatterMotion": 0.007300000172108412,
+      "TensorizerLegalizationPass": 0.00033099998836405575,
+      "VerifySupportedOps": 0.0002410000015515834,
+      "algsimp": 0.003000000026077032,
+      "batchnorm_expander": 0.0011419999646022916,
+      "boundary-marker-removal": 0.0003319999959785491,
+      "call-inliner": 0.0004710000066552311,
+      "canonicalize-boundary-marker": 0.00036800000816583633,
+      "collective-stream-id-checker": 8.399999933317304e-05,
+      "comparison-expander": 0.00036199999158270657,
+      "computation-deduplicator": 0.000446999998530373,
+      "conditional-to-select": 0.00016700000560376793,
+      "config-lowering": 0.0003150000120513141,
+      "constant_folding": 0.0003420000139158219,
+      "cse": 0.0009110000100918114,
+      "dce": 0.00011300000187475234,
+      "dynamic-slice-transpose": 0.00029799999902024865,
+      "eliminate-redundant-compare": 0.00032900000223889947,
+      "emit-offloaded-dropout": 0.0004870000120718032,
+      "flatten-call-graph": 0.00035600000410340726,
+      "fuse-send-recv": 0.002119000069797039,
+      "hilo::LegalizeAlias": 0.007114000152796507,
+      "hilo::NeuronInstCombine": 0.002025000052526593,
+      "hilo::NeuronOpFusion": 1.4999999621068127e-05,
+      "hilo::ReplaceTokenTypeWithU8Pass": 0.0002280000044265762,
+      "hilo::ScheduleFusion": 7.100000220816582e-05,
+      "hilo::SixtyFourHack": 0.0005000000237487257,
+      "hilo::VerifyAliasing": 0.000295000005280599,
+      "hlo-mac-count": 0.0010900000343099236,
+      "hlo-verifier": 0.008117999881505966,
+      "io-con-pipe-begin": 1.1000000085914508e-05,
+      "io-con-pipe-end": 9.999999974752427e-07,
+      "io-layout-normalization": 0.0011190000222995877,
+      "legalize-ccops": 4.70000013592653e-05,
+      "legalize-compare": 0.00031999999191612005,
+      "lower-argminmax-custom-call": 0.0002849999873433262,
+      "map-inline": 0.0006060000159777701,
+      "metadata-naming": 0.001366000040434301,
+      "mlir::detail::OpToOpPassAdaptor": 0.00028899998869746923,
+      "mlir::hlo::MhloToPyPenguin": 0.06228100135922432,
+      "mlir::mhlo::LowerComplexExtraPass": 0.0029380000196397305,
+      "mlir::mhlo::LowerComplexPass": 0.004118999931961298,
+      "native-to-custom-softmax": 0.000371000001905486,
+      "native-to-custom-softmax-dx": 0.0003800000122282654,
+      "operand_upcaster": 0.0007779999868944287,
+      "post-par-pipe-begin": 1.9999999949504854e-06,
+      "post-par-pipe-end": 0.0,
+      "post-partition-simplification": 0.09592299908399582,
+      "pre-hlo-begin": 3.999999989900971e-06,
+      "pre-hlo-end": 9.999999974752427e-07,
+      "replace-minimum-constant": 0.0003220000071451068,
+      "reshape-mover": 0.00015500000154133886,
+      "simplify-concat": 0.0029800001066178083,
+      "simplify-while-loops": 0.00015199999324977398,
+      "transform-variadic-reduce": 0.000577000027988106,
+      "tuple-simplifier": 0.00037200000951997936,
+      "unpack-nested-aws-ntwsr": 0.000307999987853691,
+      "unroll-while-loop": 2.700000004551839e-05
+    }
+  },
+  "cumsum": {
+    "compiletime": {
+      "CoalesceCCOp": 0.0002815723419189453,
+      "DMALocalityOpt": 0.00020623207092285156,
+      "DMAProfiler": 0.0009174346923828125,
+      "DataStreaming": 0.0003020763397216797,
+      "DoNothing": 0.00015735626220703125,
+      "ExpandISAMacro": 0.000522613525390625,
+      "FactorizeBlkDims": 0.00048041343688964844,
+      "InferPSumTensor": 0.0004773139953613281,
+      "LateLegalizeInst": 0.0004329681396484375,
+      "LateNeuronInstComb": 0.00048470497131347656,
+      "LegalizeSundaAccess": 0.0015790462493896484,
+      "LegalizeType": 0.0002918243408203125,
+      "LowerBroadcast": 0.0002551078796386719,
+      "LowerIntrinsics": 0.00031304359436035156,
+      "LowerTranspose": 0.0002570152282714844,
+      "NeuronInstComb": 0.0004956722259521484,
+      "NeuronLICM": 0.00045418739318847656,
+      "NeuronSimplifyPredicates": 0.0029306411743164063,
+      "NeuronValueNumbering": 0.00044155120849609375,
+      "SFKVectorizer": 0.002784252166748047,
+      "SimpleAllReduceTiling": 0.000247955322265625,
+      "SimplifyNeuronTensor": 0.0004968643188476563,
+      "SpillPSum": 0.0005667209625244141,
+      "WeightCoalescing": 0.000274658203125
+    }
+  },
+  "sg00": {
+    "hilo": {
+      "ArithmeticIntensity": 3.6884374618530273,
+      "HloMacCount": 14361559040.0,
+      "Traffic": 7787340288.0
+    }
+  },
+  "sg0000": {
+    "compiletime": {
+      "AGOrderingAnalysisPass": 2.5377185344696045,
+      "AffinePredicateResolution": 0.09365105628967285,
+      "AliasDependencyElimination": 0.0029387474060058594,
+      "AliasDependencyInduction": 0.9149439334869385,
+      "AliasDependencyReset": 0.9402039051055908,
+      "BFComputeCutting": 0.08754968643188477,
+      "BirCodeGenLoop": 1.745743751525879,
+      "CCOpFusion": 0.6600353717803955,
+      "CanonicalizeDAGForPGTiling": 0.18796682357788086,
+      "CanonicalizeIR": 0.11113405227661133,
+      "CoalesceCCOp": 0.17152976989746094,
+      "CommuteConcat": 0.03446626663208008,
+      "DMALocalityOpt": 0.03101348876953125,
+      "DMAProfiler": 0.07795047760009766,
+      "DMATilingProfiler": 0.07289671897888184,
+      "DataLocalityOpt": 2.36210560798645,
+      "DataStreaming": 0.11712121963500977,
+      "DeConcat": 0.018580198287963867,
+      "DeadCodeElimination": 0.034860849380493164,
+      "DeadStoreElimination": 1.3254823684692383,
+      "DelinearIndices": 0.3446221351623535,
+      "Delinearization": 0.1506659984588623,
+      "DoNothing": 0.00012350082397460938,
+      "DramToDramTranspose": 1.202136516571045,
+      "DumpGraphAndMetadata": 0.16636061668395996,
+      "EliminateDivs": 0.2120811939239502,
+      "ExpandBatchNorm": 0.08646512031555176,
+      "ExpandISAMacro": 0.07089924812316895,
+      "FactorizeBlkDims": 0.28119874000549316,
+      "FactorizeThreadAxesInFreeDims": 0.04656529426574707,
+      "FlattenMacroLoop": 0.09647870063781738,
+      "GenericAccessSimplifier": 0.03339672088623047,
+      "InferInitValue": 1.0866944789886475,
+      "InferIntrinsicOnCC": 0.36792969703674316,
+      "InferNeuronTensor": 1.8716816902160645,
+      "InferNonlocalTensors": 4.868690013885498,
+      "InferPSumTensor": 1.184112787246704,
+      "InlineNativeKernels": 0.046529531478881836,
+      "InsertIOTransposes": 0.8822572231292725,
+      "InsertLocalTransposes": 0.9775772094726563,
+      "InsertOffloadedTransposes": 0.07813811302185059,
+      "LICM": 0.09977531433105469,
+      "LateLegalizeInst": 0.1516575813293457,
+      "LateLegalizePostSplit": 0.07948708534240723,
+      "LateLowerReshapeOp": 0.044101715087890625,
+      "LateLowerTensorOp": 0.9023852348327637,
+      "LateNeuronInstComb": 0.3992629051208496,
+      "LayoutPreprocessing": 1.0312676429748535,
+      "LayoutPreprocessingAndAnalysis": 1.378429651260376,
+      "LayoutRequirementAnalysis": 0.3332996368408203,
+      "LegalizeCCOpLayout": 0.11418366432189941,
+      "LegalizeOpLevelAlias": 0.030529260635375977,
+      "LegalizePartitionReduce": 0.03310394287109375,
+      "LegalizeSundaAccess": 1.0573086738586426,
+      "LegalizeSundaMacro": 0.32631635665893555,
+      "LegalizeType": 0.15149402618408203,
+      "LocalLayoutOpt": 0.736790657043457,
+      "LoopFusion": 0.3599417209625244,
+      "LoopSplitting": 0.03815197944641113,
+      "LowerBroadcast": 0.0502932071685791,
+      "LowerCCOpBlockAxis": 0.24752306938171387,
+      "LowerComplexBroadcast": 0.15025925636291504,
+      "LowerIntrinsics": 0.5702738761901855,
+      "LowerTensorOp": 0.6496686935424805,
+      "LowerTranspose": 0.47429561614990234,
+      "MacroGeneration": 2.619112730026245,
+      "MaskPropagation": 0.13018035888671875,
+      "MemcpyElimination": 9.560661315917969,
+      "MutateDataType": 0.04511690139770508,
+      "NeuronAliasDependencyInduction": 0.01844191551208496,
+      "NeuronAliasDependencyReset": 0.032405853271484375,
+      "NeuronInstComb": 0.313948392868042,
+      "NeuronLICM": 0.2585768699645996,
+      "NeuronLoopFusion": 0.7092890739440918,
+      "NeuronLoopInterchange": 0.0486292839050293,
+      "NeuronSimplifier": 0.3838067054748535,
+      "NeuronSimplifyPredicates": 0.05862069129943848,
+      "NeuronValueNumbering": 0.09080743789672852,
+      "OptimizeAliasedCopyChain": 0.01573491096496582,
+      "OptimizeNKIKernels": 0.37357568740844727,
+      "PAGLayoutOpt": 43.78666687011719,
+      "PComputeCutting": 0.2951951026916504,
+      "PGLayoutTilingPipeline": 58.915504455566406,
+      "PGTiling": 5.879387855529785,
+      "PadElimination": 0.015177726745605469,
+      "ParAxesAnnotation": 42.794742584228516,
+      "PartialLoopFusion": 0.3791334629058838,
+      "PartialSimdFusion": 0.28896236419677734,
+      "PerfectLoopNest": 0.06396722793579102,
+      "RecognizeOpIdiom": 0.14191770553588867,
+      "Recompute": 0.008079290390014648,
+      "RelaxPredicates": 0.12253475189208984,
+      "Rematerialization": 0.15925836563110352,
+      "ReshapeWeights": 0.02360057830810547,
+      "ResolveAccessConflict": 0.23374605178833008,
+      "ResolveComplicatePredicates": 0.09412980079650879,
+      "RewriteReplicationMatmul": 0.05128669738769531,
+      "RewriteWeights": 0.069183349609375,
+      "SFKVectorizer": 5.667294502258301,
+      "SimpleAllReduceTiling": 0.059381961822509766,
+      "Simplifier": 0.11846542358398438,
+      "SimplifyMacroPredicates": 0.20037221908569336,
+      "SimplifyNeuronTensor": 0.42980480194091797,
+      "SimplifySlice": 0.03480982780456543,
+      "SimplifyTensor": 0.21754741668701172,
+      "SpillPSum": 0.5781185626983643,
+      "SplitAPUnionSets": 0.2246379852294922,
+      "SplitAccGrp": 0.041144371032714844,
+      "StaticProfiler": 0.11391091346740723,
+      "StaticTransposeLocalTensor": 0.2581653594970703,
+      "SundaISel": 1.714808702468872,
+      "TCTransform": 0.03960013389587402,
+      "TensorInitialization": 0.08710885047912598,
+      "TensorOpSimplifier": 0.5631639957427979,
+      "TensorOpTransform": 1.7516539096832275,
+      "TileCCOps": 0.2776148319244385,
+      "TilingProfiler": 0.4766700267791748,
+      "TransformConvOp": 0.05919003486633301,
+      "TritiumFusion": 1.33201265335083,
+      "ValueNumbering": 0.09897065162658691,
+      "VectorizeDMA": 0.1225745677947998,
+      "VectorizeMatMult": 0.0306241512298584,
+      "WeightCoalescing": 0.047003984451293945,
+      "ZeroSizeTensorElimination": 0.0004246234893798828
+    },
+    "tensorizer": {
+      "DMATilingProfiler::TotalInstructionsAfterTiling": 241479,
+      "StaticProfiler::AifUb": 4.066871166229248,
+      "StaticProfiler::ArithmeticIntensityTensorizer": 4.706264019012451,
+      "StaticProfiler::AverageDmaLength": 6434.17578125,
+      "StaticProfiler::AverageFractalPeUtilization": 99.41757202148438,
+      "StaticProfiler::AveragePartitionUtilization": 98.54876708984375,
+      "StaticProfiler::AveragePeUtilization": 97.57201385498047,
+      "StaticProfiler::DDRTransferBytes": 7193317884,
+      "StaticProfiler::InternalTransferBytes": 47988784,
+      "StaticProfiler::LoadExpanded": 1054211,
+      "StaticProfiler::LocalizationEfficiency": 115.72197723388672,
+      "StaticProfiler::LocalizationEfficiencyIgnoreNonlocal": 115.86865234375,
+      "StaticProfiler::StoreExpanded": 2218,
+      "StaticProfiler::TotalDMAExpanded": 1056429,
+      "StaticProfiler::TotalDynamicInstancesCount": 251892,
+      "StaticProfiler::TotalDynamicInstancesWithMmPackedCount": 247287,
+      "StaticProfiler::TotalLNCComm": 0,
+      "StaticProfiler::TotalLNCCommTransfer": 0,
+      "TilingProfiler::AveragePartitionUtilizationAfterTiling": 0,
+      "TilingProfiler::AveragePeUtilizationAfterTiling": 0,
+      "TilingProfiler::BatchnormInstructionsAfterTiling": 0,
+      "TilingProfiler::DmaInstructionsAfterTiling": 0,
+      "TilingProfiler::GenericInstructionsAfterTiling": 281,
+      "TilingProfiler::MatMultInstructionsAfterTiling": 220160,
+      "TilingProfiler::NumPfTransposes": 394,
+      "TilingProfiler::NumPfTransposesForIo": 33,
+      "TilingProfiler::NumPfTransposesForLocal": 226,
+      "TilingProfiler::NumPfTransposesForNonlocal": 135,
+      "TilingProfiler::PfTransposeInstructions": 6226,
+      "TilingProfiler::PfTransposeInstructionsForIo": 1026,
+      "TilingProfiler::PfTransposeInstructionsForLocal": 964,
+      "TilingProfiler::PfTransposeInstructionsForNonlocal": 4236,
+      "TilingProfiler::ReduceInstructionsAfterTiling": 67,
+      "TilingProfiler::SimdInstructionsAfterTiling": 6399,
+      "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_bk1/neuron_config.json

