v0.5.0.0

ONNX-MLIR v0.5.0.0 is now available with exciting new features. We thank everyone who contributed to this release!
Please visit onnx-mlir to learn more about ONNX-MLIR.

Key Updates

• ONNX 1.17.0
• PyBind 2.12.0
• Benchmark 1.8.4
• IBM z17 NNPA Telum II Support Enabled

What's Changed

• Add a python script for generating text using huggingface gpt2 by @tungld in #2983
• Remove a spike of memory usage in ScrubDisposablePass. by @imaihal in #2978
• RunONNXModel.py: Add a --cache-model=path option by @AlexandreEichenberger in #2984
• Enable check-onnx-backend-numerical-nnpa on Jenkins s390x by @tungld in #2985
• RunONNXModel.py: save compilation info into a file when using --save-model or --cache-model by @tungld in #2994
• Fix wrong total number of phases for EmitObj and EmitJNI by @tungld in #2995
• run_gpt2_from_huggingface.py: do not download the onnx data file if it exists by @tungld in #2996
• Opening binary constants files fix on zOS by @christopherlmunoz in #2991
• [NNPA] Memory reduction of stickified constant by stickifying at file writing by @imaihal in #2917
• Option to not emit the full MLIR (only emit .tmp file) by @imaihal in #2997
• RunONNXModel.py: allow to change the default model name by @tungld in #2999
• upgrade to ONNX 1.17.0 (opset 22) by @gongsu832 in #3004
• Add decomposition for ONNXSoftmaxCrossEntropyLossOp by @srcarroll in #2968
• Delay scrubbing disposable elements attrs as long as possible by @tungld in #3006
• Add limitation for BFLOAT supported ops for NNPA by @Sunny-Anand in #3008
• Test the return value of omMMapBinaryFile function and terminate the main program elegantly by @tungld in #3002
• Fix a wrong function call by @tungld in #3012
• Making runtime omunreachable static to support clang compiler by @christopherlmunoz in #3015
• Fix security vulenrabilities by @Sunny-Anand in #3019
• Do not fuse locations when normalizing constants for Add and Mul by @jorickert in #3016
• Handle full reduction over all dimensions by @tungld in #3022
• Use DisposableElementsAttr for ZHigh constant propagation by @tungld in #3013
• Re-enable diagnostic error/warning printing by @AlexandreEichenberger in #3020
• Transform SequenceAt to split for special cases by @chentong319 in #3018
• Add tolerance args to CheckONNXModel.py by @AlexandreEichenberger in #3024
• Return a failure instead of crashing if shape inference can not be run because of unraked operand types by @jorickert in #3023
• upgrade becnhmark by @Sunny-Anand in #3027
• Update llvm-project to llvm/llvm-project@01d233ff403823389f848 by @hamptonm1 in #3011
• Update llvm-project to llvm/llvm-project@af20aff35ec3 by @hamptonm1 in #3032
• Fix biasScaleShape of GroupNormalizationV21 to support ranks > 4 by @jorickert in #3030
• Merge from repo by @AlexandreEichenberger in #3033
• Update llvm-project to llvm/llvm-project@e86910337f98 by @hamptonm1 in #3037
• Best practice by @AlexandreEichenberger in #3039
• [NNPA] Fix some bugs for ReduceMin/Max by @tungld in #3038
• Skip over uninitialized DenseResourceAttrs in verifiers by @jorickert in #3041
• [NNPA] Revise compiler options for quantization by @tungld in #3043
• Update the instruction for building multiple accelerators by @tungld in #3046
• Add a document for quantization on NNPA by @tungld in #3045
• update onnx opset by @Sunny-Anand in #3050
• Remove element type restriction in softmax lowering by @srcarroll in #3051
• Fix ASAN/UBSAN issues in DimAnalysis by @jorickert in #3052
• Build light weight PyRuntime without llvm or onnx-mlir by @chentong319 in #3044
• Option to set the number of threads for parallel compilation by @imaihal in #3048