@@ -0,0 +1,213 @@
+{
+  "_attn_implementation_autoset": false,
+  "_name_or_path": "mistralai/Mistral-7B-Instruct-v0.3",
+  "add_cross_attention": false,
+  "architectures": [
+    "MistralForCausalLM"
+  ],
+  "attention_dropout": 0.0,
+  "attribute_map": {},
+  "bad_words_ids": null,
+  "begin_suppress_tokens": null,
+  "bos_token_id": 1,
+  "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": 2,
+  "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": 14336,
+  "is_decoder": false,
+  "is_encoder_decoder": false,
+  "label2id": {
+    "LABEL_0": 0,
+    "LABEL_1": 1
+  },
+  "length_penalty": 1.0,
+  "max_length": 20,
+  "max_position_embeddings": 32768,
+  "metadata": null,
+  "min_length": 0,
+  "model_type": "mistral",
+  "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": "NeuronLlamaAttention",
+    "attn_kernel_enabled": null,
+    "attn_tkg_builtin_kernel_enabled": false,
+    "attn_tkg_nki_kernel_enabled": false,
+    "batch_size": 4,
+    "bucket_n_active_tokens": false,
+    "buckets": [
+      256
+    ],
+    "cast_type": "config",
+    "cc_pipeline_tiling_factor": 1,
+    "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": false,
+    "is_prefix_caching": false,
+    "k_cache_transposed": false,
+    "kv_cache_batch_size": 4,
+    "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": 4,
+    "max_context_length": 2048,
+    "max_length": 2048,
+    "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": 1,
+    "n_positions": 2048,
+    "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": 2048,
+    "pa_num_blocks": 4,
+    "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": 2048,
+    "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": 4,
+    "speculation_length": 0,
+    "start_rank_id": 0,
+    "target": null,
+    "tile_cc": false,
+    "tkg_batch_size": 4,
+    "token_generation_buckets": [
+      256
+    ],
+    "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": 32,
+  "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-05,
+  "rope_theta": 1000000.0,
+  "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": "4.42.0.dev0",
+  "typical_p": 1.0,
+  "use_bfloat16": false,
+  "use_cache": true,
+  "vocab_size": 32768
+}

token_generation_model/_tp0_bk2/command.txt

@@ -0,0 +1 @@
+neuronx-cc compile --framework=XLA model.MODULE_2f686dc6ba7ef3326a56+6113de8c.hlo_module.pb --output model.MODULE_2f686dc6ba7ef3326a56+6113de8c.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 --verbose=35

token_generation_model/_tp0_bk2/compile_flags.MODULE_2f686dc6ba7ef3326a56+6113de8c.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=/models/mistral-7b-v0.3-instruct-neuronx/token_generation_model/_tp0_bk2/log-neuron-cc.txt"]

token_generation_model/_tp0_bk2/global_metric_store.json

@@ -0,0 +1,540 @@
+{
+  "Average": {
+    "tensorizer": {
+      "StaticProfiler::AverageFractalPeUtilization": 99.39725494384766,
+      "StaticProfiler::AveragePartitionUtilization": 98.54637908935547,
+      "StaticProfiler::AveragePeUtilization": 97.55433654785156,
+      "StaticProfiler::LocalizationEfficiency": 114.65234375,
+      "StaticProfiler::LocalizationEfficiencyIgnoreNonlocal": 114.79631042480469,
+      "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": 2.565244436264038,
+      "AffinePredicateResolution": 0.08378434181213379,
+      "AliasDependencyElimination": 0.003157377243041992,
+      "AliasDependencyInduction": 0.6296179294586182,
+      "AliasDependencyReset": 0.6559426784515381,
+      "BFComputeCutting": 0.08006763458251953,
+      "BirCodeGenLoop": 1.8334097862243652,
+      "CCOpFusion": 0.7150745391845703,
+      "CanonicalizeConv": 3.999999989900971e-06,
+      "CanonicalizeDAGForPGTiling": 0.18795156478881836,
+      "CanonicalizeForTensorizer": 0.0006089999806135893,
+      "CanonicalizeIR": 0.0996088981628418,
+      "Canonicalizer": 0.007222999818623066,
+      "CoalesceCCOp": 0.17608022689819336,
+      "CommuteConcat": 0.036467552185058594,
+      "DMALocalityOpt": 0.03174924850463867,
+      "DMAProfiler": 0.07987165451049805,
+      "DMATilingProfiler": 0.07628154754638672,
+      "DataLocalityOpt": 2.366703510284424,
+      "DataStreaming": 0.11896848678588867,
+      "DeConcat": 0.024710893630981445,
+      "DeadCodeElimination": 0.03867483139038086,
+      "DeadStoreElimination": 1.3142352104187012,
+      "DelinearIndices": 0.33995985984802246,
+      "Delinearization": 0.1597750186920166,
+      "DoNothing": 0.00028634071350097656,
+      "DramToDramTranspose": 1.1566178798675537,
+      "DumpGraphAndMetadata": 0.16752195358276367,
+      "EliminateDivs": 0.16330885887145996,
+      "ExpandBatchNorm": 0.11280179023742676,
+      "ExpandISAMacro": 0.07273101806640625,
+      "FactorizeBlkDims": 0.5880463123321533,
+      "FactorizeThreadAxesInFreeDims": 0.05299186706542969,
+      "FlattenMacroLoop": 0.08956408500671387,
+      "GenericAccessSimplifier": 0.033487796783447266,
+      "HoistCompute": 7.000000186963007e-05,
+      "IdentifyCrossPassTensors": 0.0003000000142492354,
+      "InferInitValue": 1.0861613750457764,
+      "InferIntrinsicOnCC": 0.3759927749633789,
+      "InferNeuronTensor": 1.848968744277954,
+      "InferNonlocalTensors": 4.863279342651367,
+      "InferPSumTensor": 1.23720383644104,
+      "InlineNativeKernels": 0.053342580795288086,
+      "InsertIOTransposes": 0.9273536205291748,
+      "InsertLocalTransposes": 0.7733509540557861,
+      "InsertOffloadedTransposes": 0.07759976387023926,
+      "LICM": 0.10009574890136719,
+      "LateLegalizeInst": 0.15100622177124023,
+      "LateLegalizePostSplit": 0.08344554901123047,
+      "LateLowerReshapeOp": 0.04735827445983887,
+      "LateLowerTensorOp": 0.9971563816070557,
+      "LateNeuronInstComb": 0.4363434314727783,
+      "LayoutPreprocessing": 1.0707452297210693,
+      "LayoutPreprocessingAndAnalysis": 1.4177396297454834,
+      "LayoutRequirementAnalysis": 0.33483314514160156,
+      "LegalizeCCOpLayout": 0.12552833557128906,
+      "LegalizeOpLevelAlias": 0.03364896774291992,
+      "LegalizePartitionReduce": 0.07460308074951172,
+      "LegalizeSundaAccess": 1.0685956478118896,
+      "LegalizeSundaMacro": 0.3450655937194824,
+      "LegalizeType": 0.16722607612609863,
+      "LocalLayoutOpt": 0.7447531223297119,
+      "LoopFusion": 0.3385646343231201,
+      "LoopSplitting": 0.03500247001647949,
+      "LowerBroadcast": 0.09404349327087402,
+      "LowerCCOpBlockAxis": 0.24126768112182617,
+      "LowerComplexBroadcast": 0.15259552001953125,
+      "LowerIntrinsics": 0.5980756282806396,
+      "LowerTensorOp": 0.6645591259002686,
+      "LowerTranspose": 0.4301431179046631,
+      "MacroGeneration": 2.585885763168335,
+      "MaskPropagation": 0.13078856468200684,
+      "MemcastMotion": 0.0001500000071246177,
+      "MemcpyElimination": 9.53384780883789,
+      "MutateDataType": 0.04887223243713379,
+      "NeuronAliasDependencyInduction": 0.01924896240234375,
+      "NeuronAliasDependencyReset": 0.034535884857177734,
+      "NeuronInstComb": 0.3494391441345215,
+      "NeuronLICM": 0.252178430557251,
+      "NeuronLoopFusion": 0.7606415748596191,
+      "NeuronLoopInterchange": 0.04935908317565918,
+      "NeuronSimplifier": 0.3867678642272949,
+      "NeuronSimplifyPredicates": 0.06488251686096191,
+      "NeuronValueNumbering": 0.0907585620880127,
+      "OptimizeAliasedCopyChain": 0.028003215789794922,
+      "OptimizeNKIKernels": 0.3752737045288086,
+      "PAGLayoutOpt": 43.734619140625,
+      "PComputeCutting": 0.3206455707550049,
+      "PGLayoutTilingPipeline": 58.89807891845703,
+      "PGTiling": 5.892523765563965,
+      "PadElimination": 0.015540599822998047,
+      "ParAxesAnnotation": 42.94718551635742,
+      "PartialLoopFusion": 0.41214418411254883,
+      "PartialSimdFusion": 0.3176724910736084,
+      "PenguinizeFunctions": 0.00025400001322850585,
+      "PerfectLoopNest": 0.06399226188659668,
+      "PruneFunctions": 0.0001340000017080456,
+      "RecognizeOpIdiom": 0.15635156631469727,
+      "Recompute": 0.009923696517944336,
+      "RelaxPredicates": 0.12309074401855469,
+      "Rematerialization": 0.22021842002868652,
+      "RemoveOptimizationBarriers": 0.00010699999984353781,
+      "ReshapeWeights": 0.026267051696777344,
+      "ResolveAccessConflict": 0.22898530960083008,
+      "ResolveComplicatePredicates": 0.09242081642150879,
+      "RewriteReplicationMatmul": 0.051862239837646484,
+      "RewriteWeights": 0.0745856761932373,
+      "SFKVectorizer": 6.074093341827393,
+      "ScatterMotion": 0.007073000073432922,
+      "SimpleAllReduceTiling": 0.0595548152923584,
+      "Simplifier": 0.1168510913848877,
+      "SimplifyMacroPredicates": 0.19426846504211426,
+      "SimplifyNeuronTensor": 0.43988728523254395,
+      "SimplifySlice": 0.07861804962158203,
+      "SimplifyTensor": 0.22563552856445313,
+      "SpillPSum": 0.6065900325775146,
+      "SplitAPUnionSets": 0.2475893497467041,
+      "SplitAccGrp": 0.040537118911743164,
+      "StaticProfiler": 0.12209963798522949,
+      "StaticTransposeLocalTensor": 0.26148557662963867,
+      "SundaISel": 1.426959753036499,
+      "TCTransform": 0.04041242599487305,
+      "TensorInitialization": 0.0903174877166748,
+      "TensorOpSimplifier": 0.5430335998535156,
+      "TensorOpTransform": 2.182173252105713,
+      "TensorizerLegalizationPass": 0.0002280000044265762,
+      "TileCCOps": 0.2870340347290039,
+      "TilingProfiler": 0.4631321430206299,
+      "TransformConvOp": 0.07256841659545898,
+      "TritiumFusion": 1.387312412261963,
+      "ValueNumbering": 0.10227847099304199,
+      "VectorizeDMA": 0.12761187553405762,
+      "VectorizeMatMult": 0.02546095848083496,
+      "VerifySupportedOps": 0.0002390000008745119,
+      "WeightCoalescing": 0.04958319664001465,
+      "ZeroSizeTensorElimination": 0.00046133995056152344,
+      "algsimp": 0.0017790000420063734,
+      "batchnorm_expander": 0.0009469999931752682,
+      "boundary-marker-removal": 0.0003819999983534217,
+      "call-inliner": 0.0003330000035930425,
+      "canonicalize-boundary-marker": 0.0004419999895617366,
+      "collective-stream-id-checker": 7.999999797903001e-05,
+      "comparison-expander": 0.00040600000647827983,
+      "computation-deduplicator": 0.0006760000251233578,
+      "conditional-to-select": 0.00012599999899975955,
+      "config-lowering": 0.0003220000071451068,
+      "constant_folding": 0.00023200000578071922,
+      "cse": 0.0005840000230818987,
+      "dce": 6.500000017695129e-05,
+      "dynamic-slice-transpose": 0.00021100000594742596,
+      "eliminate-redundant-compare": 0.00021699999342672527,
+      "emit-offloaded-dropout": 0.00032699998700991273,
+      "flatten-call-graph": 0.0005499999970197678,
+      "fuse-send-recv": 0.002495999913662672,
+      "hilo::LegalizeAlias": 0.0030300000216811895,
+      "hilo::NeuronInstCombine": 0.0010979999788105488,
+      "hilo::NeuronOpFusion": 0.0006970000104047358,
+      "hilo::ReplaceTokenTypeWithU8Pass": 0.00018200000340584666,
+      "hilo::ScheduleFusion": 5.8000001445179805e-05,
+      "hilo::SixtyFourHack": 0.0002469999890308827,
+      "hilo::VerifyAliasing": 0.00012199999764561653,
+      "hlo-mac-count": 0.0007970000151544809,
+      "hlo-verifier": 0.009019999764859676,
+      "io-con-pipe-begin": 7.000000096013537e-06,
+      "io-con-pipe-end": 9.999999974752427e-07,
+      "io-layout-normalization": 0.0010720000136643648,
+      "legalize-ccops": 2.5999999706982635e-05,
+      "legalize-compare": 0.0003650000144261867,
+      "lower-argminmax-custom-call": 0.0002280000044265762,
+      "map-inline": 0.0006910000229254365,
+      "metadata-naming": 0.002133999951183796,
+      "mlir::detail::OpToOpPassAdaptor": 0.00021600000036414713,
+      "mlir::hlo::MhloToPyPenguin": 0.05840799957513809,
+      "mlir::mhlo::LowerComplexExtraPass": 0.004083000123500824,
+      "mlir::mhlo::LowerComplexPass": 0.004681000020354986,
+      "native-to-custom-softmax": 0.00060299999313429,
+      "native-to-custom-softmax-dx": 0.05077800154685974,
+      "operand_upcaster": 0.0009350000182166696,
+      "post-par-pipe-begin": 9.999999974752427e-07,
+      "post-par-pipe-end": 0.0,
+      "post-partition-simplification": 0.10879900306463242,
+      "pre-hlo-begin": 1.9999999949504854e-06,
+      "pre-hlo-end": 9.999999974752427e-07,
+      "replace-minimum-constant": 0.00022600000374950469,
+      "reshape-mover": 9.300000237999484e-05,
+      "simplify-concat": 0.0023070001043379307,
+      "simplify-while-loops": 9.000000136438757e-05,
+      "transform-variadic-reduce": 0.0006370000191964209,
+      "tuple-simplifier": 0.0002800000074785203,
+      "unpack-nested-aws-ntwsr": 0.0003440000000409782,
+      "unroll-while-loop": 1.5999999959603883e-05
+    },
+    "hilo": {
+      "HloMacCount": 14495776768.0,
+      "Traffic": 7787344384.0
+    },
+    "tensorizer": {
+      "DMATilingProfiler::TotalInstructionsAfterTiling": 245899,
+      "StaticProfiler::AifUb": 4.623864650726318,
+      "StaticProfiler::ArithmeticIntensityTensorizer": 5.3013691902160645,
+      "StaticProfiler::AverageDmaLength": 6491.51513671875,
+      "StaticProfiler::DDRTransferBytes": 7260430844,
+      "StaticProfiler::InternalTransferBytes": 82595888,
+      "StaticProfiler::LoadExpanded": 1054219,
+      "StaticProfiler::StoreExpanded": 2218,
+      "StaticProfiler::TotalDMAExpanded": 1056437,
+      "StaticProfiler::TotalDynamicInstancesCount": 255460,
+      "StaticProfiler::TotalDynamicInstancesWithMmPackedCount": 250753,
+      "StaticProfiler::TotalLNCComm": 0,
+      "StaticProfiler::TotalLNCCommTransfer": 0,
+      "TilingProfiler::BatchnormInstructionsAfterTiling": 0,
+      "TilingProfiler::DmaInstructionsAfterTiling": 0,
+      "TilingProfiler::GenericInstructionsAfterTiling": 281,
+      "TilingProfiler::MatMultInstructionsAfterTiling": 222208,
+      "TilingProfiler::NumPfTransposes": 330,
+      "TilingProfiler::NumPfTransposesForIo": 33,
+      "TilingProfiler::NumPfTransposesForLocal": 162,
+      "TilingProfiler::NumPfTransposesForNonlocal": 135,
+      "TilingProfiler::PfTransposeInstructions": 7220,
+      "TilingProfiler::PfTransposeInstructionsForIo": 2052,
+      "TilingProfiler::PfTransposeInstructionsForLocal": 932,
+      "TilingProfiler::PfTransposeInstructionsForNonlocal": 4236,
+      "TilingProfiler::ReduceInstructionsAfterTiling": 163,
+      "TilingProfiler::SimdInstructionsAfterTiling": 6657,
+      "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": 3.999999989900971e-06,
+      "CanonicalizeForTensorizer": 0.0006089999806135893,
+      "Canonicalizer": 0.007222999818623066,
+      "HoistCompute": 7.000000186963007e-05,
+      "IdentifyCrossPassTensors": 0.0003000000142492354,
+      "MemcastMotion": 0.0001500000071246177,
+      "PenguinizeFunctions": 0.00025400001322850585,
+      "PruneFunctions": 0.0001340000017080456,
+      "RemoveOptimizationBarriers": 0.00010699999984353781,
+      "ScatterMotion": 0.007073000073432922,
+      "TensorizerLegalizationPass": 0.0002280000044265762,
+      "VerifySupportedOps": 0.0002390000008745119,
+      "algsimp": 0.0017790000420063734,
+      "batchnorm_expander": 0.0009469999931752682,
+      "boundary-marker-removal": 0.0003819999983534217,
+      "call-inliner": 0.0003330000035930425,
+      "canonicalize-boundary-marker": 0.0004419999895617366,
+      "collective-stream-id-checker": 7.999999797903001e-05,
+      "comparison-expander": 0.00040600000647827983,
+      "computation-deduplicator": 0.0006760000251233578,
+      "conditional-to-select": 0.00012599999899975955,
+      "config-lowering": 0.0003220000071451068,
+      "constant_folding": 0.00023200000578071922,
+      "cse": 0.0005840000230818987,
+      "dce": 6.500000017695129e-05,
+      "dynamic-slice-transpose": 0.00021100000594742596,
+      "eliminate-redundant-compare": 0.00021699999342672527,
+      "emit-offloaded-dropout": 0.00032699998700991273,
+      "flatten-call-graph": 0.0005499999970197678,
+      "fuse-send-recv": 0.002495999913662672,
+      "hilo::LegalizeAlias": 0.0030300000216811895,
+      "hilo::NeuronInstCombine": 0.0010979999788105488,
+      "hilo::NeuronOpFusion": 0.0006970000104047358,
+      "hilo::ReplaceTokenTypeWithU8Pass": 0.00018200000340584666,
+      "hilo::ScheduleFusion": 5.8000001445179805e-05,
+      "hilo::SixtyFourHack": 0.0002469999890308827,
+      "hilo::VerifyAliasing": 0.00012199999764561653,
+      "hlo-mac-count": 0.0007970000151544809,
+      "hlo-verifier": 0.009019999764859676,
+      "io-con-pipe-begin": 7.000000096013537e-06,
+      "io-con-pipe-end": 9.999999974752427e-07,
+      "io-layout-normalization": 0.0010720000136643648,
+      "legalize-ccops": 2.5999999706982635e-05,
+      "legalize-compare": 0.0003650000144261867,
+      "lower-argminmax-custom-call": 0.0002280000044265762,
+      "map-inline": 0.0006910000229254365,
+      "metadata-naming": 0.002133999951183796,
+      "mlir::detail::OpToOpPassAdaptor": 0.00021600000036414713,
+      "mlir::hlo::MhloToPyPenguin": 0.05840799957513809,
+      "mlir::mhlo::LowerComplexExtraPass": 0.004083000123500824,
+      "mlir::mhlo::LowerComplexPass": 0.004681000020354986,
+      "native-to-custom-softmax": 0.00060299999313429,
+      "native-to-custom-softmax-dx": 0.05077800154685974,
+      "operand_upcaster": 0.0009350000182166696,
+      "post-par-pipe-begin": 9.999999974752427e-07,
+      "post-par-pipe-end": 0.0,
+      "post-partition-simplification": 0.10879900306463242,
+      "pre-hlo-begin": 1.9999999949504854e-06,
+      "pre-hlo-end": 9.999999974752427e-07,
+      "replace-minimum-constant": 0.00022600000374950469,
+      "reshape-mover": 9.300000237999484e-05,
+      "simplify-concat": 0.0023070001043379307,
+      "simplify-while-loops": 9.000000136438757e-05,
+      "transform-variadic-reduce": 0.0006370000191964209,
+      "tuple-simplifier": 0.0002800000074785203,
+      "unpack-nested-aws-ntwsr": 0.0003440000000409782,
+      "unroll-while-loop": 1.5999999959603883e-05
+    }
+  },
+  "cumsum": {
+    "compiletime": {
+      "CoalesceCCOp": 0.00025391578674316406,
+      "DMALocalityOpt": 0.00021147727966308594,
+      "DMAProfiler": 0.0009748935699462891,
+      "DataStreaming": 0.0003063678741455078,
+      "DoNothing": 0.00016117095947265625,
+      "ExpandISAMacro": 0.0005402565002441406,
+      "FactorizeBlkDims": 0.00048279762268066406,
+      "InferPSumTensor": 0.0004856586456298828,
+      "LateLegalizeInst": 0.0004565715789794922,
+      "LateNeuronInstComb": 0.0005049705505371094,
+      "LegalizeSundaAccess": 0.0016589164733886719,
+      "LegalizeType": 0.000293731689453125,
+      "LowerBroadcast": 0.00036835670471191406,
+      "LowerIntrinsics": 0.0002689361572265625,
+      "LowerTranspose": 0.0002696514129638672,
+      "NeuronInstComb": 0.0005173683166503906,
+      "NeuronLICM": 0.000446319580078125,
+      "NeuronSimplifyPredicates": 0.002989530563354492,
+      "NeuronValueNumbering": 0.0004544258117675781,
+      "SFKVectorizer": 0.002811908721923828,
+      "SimpleAllReduceTiling": 0.00024199485778808594,
+      "SimplifyNeuronTensor": 0.0004520416259765625,
+      "SpillPSum": 0.0006034374237060547,
+      "WeightCoalescing": 0.0002529621124267578
+    }
+  },
+  "sg00": {
+    "hilo": {
+      "ArithmeticIntensity": 3.7229063510894775,
+      "HloMacCount": 14495776768.0,
+      "Traffic": 7787344384.0
+    }
+  },
+  "sg0000": {
+    "compiletime": {
+      "AGOrderingAnalysisPass": 2.565244436264038,
+      "AffinePredicateResolution": 0.08378434181213379,
+      "AliasDependencyElimination": 0.003157377243041992,
+      "AliasDependencyInduction": 0.6296179294586182,
+      "AliasDependencyReset": 0.6559426784515381,
+      "BFComputeCutting": 0.08006763458251953,
+      "BirCodeGenLoop": 1.8334097862243652,
+      "CCOpFusion": 0.7150745391845703,
+      "CanonicalizeDAGForPGTiling": 0.18795156478881836,
+      "CanonicalizeIR": 0.0996088981628418,
+      "CoalesceCCOp": 0.1758263111114502,
+      "CommuteConcat": 0.036467552185058594,
+      "DMALocalityOpt": 0.031537771224975586,
+      "DMAProfiler": 0.07889676094055176,
+      "DMATilingProfiler": 0.07628154754638672,
+      "DataLocalityOpt": 2.366703510284424,
+      "DataStreaming": 0.11866211891174316,
+      "DeConcat": 0.024710893630981445,
+      "DeadCodeElimination": 0.03867483139038086,
+      "DeadStoreElimination": 1.3142352104187012,
+      "DelinearIndices": 0.33995985984802246,
+      "Delinearization": 0.1597750186920166,
+      "DoNothing": 0.0001251697540283203,
+      "DramToDramTranspose": 1.1566178798675537,
+      "DumpGraphAndMetadata": 0.16752195358276367,
+      "EliminateDivs": 0.16330885887145996,
+      "ExpandBatchNorm": 0.11280179023742676,
+      "ExpandISAMacro": 0.07219076156616211,
+      "FactorizeBlkDims": 0.5875635147094727,
+      "FactorizeThreadAxesInFreeDims": 0.05299186706542969,
+      "FlattenMacroLoop": 0.08956408500671387,
+      "GenericAccessSimplifier": 0.033487796783447266,
+      "InferInitValue": 1.0861613750457764,
+      "InferIntrinsicOnCC": 0.3759927749633789,
+      "InferNeuronTensor": 1.848968744277954,
+      "InferNonlocalTensors": 4.863279342651367,
+      "InferPSumTensor": 1.2367181777954102,
+      "InlineNativeKernels": 0.053342580795288086,
+      "InsertIOTransposes": 0.9273536205291748,
+      "InsertLocalTransposes": 0.7733509540557861,
+      "InsertOffloadedTransposes": 0.07759976387023926,
+      "LICM": 0.10009574890136719,
+      "LateLegalizeInst": 0.15054965019226074,
+      "LateLegalizePostSplit": 0.08344554901123047,
+      "LateLowerReshapeOp": 0.04735827445983887,
+      "LateLowerTensorOp": 0.9971563816070557,
+      "LateNeuronInstComb": 0.4358384609222412,
+      "LayoutPreprocessing": 1.0707452297210693,
+      "LayoutPreprocessingAndAnalysis": 1.4177396297454834,
+      "LayoutRequirementAnalysis": 0.33483314514160156,
+      "LegalizeCCOpLayout": 0.12552833557128906,
+      "LegalizeOpLevelAlias": 0.03364896774291992,
+      "LegalizePartitionReduce": 0.07460308074951172,
+      "LegalizeSundaAccess": 1.066936731338501,
+      "LegalizeSundaMacro": 0.3450655937194824,
+      "LegalizeType": 0.1669323444366455,
+      "LocalLayoutOpt": 0.7447531223297119,
+      "LoopFusion": 0.3385646343231201,
+      "LoopSplitting": 0.03500247001647949,
+      "LowerBroadcast": 0.09367513656616211,
+      "LowerCCOpBlockAxis": 0.24126768112182617,
+      "LowerComplexBroadcast": 0.15259552001953125,
+      "LowerIntrinsics": 0.5978066921234131,
+      "LowerTensorOp": 0.6645591259002686,
+      "LowerTranspose": 0.4298734664916992,
+      "MacroGeneration": 2.585885763168335,
+      "MaskPropagation": 0.13078856468200684,
+      "MemcpyElimination": 9.53384780883789,
+      "MutateDataType": 0.04887223243713379,
+      "NeuronAliasDependencyInduction": 0.01924896240234375,
+      "NeuronAliasDependencyReset": 0.034535884857177734,
+      "NeuronInstComb": 0.3489217758178711,
+      "NeuronLICM": 0.25173211097717285,
+      "NeuronLoopFusion": 0.7606415748596191,
+      "NeuronLoopInterchange": 0.04935908317565918,
+      "NeuronSimplifier": 0.3867678642272949,
+      "NeuronSimplifyPredicates": 0.06189298629760742,
+      "NeuronValueNumbering": 0.09030413627624512,
+      "OptimizeAliasedCopyChain": 0.028003215789794922,
+      "OptimizeNKIKernels": 0.3752737045288086,
+      "PAGLayoutOpt": 43.734619140625,
+      "PComputeCutting": 0.3206455707550049,
+      "PGLayoutTilingPipeline": 58.89807891845703,
+      "PGTiling": 5.892523765563965,
+      "PadElimination": 0.015540599822998047,
+      "ParAxesAnnotation": 42.94718551635742,
+      "PartialLoopFusion": 0.41214418411254883,
+      "PartialSimdFusion": 0.3176724910736084,
+      "PerfectLoopNest": 0.06399226188659668,
+      "RecognizeOpIdiom": 0.15635156631469727,
+      "Recompute": 0.009923696517944336,
+      "RelaxPredicates": 0.12309074401855469,
+      "Rematerialization": 0.22021842002868652,
+      "ReshapeWeights": 0.026267051696777344,
+      "ResolveAccessConflict": 0.22898530960083008,
+      "ResolveComplicatePredicates": 0.09242081642150879,
+      "RewriteReplicationMatmul": 0.051862239837646484,
+      "RewriteWeights": 0.0745856761932373,
+      "SFKVectorizer": 6.071281433105469,
+      "SimpleAllReduceTiling": 0.05931282043457031,
+      "Simplifier": 0.1168510913848877,
+      "SimplifyMacroPredicates": 0.19426846504211426,
+      "SimplifyNeuronTensor": 0.4394352436065674,
+      "SimplifySlice": 0.07861804962158203,
+      "SimplifyTensor": 0.22563552856445313,
+      "SpillPSum": 0.6059865951538086,
+      "SplitAPUnionSets": 0.2475893497467041,
+      "SplitAccGrp": 0.040537118911743164,
+      "StaticProfiler": 0.12209963798522949,
+      "StaticTransposeLocalTensor": 0.26148557662963867,
+      "SundaISel": 1.426959753036499,
+      "TCTransform": 0.04041242599487305,
+      "TensorInitialization": 0.0903174877166748,
+      "TensorOpSimplifier": 0.5430335998535156,
+      "TensorOpTransform": 2.182173252105713,
+      "TileCCOps": 0.2870340347290039,
+      "TilingProfiler": 0.4631321430206299,
+      "TransformConvOp": 0.07256841659545898,
+      "TritiumFusion": 1.387312412261963,
+      "ValueNumbering": 0.10227847099304199,
+      "VectorizeDMA": 0.12761187553405762,
+      "VectorizeMatMult": 0.02546095848083496,
+      "WeightCoalescing": 0.04933023452758789,
+      "ZeroSizeTensorElimination": 0.00046133995056152344
+    },
+    "tensorizer": {
+      "DMATilingProfiler::TotalInstructionsAfterTiling": 245899,
+      "StaticProfiler::AifUb": 4.623864650726318,
+      "StaticProfiler::ArithmeticIntensityTensorizer": 5.3013691902160645,
+      "StaticProfiler::AverageDmaLength": 6491.51513671875,
+      "StaticProfiler::AverageFractalPeUtilization": 99.39725494384766,
+      "StaticProfiler::AveragePartitionUtilization": 98.54637908935547,
+      "StaticProfiler::AveragePeUtilization": 97.55433654785156,
+      "StaticProfiler::DDRTransferBytes": 7260430844,
+      "StaticProfiler::InternalTransferBytes": 82595888,
+      "StaticProfiler::LoadExpanded": 1054219,
+      "StaticProfiler::LocalizationEfficiency": 114.65234375,
+      "StaticProfiler::LocalizationEfficiencyIgnoreNonlocal": 114.79631042480469,
+      "StaticProfiler::StoreExpanded": 2218,
+      "StaticProfiler::TotalDMAExpanded": 1056437,
+      "StaticProfiler::TotalDynamicInstancesCount": 255460,
+      "StaticProfiler::TotalDynamicInstancesWithMmPackedCount": 250753,
+      "StaticProfiler::TotalLNCComm": 0,
+      "StaticProfiler::TotalLNCCommTransfer": 0,
+      "TilingProfiler::AveragePartitionUtilizationAfterTiling": 0,
+      "TilingProfiler::AveragePeUtilizationAfterTiling": 0,
+      "TilingProfiler::BatchnormInstructionsAfterTiling": 0,
+      "TilingProfiler::DmaInstructionsAfterTiling": 0,
+      "TilingProfiler::GenericInstructionsAfterTiling": 281,
+      "TilingProfiler::MatMultInstructionsAfterTiling": 222208,
+      "TilingProfiler::NumPfTransposes": 330,
+      "TilingProfiler::NumPfTransposesForIo": 33,
+      "TilingProfiler::NumPfTransposesForLocal": 162,
+      "TilingProfiler::NumPfTransposesForNonlocal": 135,
+      "TilingProfiler::PfTransposeInstructions": 7220,
+      "TilingProfiler::PfTransposeInstructionsForIo": 2052,
+      "TilingProfiler::PfTransposeInstructionsForLocal": 932,
+      "TilingProfiler::PfTransposeInstructionsForNonlocal": 4236,
+      "TilingProfiler::ReduceInstructionsAfterTiling": 163,
+      "TilingProfiler::SimdInstructionsAfterTiling": 6657,
+      "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_bk2/neuron_config.json