• Update onnx requirement to 1.17.0 by @jorickert in #3054
• Optimization for Roberta unstick->reshape->transpose->reshape->stick by @AlexandreEichenberger in #3056
• Extend GridSample support by @jorickert in #3060
• Remove the pattern unstick_4ds_squeeze_stick_3ds by @tungld in #3062
• Instrumentation cleanup when operation was removed by @AlexandreEichenberger in #3061
• Add support for ONNX.shape with permutation pattern by @AlexandreEichenberger in #3066
• Update docker image to point to github registry in devcontainer-example by @jorickert in #3055
• Parallelization of ConstProp compilation by @imaihal in #3042
• Bump various ops to opset 22, adding bf16 support by @jorickert in #3059
• Bump onnx.Cast to opset 21 , adding int/uint4 support by @jorickert in #3057
• Add runtime check for Gather Op by @chentong319 in #3069
• fix weak hash by @Sunny-Anand in #3070
• Remove the compile option -nnpa-clip-to-dlfloat-range by @tungld in #3075
• Matmul CPU performance regression by @AlexandreEichenberger in #3072
• ZHigh to ONNX optimization is default on. Switch flag from enable to disable by @AlexandreEichenberger in #3074
• Since compiler generated stick/unstick is default on, change new option to disable it by @AlexandreEichenberger in #3073
• Add lit tests for KrnlMatmulOp lowering (Krnl to affine) by @AlexandreEichenberger in #3076
• Upgrading llvm and stablehlo hash by @christopherlmunoz in #3053
• Don't try to free static array in mnist example by @Zentrik in #3049
• Handle out-of-bound value for Gather alike operation by @chentong319 in #3077
• Extend instrumentSignature to print data by @chentong319 in #3078
• Modifying RunONNXModel.py to better support external performance profiling tools by @AlexandreEichenberger in #3082
• Add support to use either docker or local compiler to compile a model by @chentong319 in #3081
• Use docker and podman package in python driver by @chentong319 in #3087
• update pybind11 to version 2.12.0 by @chentong319 in #3088
• Bump Upsample to Opset 10 and change the opset versioning to allow to skip over opset versions if a newer, backwards compatible one exists. by @jorickert in #3065
• Improve scripts by @AlexandreEichenberger in #3089
• Add result type inference to RandomNormalLike and fix wrong hardcodings for dtypes by @jorickert in #3091
• Bump various ops to opset 21, adding int4/uint4 and 8 bit float support. by @jorickert in #3064
• Added minimal support to do some timing of OM Runtime functionality by @AlexandreEichenberger in #3095
• Including __errno_location call for MVS by @christopherlmunoz in #3099
• Rewriting pattern to remove WhereOp and EqualOp. by @imaihal in #3094
• Enable NNPA saturation by default and change the option to --nnpa-disable-saturation by @tungld in #3101
• removing weak attribute of errorno by @christopherlmunoz in #3103
• Fix the custom build link for docs/Docker.md by @qjivy in #3104
• Python driver for torch model by @chentong319 in #3093
• Cherry pick updates from main for z17 and fix for ZHighConstantPropagation in QunarizedStick by @Sunny-Anand in #3133
• [cherry-pick]fix CVE-2025-32434 (#3135) by @Sunny-Anand in #3138
• Removing duplicate file (#3146) by @Sunny-Anand in #3158

New Contributors

• @Zentrik made their first contribution in #3049
• @qjivy made their first contribution in #3104

NNPA Contributors for z17

• @AlexandreEichenberger
• @gongsu832
• @imaihal
• Megan Hampton [email protected]
• Mike Essenmacher [email protected]
• @Sunny-Anand
• @chentong319
• @tungld
• @negiyas
• @christopherlmunoz

Full Changelog: v0.4.3.0...v0.5.0.0