@@ -0,0 +1,213 @@
+{
+  "_attn_implementation_autoset": false,
+  "_name_or_path": "mistralai/Mistral-7B-Instruct-v0.3",
+  "add_cross_attention": false,
+  "architectures": [
+    "MistralForCausalLM"
+  ],
+  "attention_dropout": 0.0,
+  "attribute_map": {},
+  "bad_words_ids": null,
+  "begin_suppress_tokens": null,
+  "bos_token_id": 1,
+  "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": 2,
+  "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": 14336,
+  "is_decoder": false,
+  "is_encoder_decoder": false,
+  "label2id": {
+    "LABEL_0": 0,
+    "LABEL_1": 1
+  },
+  "length_penalty": 1.0,
+  "max_length": 20,
+  "max_position_embeddings": 32768,
+  "metadata": null,
+  "min_length": 0,
+  "model_type": "mistral",
+  "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": "NeuronLlamaAttention",
+    "attn_kernel_enabled": null,
+    "attn_tkg_builtin_kernel_enabled": false,
+    "attn_tkg_nki_kernel_enabled": false,
+    "batch_size": 4,
+    "bucket_n_active_tokens": false,
+    "buckets": [
+      512
+    ],
+    "cast_type": "config",
+    "cc_pipeline_tiling_factor": 1,
+    "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": false,
+    "is_prefix_caching": false,
+    "k_cache_transposed": false,
+    "kv_cache_batch_size": 4,
+    "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": 4,
+    "max_context_length": 2048,
+    "max_length": 2048,
+    "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": 1,
+    "n_positions": 2048,
+    "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": 2048,
+    "pa_num_blocks": 4,
+    "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": 2048,
+    "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": 4,
+    "speculation_length": 0,
+    "start_rank_id": 0,
+    "target": null,
+    "tile_cc": false,
+    "tkg_batch_size": 4,
+    "token_generation_buckets": [
+      512
+    ],
+    "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": 32,
+  "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-05,
+  "rope_theta": 1000000.0,
+  "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": "4.42.0.dev0",
+  "typical_p": 1.0,
+  "use_bfloat16": false,
+  "use_cache": true,
+  "vocab_size": 32768
+}

token_generation_model/_tp0_bk3/command.txt

@@ -0,0 +1 @@
+neuronx-cc compile --framework=XLA model.MODULE_668122c92a86c0ce6817+f94fe8ed.hlo_module.pb --output model.MODULE_668122c92a86c0ce6817+f94fe8ed.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 --verbose=35

token_generation_model/_tp0_bk3/compile_flags.MODULE_668122c92a86c0ce6817+f94fe8ed.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=/models/mistral-7b-v0.3-instruct-neuronx/token_generation_model/_tp0_bk3/log-neuron-cc.txt"]

token_generation_model/_tp0_bk3/global_metric_store.json

@@ -0,0 +1,540 @@
+{
+  "Average": {
+    "tensorizer": {
+      "StaticProfiler::AverageFractalPeUtilization": 99.4054183959961,
+      "StaticProfiler::AveragePartitionUtilization": 98.58908081054688,
+      "StaticProfiler::AveragePeUtilization": 97.55974578857422,
+      "StaticProfiler::LocalizationEfficiency": 112.5713119506836,
+      "StaticProfiler::LocalizationEfficiencyIgnoreNonlocal": 112.71009826660156,
+      "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": 2.305126667022705,
+      "AffinePredicateResolution": 0.09757089614868164,
+      "AliasDependencyElimination": 0.0027077198028564453,
+      "AliasDependencyInduction": 0.6443400382995605,
+      "AliasDependencyReset": 0.6698822975158691,
+      "BFComputeCutting": 0.07641983032226563,
+      "BirCodeGenLoop": 2.047102689743042,
+      "CCOpFusion": 0.708749532699585,
+      "CanonicalizeConv": 0.0,
+      "CanonicalizeDAGForPGTiling": 0.18178820610046387,
+      "CanonicalizeForTensorizer": 0.0006220000213943422,
+      "CanonicalizeIR": 0.11043667793273926,
+      "Canonicalizer": 0.006990000139921904,
+      "CoalesceCCOp": 0.1608116626739502,
+      "CommuteConcat": 0.036214590072631836,
+      "DMALocalityOpt": 0.02865910530090332,
+      "DMAProfiler": 0.07138633728027344,
+      "DMATilingProfiler": 0.06614851951599121,
+      "DataLocalityOpt": 2.321251392364502,
+      "DataStreaming": 0.1097254753112793,
+      "DeConcat": 0.02409839630126953,
+      "DeadCodeElimination": 0.035909414291381836,
+      "DeadStoreElimination": 1.309647560119629,
+      "DelinearIndices": 0.3166069984436035,
+      "Delinearization": 0.12859582901000977,
+      "DoNothing": 0.00027942657470703125,
+      "DramToDramTranspose": 1.095766544342041,
+      "DumpGraphAndMetadata": 0.16588640213012695,
+      "EliminateDivs": 0.2634701728820801,
+      "ExpandBatchNorm": 0.09694433212280273,
+      "ExpandISAMacro": 0.06791496276855469,
+      "FactorizeBlkDims": 0.3919196128845215,
+      "FactorizeThreadAxesInFreeDims": 0.04469728469848633,
+      "FlattenMacroLoop": 0.07828259468078613,
+      "GenericAccessSimplifier": 0.03183865547180176,
+      "HoistCompute": 0.00010399999882793054,
+      "IdentifyCrossPassTensors": 0.0005099999834783375,
+      "InferInitValue": 1.2782227993011475,
+      "InferIntrinsicOnCC": 0.35162997245788574,
+      "InferNeuronTensor": 1.7735645771026611,
+      "InferNonlocalTensors": 4.640493392944336,
+      "InferPSumTensor": 0.9281036853790283,
+      "InlineNativeKernels": 0.04500627517700195,
+      "InsertIOTransposes": 0.8502795696258545,
+      "InsertLocalTransposes": 0.7141804695129395,
+      "InsertOffloadedTransposes": 0.06985926628112793,
+      "LICM": 0.09515905380249023,
+      "LateLegalizeInst": 0.13829779624938965,
+      "LateLegalizePostSplit": 0.07408785820007324,
+      "LateLowerReshapeOp": 0.03981423377990723,
+      "LateLowerTensorOp": 0.9939045906066895,
+      "LateNeuronInstComb": 0.3960134983062744,
+      "LayoutPreprocessing": 0.9070339202880859,
+      "LayoutPreprocessingAndAnalysis": 1.2123572826385498,
+      "LayoutRequirementAnalysis": 0.2940239906311035,
+      "LegalizeCCOpLayout": 0.1556408405303955,
+      "LegalizeOpLevelAlias": 0.03197050094604492,
+      "LegalizePartitionReduce": 0.06994032859802246,
+      "LegalizeSundaAccess": 1.0174872875213623,
+      "LegalizeSundaMacro": 0.344529390335083,
+      "LegalizeType": 0.1504347324371338,
+      "LocalLayoutOpt": 0.7021169662475586,
+      "LoopFusion": 0.3479132652282715,
+      "LoopSplitting": 0.031319618225097656,
+      "LowerBroadcast": 0.08377981185913086,
+      "LowerCCOpBlockAxis": 0.21303391456604004,
+      "LowerComplexBroadcast": 0.13837647438049316,
+      "LowerIntrinsics": 0.8480772972106934,
+      "LowerTensorOp": 0.6734035015106201,
+      "LowerTranspose": 0.4170114994049072,
+      "MacroGeneration": 2.3995394706726074,
+      "MaskPropagation": 0.11813521385192871,
+      "MemcastMotion": 0.00015700000221841037,
+      "MemcpyElimination": 9.78618049621582,
+      "MutateDataType": 0.04571127891540527,
+      "NeuronAliasDependencyInduction": 0.018005847930908203,
+      "NeuronAliasDependencyReset": 0.033030033111572266,
+      "NeuronInstComb": 0.31530284881591797,
+      "NeuronLICM": 0.2629721164703369,
+      "NeuronLoopFusion": 0.7358701229095459,
+      "NeuronLoopInterchange": 0.04501771926879883,
+      "NeuronSimplifier": 0.37318873405456543,
+      "NeuronSimplifyPredicates": 0.059458255767822266,
+      "NeuronValueNumbering": 0.08360004425048828,
+      "OptimizeAliasedCopyChain": 0.016149282455444336,
+      "OptimizeNKIKernels": 0.36123156547546387,
+      "PAGLayoutOpt": 43.12562561035156,
+      "PComputeCutting": 0.31317949295043945,
+      "PGLayoutTilingPipeline": 57.15171813964844,
+      "PGTiling": 5.401483535766602,
+      "PadElimination": 0.01123952865600586,
+      "ParAxesAnnotation": 42.39716339111328,
+      "PartialLoopFusion": 0.4405984878540039,
+      "PartialSimdFusion": 0.33167028427124023,
+      "PenguinizeFunctions": 0.00023700000019744039,
+      "PerfectLoopNest": 0.057257890701293945,
+      "PruneFunctions": 0.00019799999427050352,
+      "RecognizeOpIdiom": 0.14631319046020508,
+      "Recompute": 0.008225679397583008,
+      "RelaxPredicates": 0.11752939224243164,
+      "Rematerialization": 0.1648578643798828,
+      "RemoveOptimizationBarriers": 0.0005300000193528831,
+      "ReshapeWeights": 0.02127981185913086,
+      "ResolveAccessConflict": 0.24170351028442383,
+      "ResolveComplicatePredicates": 0.05225372314453125,
+      "RewriteReplicationMatmul": 0.042751312255859375,
+      "RewriteWeights": 0.06114816665649414,
+      "SFKVectorizer": 6.08488655090332,
+      "ScatterMotion": 0.007275999989360571,
+      "SimpleAllReduceTiling": 0.0534822940826416,
+      "Simplifier": 0.11982178688049316,
+      "SimplifyMacroPredicates": 0.18191838264465332,
+      "SimplifyNeuronTensor": 0.40709972381591797,
+      "SimplifySlice": 0.07960724830627441,
+      "SimplifyTensor": 0.19757962226867676,
+      "SpillPSum": 0.577000617980957,
+      "SplitAPUnionSets": 0.2315382957458496,
+      "SplitAccGrp": 0.03763127326965332,
+      "StaticProfiler": 0.11375260353088379,
+      "StaticTransposeLocalTensor": 0.2317829132080078,
+      "SundaISel": 1.3966825008392334,
+      "TCTransform": 0.039675235748291016,
+      "TensorInitialization": 0.13313651084899902,
+      "TensorOpSimplifier": 0.5821249485015869,
+      "TensorOpTransform": 2.0534775257110596,
+      "TensorizerLegalizationPass": 0.0002500000118743628,
+      "TileCCOps": 0.2945268154144287,
+      "TilingProfiler": 0.41756296157836914,
+      "TransformConvOp": 0.069244384765625,
+      "TritiumFusion": 1.3019096851348877,
+      "ValueNumbering": 0.09937286376953125,
+      "VectorizeDMA": 0.11787676811218262,
+      "VectorizeMatMult": 0.0218658447265625,
+      "VerifySupportedOps": 0.00025599999935366213,
+      "WeightCoalescing": 0.046278953552246094,
+      "ZeroSizeTensorElimination": 0.00039458274841308594,
+      "algsimp": 0.0017810000572353601,
+      "batchnorm_expander": 0.0009689999860711396,
+      "boundary-marker-removal": 0.00036299999919719994,
+      "call-inliner": 0.0004079999926034361,
+      "canonicalize-boundary-marker": 0.0004490000137593597,
+      "collective-stream-id-checker": 0.00010399999882793054,
+      "comparison-expander": 0.0004090000002179295,
+      "computation-deduplicator": 0.0005169999785721302,
+      "conditional-to-select": 0.00012099999730708078,
+      "config-lowering": 0.0003060000017285347,
+      "constant_folding": 0.00023099999816622585,
+      "cse": 0.0005119999987073243,
+      "dce": 6.900000153109431e-05,
+      "dynamic-slice-transpose": 0.00020799999765586108,
+      "eliminate-redundant-compare": 0.00024300000222865492,
+      "emit-offloaded-dropout": 0.0003650000144261867,
+      "flatten-call-graph": 0.0003650000144261867,
+      "fuse-send-recv": 0.002230999991297722,
+      "hilo::LegalizeAlias": 0.0036810000892728567,
+      "hilo::NeuronInstCombine": 0.001101000001654029,
+      "hilo::NeuronOpFusion": 0.00039599998854100704,
+      "hilo::ReplaceTokenTypeWithU8Pass": 0.0004830000107176602,
+      "hilo::ScheduleFusion": 5.900000178371556e-05,
+      "hilo::SixtyFourHack": 0.00026000000070780516,
+      "hilo::VerifyAliasing": 0.00015300000086426735,
+      "hlo-mac-count": 0.0007660000119358301,
+      "hlo-verifier": 0.00834800023585558,
+      "io-con-pipe-begin": 7.999999979801942e-06,
+      "io-con-pipe-end": 9.999999974752427e-07,
+      "io-layout-normalization": 0.0012280000373721123,
+      "legalize-ccops": 2.5999999706982635e-05,
+      "legalize-compare": 0.00031900001340545714,
+      "lower-argminmax-custom-call": 0.00019999999494757503,
+      "map-inline": 0.000674000009894371,
+      "metadata-naming": 0.0016090000281110406,
+      "mlir::detail::OpToOpPassAdaptor": 0.00021300000662449747,
+      "mlir::hlo::MhloToPyPenguin": 0.05411199852824211,
+      "mlir::mhlo::LowerComplexExtraPass": 0.00292299990542233,
+      "mlir::mhlo::LowerComplexPass": 0.0035620001144707203,
+      "native-to-custom-softmax": 0.00039999998989515007,
+      "native-to-custom-softmax-dx": 0.00041599999531172216,
+      "operand_upcaster": 0.0007249999907799065,
+      "post-par-pipe-begin": 9.999999974752427e-07,
+      "post-par-pipe-end": 0.0,
+      "post-partition-simplification": 0.10395599901676178,
+      "pre-hlo-begin": 3.000000106112566e-06,
+      "pre-hlo-end": 9.999999974752427e-07,
+      "replace-minimum-constant": 0.0002280000044265762,
+      "reshape-mover": 9.699999645818025e-05,
+      "simplify-concat": 0.002274000085890293,
+      "simplify-while-loops": 9.200000204145908e-05,
+      "transform-variadic-reduce": 0.0007149999728426337,
+      "tuple-simplifier": 0.0002570000069681555,
+      "unpack-nested-aws-ntwsr": 0.0003380000125616789,
+      "unroll-while-loop": 1.9999999494757503e-05
+    },
+    "hilo": {
+      "HloMacCount": 14764212224.0,
+      "Traffic": 7787352576.0
+    },
+    "tensorizer": {
+      "DMATilingProfiler::TotalInstructionsAfterTiling": 254771,
+      "StaticProfiler::AifUb": 5.7534074783325195,
+      "StaticProfiler::ArithmeticIntensityTensorizer": 6.476686000823975,
+      "StaticProfiler::AverageDmaLength": 6606.11767578125,
+      "StaticProfiler::DDRTransferBytes": 7394656764,
+      "StaticProfiler::InternalTransferBytes": 153915440,
+      "StaticProfiler::LoadExpanded": 1054235,
+      "StaticProfiler::StoreExpanded": 2218,
+      "StaticProfiler::TotalDMAExpanded": 1056453,
+      "StaticProfiler::TotalDynamicInstancesCount": 262628,
+      "StaticProfiler::TotalDynamicInstancesWithMmPackedCount": 257749,
+      "StaticProfiler::TotalLNCComm": 0,
+      "StaticProfiler::TotalLNCCommTransfer": 0,
+      "TilingProfiler::BatchnormInstructionsAfterTiling": 0,
+      "TilingProfiler::DmaInstructionsAfterTiling": 0,
+      "TilingProfiler::GenericInstructionsAfterTiling": 281,
+      "TilingProfiler::MatMultInstructionsAfterTiling": 226304,
+      "TilingProfiler::NumPfTransposes": 330,
+      "TilingProfiler::NumPfTransposesForIo": 33,
+      "TilingProfiler::NumPfTransposesForLocal": 162,
+      "TilingProfiler::NumPfTransposesForNonlocal": 135,
+      "TilingProfiler::PfTransposeInstructions": 9400,
+      "TilingProfiler::PfTransposeInstructionsForIo": 4104,
+      "TilingProfiler::PfTransposeInstructionsForLocal": 1060,
+      "TilingProfiler::PfTransposeInstructionsForNonlocal": 4236,
+      "TilingProfiler::ReduceInstructionsAfterTiling": 323,
+      "TilingProfiler::SimdInstructionsAfterTiling": 7045,
+      "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": 0.0,
+      "CanonicalizeForTensorizer": 0.0006220000213943422,
+      "Canonicalizer": 0.006990000139921904,
+      "HoistCompute": 0.00010399999882793054,
+      "IdentifyCrossPassTensors": 0.0005099999834783375,
+      "MemcastMotion": 0.00015700000221841037,
+      "PenguinizeFunctions": 0.00023700000019744039,
+      "PruneFunctions": 0.00019799999427050352,
+      "RemoveOptimizationBarriers": 0.0005300000193528831,
+      "ScatterMotion": 0.007275999989360571,
+      "TensorizerLegalizationPass": 0.0002500000118743628,
+      "VerifySupportedOps": 0.00025599999935366213,
+      "algsimp": 0.0017810000572353601,
+      "batchnorm_expander": 0.0009689999860711396,
+      "boundary-marker-removal": 0.00036299999919719994,
+      "call-inliner": 0.0004079999926034361,
+      "canonicalize-boundary-marker": 0.0004490000137593597,
+      "collective-stream-id-checker": 0.00010399999882793054,
+      "comparison-expander": 0.0004090000002179295,
+      "computation-deduplicator": 0.0005169999785721302,
+      "conditional-to-select": 0.00012099999730708078,
+      "config-lowering": 0.0003060000017285347,
+      "constant_folding": 0.00023099999816622585,
+      "cse": 0.0005119999987073243,
+      "dce": 6.900000153109431e-05,
+      "dynamic-slice-transpose": 0.00020799999765586108,
+      "eliminate-redundant-compare": 0.00024300000222865492,
+      "emit-offloaded-dropout": 0.0003650000144261867,
+      "flatten-call-graph": 0.0003650000144261867,
+      "fuse-send-recv": 0.002230999991297722,
+      "hilo::LegalizeAlias": 0.0036810000892728567,
+      "hilo::NeuronInstCombine": 0.001101000001654029,
+      "hilo::NeuronOpFusion": 0.00039599998854100704,
+      "hilo::ReplaceTokenTypeWithU8Pass": 0.0004830000107176602,
+      "hilo::ScheduleFusion": 5.900000178371556e-05,
+      "hilo::SixtyFourHack": 0.00026000000070780516,
+      "hilo::VerifyAliasing": 0.00015300000086426735,
+      "hlo-mac-count": 0.0007660000119358301,
+      "hlo-verifier": 0.00834800023585558,
+      "io-con-pipe-begin": 7.999999979801942e-06,
+      "io-con-pipe-end": 9.999999974752427e-07,
+      "io-layout-normalization": 0.0012280000373721123,
+      "legalize-ccops": 2.5999999706982635e-05,
+      "legalize-compare": 0.00031900001340545714,
+      "lower-argminmax-custom-call": 0.00019999999494757503,
+      "map-inline": 0.000674000009894371,
+      "metadata-naming": 0.0016090000281110406,
+      "mlir::detail::OpToOpPassAdaptor": 0.00021300000662449747,
+      "mlir::hlo::MhloToPyPenguin": 0.05411199852824211,
+      "mlir::mhlo::LowerComplexExtraPass": 0.00292299990542233,
+      "mlir::mhlo::LowerComplexPass": 0.0035620001144707203,
+      "native-to-custom-softmax": 0.00039999998989515007,
+      "native-to-custom-softmax-dx": 0.00041599999531172216,
+      "operand_upcaster": 0.0007249999907799065,
+      "post-par-pipe-begin": 9.999999974752427e-07,
+      "post-par-pipe-end": 0.0,
+      "post-partition-simplification": 0.10395599901676178,
+      "pre-hlo-begin": 3.000000106112566e-06,
+      "pre-hlo-end": 9.999999974752427e-07,
+      "replace-minimum-constant": 0.0002280000044265762,
+      "reshape-mover": 9.699999645818025e-05,
+      "simplify-concat": 0.002274000085890293,
+      "simplify-while-loops": 9.200000204145908e-05,
+      "transform-variadic-reduce": 0.0007149999728426337,
+      "tuple-simplifier": 0.0002570000069681555,
+      "unpack-nested-aws-ntwsr": 0.0003380000125616789,
+      "unroll-while-loop": 1.9999999494757503e-05
+    }
+  },
+  "cumsum": {
+    "compiletime": {
+      "CoalesceCCOp": 0.0002467632293701172,
+      "DMALocalityOpt": 0.0002090930938720703,
+      "DMAProfiler": 0.0008313655853271484,
+      "DataStreaming": 0.0002913475036621094,
+      "DoNothing": 0.00015354156494140625,
+      "ExpandISAMacro": 0.0005238056182861328,
+      "FactorizeBlkDims": 0.00044465065002441406,
+      "InferPSumTensor": 0.00044846534729003906,
+      "LateLegalizeInst": 0.00041985511779785156,
+      "LateNeuronInstComb": 0.0004794597625732422,
+      "LegalizeSundaAccess": 0.0014634132385253906,
+      "LegalizeType": 0.00027251243591308594,
+      "LowerBroadcast": 0.00026226043701171875,
+      "LowerIntrinsics": 0.0002529621124267578,
+      "LowerTranspose": 0.0002646446228027344,
+      "NeuronInstComb": 0.0004792213439941406,
+      "NeuronLICM": 0.0004215240478515625,
+      "NeuronSimplifyPredicates": 0.002805471420288086,
+      "NeuronValueNumbering": 0.0004172325134277344,
+      "SFKVectorizer": 0.002515554428100586,
+      "SimpleAllReduceTiling": 0.00023794174194335938,
+      "SimplifyNeuronTensor": 0.0004241466522216797,
+      "SpillPSum": 0.0005099773406982422,
+      "WeightCoalescing": 0.00024390220642089844
+    }
+  },
+  "sg00": {
+    "hilo": {
+      "ArithmeticIntensity": 3.7918436527252197,
+      "HloMacCount": 14764212224.0,
+      "Traffic": 7787352576.0
+    }
+  },
+  "sg0000": {
+    "compiletime": {
+      "AGOrderingAnalysisPass": 2.305126667022705,
+      "AffinePredicateResolution": 0.09757089614868164,
+      "AliasDependencyElimination": 0.0027077198028564453,
+      "AliasDependencyInduction": 0.6443400382995605,
+      "AliasDependencyReset": 0.6698822975158691,
+      "BFComputeCutting": 0.07641983032226563,
+      "BirCodeGenLoop": 2.047102689743042,
+      "CCOpFusion": 0.708749532699585,
+      "CanonicalizeDAGForPGTiling": 0.18178820610046387,
+      "CanonicalizeIR": 0.11043667793273926,
+      "CoalesceCCOp": 0.16056489944458008,
+      "CommuteConcat": 0.036214590072631836,
+      "DMALocalityOpt": 0.02845001220703125,
+      "DMAProfiler": 0.07055497169494629,
+      "DMATilingProfiler": 0.06614851951599121,
+      "DataLocalityOpt": 2.321251392364502,
+      "DataStreaming": 0.10943412780761719,
+      "DeConcat": 0.02409839630126953,
+      "DeadCodeElimination": 0.035909414291381836,
+      "DeadStoreElimination": 1.309647560119629,
+      "DelinearIndices": 0.3166069984436035,
+      "Delinearization": 0.12859582901000977,
+      "DoNothing": 0.000125885009765625,
+      "DramToDramTranspose": 1.095766544342041,
+      "DumpGraphAndMetadata": 0.16588640213012695,
+      "EliminateDivs": 0.2634701728820801,
+      "ExpandBatchNorm": 0.09694433212280273,
+      "ExpandISAMacro": 0.06739115715026855,
+      "FactorizeBlkDims": 0.39147496223449707,
+      "FactorizeThreadAxesInFreeDims": 0.04469728469848633,
+      "FlattenMacroLoop": 0.07828259468078613,
+      "GenericAccessSimplifier": 0.03183865547180176,
+      "InferInitValue": 1.2782227993011475,
+      "InferIntrinsicOnCC": 0.35162997245788574,
+      "InferNeuronTensor": 1.7735645771026611,
+      "InferNonlocalTensors": 4.640493392944336,
+      "InferPSumTensor": 0.9276552200317383,
+      "InlineNativeKernels": 0.04500627517700195,
+      "InsertIOTransposes": 0.8502795696258545,
+      "InsertLocalTransposes": 0.7141804695129395,
+      "InsertOffloadedTransposes": 0.06985926628112793,
+      "LICM": 0.09515905380249023,
+      "LateLegalizeInst": 0.1378779411315918,
+      "LateLegalizePostSplit": 0.07408785820007324,
+      "LateLowerReshapeOp": 0.03981423377990723,
+      "LateLowerTensorOp": 0.9939045906066895,
+      "LateNeuronInstComb": 0.39553403854370117,
+      "LayoutPreprocessing": 0.9070339202880859,
+      "LayoutPreprocessingAndAnalysis": 1.2123572826385498,
+      "LayoutRequirementAnalysis": 0.2940239906311035,
+      "LegalizeCCOpLayout": 0.1556408405303955,
+      "LegalizeOpLevelAlias": 0.03197050094604492,
+      "LegalizePartitionReduce": 0.06994032859802246,
+      "LegalizeSundaAccess": 1.016023874282837,
+      "LegalizeSundaMacro": 0.344529390335083,
+      "LegalizeType": 0.1501622200012207,
+      "LocalLayoutOpt": 0.7021169662475586,
+      "LoopFusion": 0.3479132652282715,
+      "LoopSplitting": 0.031319618225097656,
+      "LowerBroadcast": 0.08351755142211914,
+      "LowerCCOpBlockAxis": 0.21303391456604004,
+      "LowerComplexBroadcast": 0.13837647438049316,
+      "LowerIntrinsics": 0.8478243350982666,
+      "LowerTensorOp": 0.6734035015106201,
+      "LowerTranspose": 0.4167468547821045,
+      "MacroGeneration": 2.3995394706726074,
+      "MaskPropagation": 0.11813521385192871,
+      "MemcpyElimination": 9.78618049621582,
+      "MutateDataType": 0.04571127891540527,
+      "NeuronAliasDependencyInduction": 0.018005847930908203,
+      "NeuronAliasDependencyReset": 0.033030033111572266,
+      "NeuronInstComb": 0.31482362747192383,
+      "NeuronLICM": 0.26255059242248535,
+      "NeuronLoopFusion": 0.7358701229095459,
+      "NeuronLoopInterchange": 0.04501771926879883,
+      "NeuronSimplifier": 0.37318873405456543,
+      "NeuronSimplifyPredicates": 0.05665278434753418,
+      "NeuronValueNumbering": 0.08318281173706055,
+      "OptimizeAliasedCopyChain": 0.016149282455444336,
+      "OptimizeNKIKernels": 0.36123156547546387,
+      "PAGLayoutOpt": 43.12562561035156,
+      "PComputeCutting": 0.31317949295043945,
+      "PGLayoutTilingPipeline": 57.15171813964844,
+      "PGTiling": 5.401483535766602,
+      "PadElimination": 0.01123952865600586,
+      "ParAxesAnnotation": 42.39716339111328,
+      "PartialLoopFusion": 0.4405984878540039,
+      "PartialSimdFusion": 0.33167028427124023,
+      "PerfectLoopNest": 0.057257890701293945,
+      "RecognizeOpIdiom": 0.14631319046020508,
+      "Recompute": 0.008225679397583008,
+      "RelaxPredicates": 0.11752939224243164,
+      "Rematerialization": 0.1648578643798828,
+      "ReshapeWeights": 0.02127981185913086,
+      "ResolveAccessConflict": 0.24170351028442383,
+      "ResolveComplicatePredicates": 0.05225372314453125,
+      "RewriteReplicationMatmul": 0.042751312255859375,
+      "RewriteWeights": 0.06114816665649414,
+      "SFKVectorizer": 6.082370758056641,
+      "SimpleAllReduceTiling": 0.05324435234069824,
+      "Simplifier": 0.11982178688049316,
+      "SimplifyMacroPredicates": 0.18191838264465332,
+      "SimplifyNeuronTensor": 0.4066755771636963,
+      "SimplifySlice": 0.07960724830627441,
+      "SimplifyTensor": 0.19757962226867676,
+      "SpillPSum": 0.5764906406402588,
+      "SplitAPUnionSets": 0.2315382957458496,
+      "SplitAccGrp": 0.03763127326965332,
+      "StaticProfiler": 0.11375260353088379,
+      "StaticTransposeLocalTensor": 0.2317829132080078,
+      "SundaISel": 1.3966825008392334,
+      "TCTransform": 0.039675235748291016,
+      "TensorInitialization": 0.13313651084899902,
+      "TensorOpSimplifier": 0.5821249485015869,
+      "TensorOpTransform": 2.0534775257110596,
+      "TileCCOps": 0.2945268154144287,
+      "TilingProfiler": 0.41756296157836914,
+      "TransformConvOp": 0.069244384765625,
+      "TritiumFusion": 1.3019096851348877,
+      "ValueNumbering": 0.09937286376953125,
+      "VectorizeDMA": 0.11787676811218262,
+      "VectorizeMatMult": 0.0218658447265625,
+      "WeightCoalescing": 0.046035051345825195,
+      "ZeroSizeTensorElimination": 0.00039458274841308594
+    },
+    "tensorizer": {
+      "DMATilingProfiler::TotalInstructionsAfterTiling": 254771,
+      "StaticProfiler::AifUb": 5.7534074783325195,
+      "StaticProfiler::ArithmeticIntensityTensorizer": 6.476686000823975,
+      "StaticProfiler::AverageDmaLength": 6606.11767578125,
+      "StaticProfiler::AverageFractalPeUtilization": 99.4054183959961,
+      "StaticProfiler::AveragePartitionUtilization": 98.58908081054688,
+      "StaticProfiler::AveragePeUtilization": 97.55974578857422,
+      "StaticProfiler::DDRTransferBytes": 7394656764,
+      "StaticProfiler::InternalTransferBytes": 153915440,
+      "StaticProfiler::LoadExpanded": 1054235,
+      "StaticProfiler::LocalizationEfficiency": 112.5713119506836,
+      "StaticProfiler::LocalizationEfficiencyIgnoreNonlocal": 112.71009826660156,
+      "StaticProfiler::StoreExpanded": 2218,
+      "StaticProfiler::TotalDMAExpanded": 1056453,
+      "StaticProfiler::TotalDynamicInstancesCount": 262628,
+      "StaticProfiler::TotalDynamicInstancesWithMmPackedCount": 257749,
+      "StaticProfiler::TotalLNCComm": 0,
+      "StaticProfiler::TotalLNCCommTransfer": 0,
+      "TilingProfiler::AveragePartitionUtilizationAfterTiling": 0,
+      "TilingProfiler::AveragePeUtilizationAfterTiling": 0,
+      "TilingProfiler::BatchnormInstructionsAfterTiling": 0,
+      "TilingProfiler::DmaInstructionsAfterTiling": 0,
+      "TilingProfiler::GenericInstructionsAfterTiling": 281,
+      "TilingProfiler::MatMultInstructionsAfterTiling": 226304,
+      "TilingProfiler::NumPfTransposes": 330,
+      "TilingProfiler::NumPfTransposesForIo": 33,
+      "TilingProfiler::NumPfTransposesForLocal": 162,
+      "TilingProfiler::NumPfTransposesForNonlocal": 135,
+      "TilingProfiler::PfTransposeInstructions": 9400,
+      "TilingProfiler::PfTransposeInstructionsForIo": 4104,
+      "TilingProfiler::PfTransposeInstructionsForLocal": 1060,
+      "TilingProfiler::PfTransposeInstructionsForNonlocal": 4236,
+      "TilingProfiler::ReduceInstructionsAfterTiling": 323,
+      "TilingProfiler::SimdInstructionsAfterTiling": 7045,
+      "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_bk3/neuron_config.json

@@ -0,0 +1,213 @@
+{
+  "_attn_implementation_autoset": false,
+  "_name_or_path": "mistralai/Mistral-7B-Instruct-v0.3",
+  "add_cross_attention": false,
+  "architectures": [
+    "MistralForCausalLM"
+  ],
+  "attention_dropout": 0.0,
+  "attribute_map": {},
+  "bad_words_ids": null,
+  "begin_suppress_tokens": null,
+  "bos_token_id": 1,
+  "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": 2,
+  "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": 14336,
+  "is_decoder": false,
+  "is_encoder_decoder": false,
+  "label2id": {
+    "LABEL_0": 0,
+    "LABEL_1": 1
+  },
+  "length_penalty": 1.0,
+  "max_length": 20,
+  "max_position_embeddings": 32768,
+  "metadata": null,
+  "min_length": 0,
+  "model_type": "mistral",
+  "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": "NeuronLlamaAttention",
+    "attn_kernel_enabled": null,
+    "attn_tkg_builtin_kernel_enabled": false,
+    "attn_tkg_nki_kernel_enabled": false,
+    "batch_size": 4,
+    "bucket_n_active_tokens": false,
+    "buckets": [
+      1024
+    ],
+    "cast_type": "config",
+    "cc_pipeline_tiling_factor": 1,
+    "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": false,
+    "is_prefix_caching": false,
+    "k_cache_transposed": false,
+    "kv_cache_batch_size": 4,
+    "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": 4,
+    "max_context_length": 2048,
+    "max_length": 2048,
+    "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": 1,
+    "n_positions": 2048,
+    "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": 2048,
+    "pa_num_blocks": 4,
+    "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": 2048,
+    "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": 4,
+    "speculation_length": 0,
+    "start_rank_id": 0,
+    "target": null,
+    "tile_cc": false,
+    "tkg_batch_size": 4,
+    "token_generation_buckets": [
+      1024
+    ],
+    "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": 32,
+  "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-05,
+  "rope_theta": 1000000.0,
+  "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": "4.42.0.dev0",
+  "typical_p": 1.0,
+  "use_bfloat16": false,
+  "use_cache": true,
+  "vocab_size": 32768
+}

token_generation_model/_tp0_bk4/command.txt

@@ -0,0 +1 @@
+neuronx-cc compile --framework=XLA model.MODULE_fb6decaa94b1936d08da+1b5847e3.hlo_module.pb --output model.MODULE_fb6decaa94b1936d08da+1b5847e3.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 --verbose=35

token_generation_model/_tp0_bk4/compile_flags.MODULE_fb6decaa94b1936d08da+1b5847e3.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=/models/mistral-7b-v0.3-instruct-neuronx/token_generation_model/_tp0_bk4/log-neuron-cc.txt"]

token_generation_model/_tp0_bk4/global_metric_store.json

@@ -0,0 +1,540 @@
+{
+  "Average": {
+    "tensorizer": {
+      "StaticProfiler::AverageFractalPeUtilization": 99.4205093383789,
+      "StaticProfiler::AveragePartitionUtilization": 98.67665100097656,
+      "StaticProfiler::AveragePeUtilization": 97.56974029541016,
+      "StaticProfiler::LocalizationEfficiency": 108.62796020507813,
+      "StaticProfiler::LocalizationEfficiencyIgnoreNonlocal": 108.75718688964844,
+      "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": 2.275869369506836,
+      "AffinePredicateResolution": 0.10508346557617188,
+      "AliasDependencyElimination": 0.0029473304748535156,
+      "AliasDependencyInduction": 1.0161068439483643,
+      "AliasDependencyReset": 1.0964906215667725,
+      "BFComputeCutting": 0.3273308277130127,
+      "BirCodeGenLoop": 1.817042350769043,
+      "CCOpFusion": 0.759657621383667,
+      "CanonicalizeConv": 1.8000000636675395e-05,
+      "CanonicalizeDAGForPGTiling": 0.18803882598876953,
+      "CanonicalizeForTensorizer": 0.0005799999926239252,
+      "CanonicalizeIR": 0.11030387878417969,
+      "Canonicalizer": 0.007720999885350466,
+      "CoalesceCCOp": 0.17146921157836914,
+      "CommuteConcat": 0.03537917137145996,
+      "DMALocalityOpt": 0.03381848335266113,
+      "DMAProfiler": 0.08129215240478516,
+      "DMATilingProfiler": 0.07448625564575195,
+      "DataLocalityOpt": 2.595104932785034,
+      "DataStreaming": 0.12432718276977539,
+      "DeConcat": 0.026020050048828125,
+      "DeadCodeElimination": 0.03551626205444336,
+      "DeadStoreElimination": 1.3150177001953125,
+      "DelinearIndices": 0.3447244167327881,
+      "Delinearization": 0.1470789909362793,
+      "DoNothing": 0.00028705596923828125,
+      "DramToDramTranspose": 1.1585099697113037,
+      "DumpGraphAndMetadata": 0.16599822044372559,
+      "EliminateDivs": 0.16966032981872559,
+      "ExpandBatchNorm": 0.08444547653198242,
+      "ExpandISAMacro": 0.07590079307556152,
+      "FactorizeBlkDims": 0.41552281379699707,
+      "FactorizeThreadAxesInFreeDims": 0.05351138114929199,
+      "FlattenMacroLoop": 0.08939695358276367,
+      "GenericAccessSimplifier": 0.03192710876464844,
+      "HoistCompute": 9.600000339560211e-05,
+      "IdentifyCrossPassTensors": 0.00026000000070780516,
+      "InferInitValue": 1.1084926128387451,
+      "InferIntrinsicOnCC": 0.36765527725219727,
+      "InferNeuronTensor": 1.879042387008667,
+      "InferNonlocalTensors": 4.959750175476074,
+      "InferPSumTensor": 0.9908087253570557,
+      "InlineNativeKernels": 0.04849410057067871,
+      "InsertIOTransposes": 0.9103872776031494,
+      "InsertLocalTransposes": 0.7676966190338135,
+      "InsertOffloadedTransposes": 0.0744314193725586,
+      "LICM": 0.10265731811523438,
+      "LateLegalizeInst": 0.1494448184967041,
+      "LateLegalizePostSplit": 0.08162188529968262,
+      "LateLowerReshapeOp": 0.043623924255371094,
+      "LateLowerTensorOp": 0.7195582389831543,
+      "LateNeuronInstComb": 0.4356348514556885,
+      "LayoutPreprocessing": 1.0318124294281006,
+      "LayoutPreprocessingAndAnalysis": 1.369814157485962,
+      "LayoutRequirementAnalysis": 0.32599711418151855,
+      "LegalizeCCOpLayout": 0.11543631553649902,
+      "LegalizeOpLevelAlias": 0.042886972427368164,
+      "LegalizePartitionReduce": 0.07607841491699219,
+      "LegalizeSundaAccess": 1.0659985542297363,
+      "LegalizeSundaMacro": 0.3716259002685547,
+      "LegalizeType": 0.17189598083496094,
+      "LocalLayoutOpt": 0.7340552806854248,
+      "LoopFusion": 0.35437893867492676,
+      "LoopSplitting": 0.03284788131713867,
+      "LowerBroadcast": 0.0972299575805664,
+      "LowerCCOpBlockAxis": 0.4246203899383545,
+      "LowerComplexBroadcast": 0.1469728946685791,
+      "LowerIntrinsics": 0.607001543045044,
+      "LowerTensorOp": 0.625375509262085,
+      "LowerTranspose": 0.4381232261657715,
+      "MacroGeneration": 2.3300650119781494,
+      "MaskPropagation": 0.13726162910461426,
+      "MemcastMotion": 0.00020199999562464654,
+      "MemcpyElimination": 9.993162155151367,
+      "MutateDataType": 0.04567551612854004,
+      "NeuronAliasDependencyInduction": 0.018932819366455078,
+      "NeuronAliasDependencyReset": 0.03390645980834961,
+      "NeuronInstComb": 0.35080647468566895,
+      "NeuronLICM": 0.26523780822753906,
+      "NeuronLoopFusion": 0.7789182662963867,
+      "NeuronLoopInterchange": 0.04993152618408203,
+      "NeuronSimplifier": 0.3916890621185303,
+      "NeuronSimplifyPredicates": 0.06646299362182617,
+      "NeuronValueNumbering": 0.09050369262695313,
+      "OptimizeAliasedCopyChain": 0.024803638458251953,
+      "OptimizeNKIKernels": 0.3809840679168701,
+      "PAGLayoutOpt": 44.665672302246094,
+      "PComputeCutting": 0.327991247177124,
+      "PGLayoutTilingPipeline": 59.74672317504883,
+      "PGTiling": 5.59335994720459,
+      "PadElimination": 0.014229774475097656,
+      "ParAxesAnnotation": 43.884578704833984,
+      "PartialLoopFusion": 0.4160308837890625,
+      "PartialSimdFusion": 0.35861706733703613,
+      "PenguinizeFunctions": 0.00023999999393709004,
+      "PerfectLoopNest": 0.06206011772155762,
+      "PruneFunctions": 0.00014600000577047467,
+      "RecognizeOpIdiom": 0.14852476119995117,
+      "Recompute": 0.00867772102355957,
+      "RelaxPredicates": 0.12411117553710938,
+      "Rematerialization": 0.15559935569763184,
+      "RemoveOptimizationBarriers": 0.00010399999882793054,
+      "ReshapeWeights": 0.0252685546875,
+      "ResolveAccessConflict": 0.23229002952575684,
+      "ResolveComplicatePredicates": 0.08260774612426758,
+      "RewriteReplicationMatmul": 0.04953479766845703,
+      "RewriteWeights": 0.06951451301574707,
+      "SFKVectorizer": 5.916876792907715,
+      "ScatterMotion": 0.009108999744057655,
+      "SimpleAllReduceTiling": 0.05846905708312988,
+      "Simplifier": 0.12264633178710938,
+      "SimplifyMacroPredicates": 0.19836044311523438,
+      "SimplifyNeuronTensor": 0.7509462833404541,
+      "SimplifySlice": 0.033946990966796875,
+      "SimplifyTensor": 0.20671844482421875,
+      "SpillPSum": 0.6513323783874512,
+      "SplitAPUnionSets": 0.25449252128601074,
+      "SplitAccGrp": 0.04361081123352051,
+      "StaticProfiler": 0.44228219985961914,
+      "StaticTransposeLocalTensor": 0.25648951530456543,
+      "SundaISel": 1.4593234062194824,
+      "TCTransform": 0.0390927791595459,
+      "TensorInitialization": 0.1486976146697998,
+      "TensorOpSimplifier": 0.5177226066589355,
+      "TensorOpTransform": 1.8292319774627686,
+      "TensorizerLegalizationPass": 0.00022499999613501132,
+      "TileCCOps": 0.2758169174194336,
+      "TilingProfiler": 0.45922422409057617,
+      "TransformConvOp": 0.05972599983215332,
+      "TritiumFusion": 1.8706681728363037,
+      "ValueNumbering": 0.09980320930480957,
+      "VectorizeDMA": 0.12941217422485352,
+      "VectorizeMatMult": 0.02786874771118164,
+      "VerifySupportedOps": 0.0002690000110305846,
+      "WeightCoalescing": 0.04975485801696777,
+      "ZeroSizeTensorElimination": 0.00039505958557128906,
+      "algsimp": 0.0017940000398084521,
+      "batchnorm_expander": 0.0008820000221021473,
+      "boundary-marker-removal": 0.00037900000461377203,
+      "call-inliner": 0.0003760000108741224,
+      "canonicalize-boundary-marker": 0.0004459999909158796,
+      "collective-stream-id-checker": 8.099999831756577e-05,
+      "comparison-expander": 0.00043700000969693065,
+      "computation-deduplicator": 0.0005830000154674053,
+      "conditional-to-select": 0.00013899999612476677,
+      "config-lowering": 0.0003220000071451068,
+      "constant_folding": 0.00022699999681208283,
+      "cse": 0.0005879999953322113,
+      "dce": 0.00010299999848939478,
+      "dynamic-slice-transpose": 0.00020300000323913991,
+      "eliminate-redundant-compare": 0.00022000000171829015,
+      "emit-offloaded-dropout": 0.000371000001905486,
+      "flatten-call-graph": 0.0004889999981969595,
+      "fuse-send-recv": 0.0023360000923275948,
+      "hilo::LegalizeAlias": 0.003705000039190054,
+      "hilo::NeuronInstCombine": 0.0010809999657794833,
+      "hilo::NeuronOpFusion": 0.0005680000176653266,
+      "hilo::ReplaceTokenTypeWithU8Pass": 0.0003440000000409782,
+      "hilo::ScheduleFusion": 6.199999916134402e-05,
+      "hilo::SixtyFourHack": 0.00025599999935366213,
+      "hilo::VerifyAliasing": 0.00011899999663000926,
+      "hlo-mac-count": 0.0007440000190399587,
+      "hlo-verifier": 0.008407999761402607,
+      "io-con-pipe-begin": 7.000000096013537e-06,
+      "io-con-pipe-end": 9.999999974752427e-07,
+      "io-layout-normalization": 0.0013689999468624592,
+      "legalize-ccops": 2.4000000848900527e-05,
+      "legalize-compare": 0.0004239999980200082,
+      "lower-argminmax-custom-call": 0.00019700000120792538,
+      "map-inline": 0.0007559999939985573,
+      "metadata-naming": 0.0019259999971836805,
+      "mlir::detail::OpToOpPassAdaptor": 0.00028199999360367656,
+      "mlir::hlo::MhloToPyPenguin": 0.058632999658584595,
+      "mlir::mhlo::LowerComplexExtraPass": 0.0029980000108480453,
+      "mlir::mhlo::LowerComplexPass": 0.0030219999607652426,
+      "native-to-custom-softmax": 0.00038400001358240843,
+      "native-to-custom-softmax-dx": 0.0004239999980200082,
+      "operand_upcaster": 0.0008040000102482736,
+      "post-par-pipe-begin": 9.999999974752427e-07,
+      "post-par-pipe-end": 0.0,
+      "post-partition-simplification": 0.10732399672269821,
+      "pre-hlo-begin": 3.000000106112566e-06,
+      "pre-hlo-end": 9.999999974752427e-07,
+      "replace-minimum-constant": 0.0003169999981764704,
+      "reshape-mover": 9.300000237999484e-05,
+      "simplify-concat": 0.0022960000205785036,
+      "simplify-while-loops": 7.899999764049426e-05,
+      "transform-variadic-reduce": 0.000707999977748841,
+      "tuple-simplifier": 0.0002300000051036477,
+      "unpack-nested-aws-ntwsr": 0.000371000001905486,
+      "unroll-while-loop": 1.8000000636675395e-05
+    },
+    "hilo": {
+      "HloMacCount": 15301083136.0,
+      "Traffic": 7787368960.0
+    },
+    "tensorizer": {
+      "DMATilingProfiler::TotalInstructionsAfterTiling": 272515,
+      "StaticProfiler::AifUb": 8.01248550415039,
+      "StaticProfiler::ArithmeticIntensityTensorizer": 8.703800201416016,
+      "StaticProfiler::AverageDmaLength": 5833.71923828125,
+      "StaticProfiler::DDRTransferBytes": 7663108604,
+      "StaticProfiler::InternalTransferBytes": 296554544,
+      "StaticProfiler::LoadExpanded": 1250875,
+      "StaticProfiler::StoreExpanded": 2218,
+      "StaticProfiler::TotalDMAExpanded": 1253093,
+      "StaticProfiler::TotalDynamicInstancesCount": 278788,
+      "StaticProfiler::TotalDynamicInstancesWithMmPackedCount": 273565,
+      "StaticProfiler::TotalLNCComm": 0,
+      "StaticProfiler::TotalLNCCommTransfer": 0,
+      "TilingProfiler::BatchnormInstructionsAfterTiling": 0,
+      "TilingProfiler::DmaInstructionsAfterTiling": 0,
+      "TilingProfiler::GenericInstructionsAfterTiling": 281,
+      "TilingProfiler::MatMultInstructionsAfterTiling": 234496,
+      "TilingProfiler::NumPfTransposes": 330,
+      "TilingProfiler::NumPfTransposesForIo": 33,
+      "TilingProfiler::NumPfTransposesForLocal": 162,
+      "TilingProfiler::NumPfTransposesForNonlocal": 135,
+      "TilingProfiler::PfTransposeInstructions": 13760,
+      "TilingProfiler::PfTransposeInstructionsForIo": 8208,
+      "TilingProfiler::PfTransposeInstructionsForLocal": 1316,
+      "TilingProfiler::PfTransposeInstructionsForNonlocal": 4236,
+      "TilingProfiler::ReduceInstructionsAfterTiling": 643,
+      "TilingProfiler::SimdInstructionsAfterTiling": 7821,
+      "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": 1.8000000636675395e-05,
+      "CanonicalizeForTensorizer": 0.0005799999926239252,
+      "Canonicalizer": 0.007720999885350466,
+      "HoistCompute": 9.600000339560211e-05,
+      "IdentifyCrossPassTensors": 0.00026000000070780516,
+      "MemcastMotion": 0.00020199999562464654,
+      "PenguinizeFunctions": 0.00023999999393709004,
+      "PruneFunctions": 0.00014600000577047467,
+      "RemoveOptimizationBarriers": 0.00010399999882793054,
+      "ScatterMotion": 0.009108999744057655,
+      "TensorizerLegalizationPass": 0.00022499999613501132,
+      "VerifySupportedOps": 0.0002690000110305846,
+      "algsimp": 0.0017940000398084521,
+      "batchnorm_expander": 0.0008820000221021473,
+      "boundary-marker-removal": 0.00037900000461377203,
+      "call-inliner": 0.0003760000108741224,
+      "canonicalize-boundary-marker": 0.0004459999909158796,
+      "collective-stream-id-checker": 8.099999831756577e-05,
+      "comparison-expander": 0.00043700000969693065,
+      "computation-deduplicator": 0.0005830000154674053,
+      "conditional-to-select": 0.00013899999612476677,
+      "config-lowering": 0.0003220000071451068,
+      "constant_folding": 0.00022699999681208283,
+      "cse": 0.0005879999953322113,
+      "dce": 0.00010299999848939478,
+      "dynamic-slice-transpose": 0.00020300000323913991,
+      "eliminate-redundant-compare": 0.00022000000171829015,
+      "emit-offloaded-dropout": 0.000371000001905486,
+      "flatten-call-graph": 0.0004889999981969595,
+      "fuse-send-recv": 0.0023360000923275948,
+      "hilo::LegalizeAlias": 0.003705000039190054,
+      "hilo::NeuronInstCombine": 0.0010809999657794833,
+      "hilo::NeuronOpFusion": 0.0005680000176653266,
+      "hilo::ReplaceTokenTypeWithU8Pass": 0.0003440000000409782,
+      "hilo::ScheduleFusion": 6.199999916134402e-05,
+      "hilo::SixtyFourHack": 0.00025599999935366213,
+      "hilo::VerifyAliasing": 0.00011899999663000926,
+      "hlo-mac-count": 0.0007440000190399587,
+      "hlo-verifier": 0.008407999761402607,
+      "io-con-pipe-begin": 7.000000096013537e-06,
+      "io-con-pipe-end": 9.999999974752427e-07,
+      "io-layout-normalization": 0.0013689999468624592,
+      "legalize-ccops": 2.4000000848900527e-05,
+      "legalize-compare": 0.0004239999980200082,
+      "lower-argminmax-custom-call": 0.00019700000120792538,
+      "map-inline": 0.0007559999939985573,
+      "metadata-naming": 0.0019259999971836805,
+      "mlir::detail::OpToOpPassAdaptor": 0.00028199999360367656,
+      "mlir::hlo::MhloToPyPenguin": 0.058632999658584595,
+      "mlir::mhlo::LowerComplexExtraPass": 0.0029980000108480453,
+      "mlir::mhlo::LowerComplexPass": 0.0030219999607652426,
+      "native-to-custom-softmax": 0.00038400001358240843,
+      "native-to-custom-softmax-dx": 0.0004239999980200082,
+      "operand_upcaster": 0.0008040000102482736,
+      "post-par-pipe-begin": 9.999999974752427e-07,
+      "post-par-pipe-end": 0.0,
+      "post-partition-simplification": 0.10732399672269821,
+      "pre-hlo-begin": 3.000000106112566e-06,
+      "pre-hlo-end": 9.999999974752427e-07,
+      "replace-minimum-constant": 0.0003169999981764704,
+      "reshape-mover": 9.300000237999484e-05,
+      "simplify-concat": 0.0022960000205785036,
+      "simplify-while-loops": 7.899999764049426e-05,
+      "transform-variadic-reduce": 0.000707999977748841,
+      "tuple-simplifier": 0.0002300000051036477,
+      "unpack-nested-aws-ntwsr": 0.000371000001905486,
+      "unroll-while-loop": 1.8000000636675395e-05
+    }
+  },
+  "cumsum": {
+    "compiletime": {
+      "CoalesceCCOp": 0.00024366378784179688,
+      "DMALocalityOpt": 0.00020456314086914063,
+      "DMAProfiler": 0.000926971435546875,
+      "DataStreaming": 0.00028705596923828125,
+      "DoNothing": 0.00016045570373535156,
+      "ExpandISAMacro": 0.0005159378051757813,
+      "FactorizeBlkDims": 0.0005528926849365234,
+      "InferPSumTensor": 0.0005333423614501953,
+      "LateLegalizeInst": 0.00042557716369628906,
+      "LateNeuronInstComb": 0.0005631446838378906,
+      "LegalizeSundaAccess": 0.0017147064208984375,
+      "LegalizeType": 0.0003027915954589844,
+      "LowerBroadcast": 0.00026917457580566406,
+      "LowerIntrinsics": 0.00027251243591308594,
+      "LowerTranspose": 0.0002696514129638672,
+      "NeuronInstComb": 0.0005166530609130859,
+      "NeuronLICM": 0.0004801750183105469,
+      "NeuronSimplifyPredicates": 0.002925872802734375,
+      "NeuronValueNumbering": 0.0004703998565673828,
+      "SFKVectorizer": 0.002753019332885742,
+      "SimpleAllReduceTiling": 0.00023436546325683594,
+      "SimplifyNeuronTensor": 0.0004436969757080078,
+      "SpillPSum": 0.0006051063537597656,
+      "WeightCoalescing": 0.0002589225769042969
+    }
+  },
+  "sg00": {
+    "hilo": {
+      "ArithmeticIntensity": 3.929718255996704,
+      "HloMacCount": 15301083136.0,
+      "Traffic": 7787368960.0
+    }
+  },
+  "sg0000": {
+    "compiletime": {
+      "AGOrderingAnalysisPass": 2.275869369506836,
+      "AffinePredicateResolution": 0.10508346557617188,
+      "AliasDependencyElimination": 0.0029473304748535156,
+      "AliasDependencyInduction": 1.0161068439483643,
+      "AliasDependencyReset": 1.0964906215667725,
+      "BFComputeCutting": 0.3273308277130127,
+      "BirCodeGenLoop": 1.817042350769043,
+      "CCOpFusion": 0.759657621383667,
+      "CanonicalizeDAGForPGTiling": 0.18803882598876953,
+      "CanonicalizeIR": 0.11030387878417969,
+      "CoalesceCCOp": 0.17122554779052734,
+      "CommuteConcat": 0.03537917137145996,
+      "DMALocalityOpt": 0.03361392021179199,
+      "DMAProfiler": 0.08036518096923828,
+      "DMATilingProfiler": 0.07448625564575195,
+      "DataLocalityOpt": 2.595104932785034,
+      "DataStreaming": 0.12404012680053711,
+      "DeConcat": 0.026020050048828125,
+      "DeadCodeElimination": 0.03551626205444336,
+      "DeadStoreElimination": 1.3150177001953125,
+      "DelinearIndices": 0.3447244167327881,
+      "Delinearization": 0.1470789909362793,
+      "DoNothing": 0.0001266002655029297,
+      "DramToDramTranspose": 1.1585099697113037,
+      "DumpGraphAndMetadata": 0.16599822044372559,
+      "EliminateDivs": 0.16966032981872559,
+      "ExpandBatchNorm": 0.08444547653198242,
+      "ExpandISAMacro": 0.07538485527038574,
+      "FactorizeBlkDims": 0.41496992111206055,
+      "FactorizeThreadAxesInFreeDims": 0.05351138114929199,
+      "FlattenMacroLoop": 0.08939695358276367,
+      "GenericAccessSimplifier": 0.03192710876464844,
+      "InferInitValue": 1.1084926128387451,
+      "InferIntrinsicOnCC": 0.36765527725219727,
+      "InferNeuronTensor": 1.879042387008667,
+      "InferNonlocalTensors": 4.959750175476074,
+      "InferPSumTensor": 0.9902753829956055,
+      "InlineNativeKernels": 0.04849410057067871,
+      "InsertIOTransposes": 0.9103872776031494,
+      "InsertLocalTransposes": 0.7676966190338135,
+      "InsertOffloadedTransposes": 0.0744314193725586,
+      "LICM": 0.10265731811523438,
+      "LateLegalizeInst": 0.1490192413330078,
+      "LateLegalizePostSplit": 0.08162188529968262,
+      "LateLowerReshapeOp": 0.043623924255371094,
+      "LateLowerTensorOp": 0.7195582389831543,
+      "LateNeuronInstComb": 0.4350717067718506,
+      "LayoutPreprocessing": 1.0318124294281006,
+      "LayoutPreprocessingAndAnalysis": 1.369814157485962,
+      "LayoutRequirementAnalysis": 0.32599711418151855,
+      "LegalizeCCOpLayout": 0.11543631553649902,
+      "LegalizeOpLevelAlias": 0.042886972427368164,
+      "LegalizePartitionReduce": 0.07607841491699219,
+      "LegalizeSundaAccess": 1.064283847808838,
+      "LegalizeSundaMacro": 0.3716259002685547,
+      "LegalizeType": 0.17159318923950195,
+      "LocalLayoutOpt": 0.7340552806854248,
+      "LoopFusion": 0.35437893867492676,
+      "LoopSplitting": 0.03284788131713867,
+      "LowerBroadcast": 0.09696078300476074,
+      "LowerCCOpBlockAxis": 0.4246203899383545,
+      "LowerComplexBroadcast": 0.1469728946685791,
+      "LowerIntrinsics": 0.6067290306091309,
+      "LowerTensorOp": 0.625375509262085,
+      "LowerTranspose": 0.4378535747528076,
+      "MacroGeneration": 2.3300650119781494,
+      "MaskPropagation": 0.13726162910461426,
+      "MemcpyElimination": 9.993162155151367,
+      "MutateDataType": 0.04567551612854004,
+      "NeuronAliasDependencyInduction": 0.018932819366455078,
+      "NeuronAliasDependencyReset": 0.03390645980834961,
+      "NeuronInstComb": 0.35028982162475586,
+      "NeuronLICM": 0.2647576332092285,
+      "NeuronLoopFusion": 0.7789182662963867,
+      "NeuronLoopInterchange": 0.04993152618408203,
+      "NeuronSimplifier": 0.3916890621185303,
+      "NeuronSimplifyPredicates": 0.0635371208190918,
+      "NeuronValueNumbering": 0.09003329277038574,
+      "OptimizeAliasedCopyChain": 0.024803638458251953,
+      "OptimizeNKIKernels": 0.3809840679168701,
+      "PAGLayoutOpt": 44.665672302246094,
+      "PComputeCutting": 0.327991247177124,
+      "PGLayoutTilingPipeline": 59.74672317504883,
+      "PGTiling": 5.59335994720459,
+      "PadElimination": 0.014229774475097656,
+      "ParAxesAnnotation": 43.884578704833984,
+      "PartialLoopFusion": 0.4160308837890625,
+      "PartialSimdFusion": 0.35861706733703613,
+      "PerfectLoopNest": 0.06206011772155762,
+      "RecognizeOpIdiom": 0.14852476119995117,
+      "Recompute": 0.00867772102355957,
+      "RelaxPredicates": 0.12411117553710938,
+      "Rematerialization": 0.15559935569763184,
+      "ReshapeWeights": 0.0252685546875,
+      "ResolveAccessConflict": 0.23229002952575684,
+      "ResolveComplicatePredicates": 0.08260774612426758,
+      "RewriteReplicationMatmul": 0.04953479766845703,
+      "RewriteWeights": 0.06951451301574707,
+      "SFKVectorizer": 5.91412353515625,
+      "SimpleAllReduceTiling": 0.05823469161987305,
+      "Simplifier": 0.12264633178710938,
+      "SimplifyMacroPredicates": 0.19836044311523438,
+      "SimplifyNeuronTensor": 0.7505025863647461,
+      "SimplifySlice": 0.033946990966796875,
+      "SimplifyTensor": 0.20671844482421875,
+      "SpillPSum": 0.6507272720336914,
+      "SplitAPUnionSets": 0.25449252128601074,
+      "SplitAccGrp": 0.04361081123352051,
+      "StaticProfiler": 0.44228219985961914,
+      "StaticTransposeLocalTensor": 0.25648951530456543,
+      "SundaISel": 1.4593234062194824,
+      "TCTransform": 0.0390927791595459,
+      "TensorInitialization": 0.1486976146697998,
+      "TensorOpSimplifier": 0.5177226066589355,
+      "TensorOpTransform": 1.8292319774627686,
+      "TileCCOps": 0.2758169174194336,
+      "TilingProfiler": 0.45922422409057617,
+      "TransformConvOp": 0.05972599983215332,
+      "TritiumFusion": 1.8706681728363037,
+      "ValueNumbering": 0.09980320930480957,
+      "VectorizeDMA": 0.12941217422485352,
+      "VectorizeMatMult": 0.02786874771118164,
+      "WeightCoalescing": 0.04949593544006348,
+      "ZeroSizeTensorElimination": 0.00039505958557128906
+    },
+    "tensorizer": {
+      "DMATilingProfiler::TotalInstructionsAfterTiling": 272515,
+      "StaticProfiler::AifUb": 8.01248550415039,
+      "StaticProfiler::ArithmeticIntensityTensorizer": 8.703800201416016,
+      "StaticProfiler::AverageDmaLength": 5833.71923828125,
+      "StaticProfiler::AverageFractalPeUtilization": 99.4205093383789,
+      "StaticProfiler::AveragePartitionUtilization": 98.67665100097656,
+      "StaticProfiler::AveragePeUtilization": 97.56974029541016,
+      "StaticProfiler::DDRTransferBytes": 7663108604,
+      "StaticProfiler::InternalTransferBytes": 296554544,
+      "StaticProfiler::LoadExpanded": 1250875,
+      "StaticProfiler::LocalizationEfficiency": 108.62796020507813,
+      "StaticProfiler::LocalizationEfficiencyIgnoreNonlocal": 108.75718688964844,
+      "StaticProfiler::StoreExpanded": 2218,
+      "StaticProfiler::TotalDMAExpanded": 1253093,
+      "StaticProfiler::TotalDynamicInstancesCount": 278788,
+      "StaticProfiler::TotalDynamicInstancesWithMmPackedCount": 273565,
+      "StaticProfiler::TotalLNCComm": 0,
+      "StaticProfiler::TotalLNCCommTransfer": 0,
+      "TilingProfiler::AveragePartitionUtilizationAfterTiling": 0,
+      "TilingProfiler::AveragePeUtilizationAfterTiling": 0,
+      "TilingProfiler::BatchnormInstructionsAfterTiling": 0,
+      "TilingProfiler::DmaInstructionsAfterTiling": 0,
+      "TilingProfiler::GenericInstructionsAfterTiling": 281,
+      "TilingProfiler::MatMultInstructionsAfterTiling": 234496,
+      "TilingProfiler::NumPfTransposes": 330,
+      "TilingProfiler::NumPfTransposesForIo": 33,
+      "TilingProfiler::NumPfTransposesForLocal": 162,
+      "TilingProfiler::NumPfTransposesForNonlocal": 135,
+      "TilingProfiler::PfTransposeInstructions": 13760,
+      "TilingProfiler::PfTransposeInstructionsForIo": 8208,
+      "TilingProfiler::PfTransposeInstructionsForLocal": 1316,
+      "TilingProfiler::PfTransposeInstructionsForNonlocal": 4236,
+      "TilingProfiler::ReduceInstructionsAfterTiling": 643,
+      "TilingProfiler::SimdInstructionsAfterTiling": 7821,
+      "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_bk4/neuron_config.json

@@ -0,0 +1,213 @@
+{
+  "_attn_implementation_autoset": false,
+  "_name_or_path": "mistralai/Mistral-7B-Instruct-v0.3",
+  "add_cross_attention": false,
+  "architectures": [
+    "MistralForCausalLM"
+  ],
+  "attention_dropout": 0.0,
+  "attribute_map": {},
+  "bad_words_ids": null,
+  "begin_suppress_tokens": null,
+  "bos_token_id": 1,
+  "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": 2,
+  "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": 14336,
+  "is_decoder": false,
+  "is_encoder_decoder": false,
+  "label2id": {
+    "LABEL_0": 0,
+    "LABEL_1": 1
+  },
+  "length_penalty": 1.0,
+  "max_length": 20,
+  "max_position_embeddings": 32768,
+  "metadata": null,
+  "min_length": 0,
+  "model_type": "mistral",
+  "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": "NeuronLlamaAttention",
+    "attn_kernel_enabled": null,
+    "attn_tkg_builtin_kernel_enabled": false,
+    "attn_tkg_nki_kernel_enabled": false,
+    "batch_size": 4,
+    "bucket_n_active_tokens": false,
+    "buckets": [
+      2048
+    ],
+    "cast_type": "config",
+    "cc_pipeline_tiling_factor": 1,
+    "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": false,
+    "is_prefix_caching": false,
+    "k_cache_transposed": false,
+    "kv_cache_batch_size": 4,
+    "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": 4,
+    "max_context_length": 2048,
+    "max_length": 2048,
+    "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": 1,
+    "n_positions": 2048,
+    "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": 2048,
+    "pa_num_blocks": 4,
+    "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": 2048,
+    "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": 4,
+    "speculation_length": 0,
+    "start_rank_id": 0,
+    "target": null,
+    "tile_cc": false,
+    "tkg_batch_size": 4,
+    "token_generation_buckets": [
+      2048
+    ],
+    "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": 32,
+  "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-05,
+  "rope_theta": 1000000.0,
+  "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": "4.42.0.dev0",
+  "typical_p": 1.0,
+  "use_bfloat16": false,
+  "use_cache": true,
+  "vocab_size": 32768
+}