v0.2.3

What's Changed

• [CI] Fail the distributed CI job if any unit test fails by @yf225 in #1125
• Add DE-Surrogate hybrid autotuner algorithm + early stopping option for DE and DE-Surrogate by @FranciscoThiesen in #1096
• Update AGENTS.md by @jansel in #1128
• Add Settings.persistent_reserved_sms by @jansel in #1129
• Add Settings.autotune_force_persistent by @jansel in #1130
• [CI] Fix fbcode test_breakpoint error by @yf225 in #1132
• Auto-select index_dtype by @jansel in #1131
• Support tuple indexing by hl.static_range iterator by @yf225 in #1134
• Fix CI to surface errors correctly, fix all existing errors by @yf225 in #1138
• Workaround TRITON_INTERPRET bug breaking tests by @jansel in #1139
• Fix size 0 tensor handling by @jansel in #1140
• [Benchmark CI] Print generated Triton code for the best config by @yf225 in #1142
• Use pyrefly for type checking by @rchen152 in #1143
• fix pyrefly errors by @oulgen in #1144
• [CI] Skip TestBreakpoint in ref-eager CI job by @yf225 in #1141
• Bump pyrefly to 0.42.1 and remove 'sed' workaround. by @rchen152 in #1145

New Contributors

• @FranciscoThiesen made their first contribution in #1096
• @rchen152 made their first contribution in #1143

Full Changelog: v0.2.2...v0.2.3

v0.2.2

What's Changed

• [Benchmark] Update welford torch.compile function name by @yf225 in #1029
• chore: Bump actions/upload-artifact from 4 to 5 by @dependabot[bot] in #1030
• chore: Bump actions/download-artifact from 5 to 6 by @dependabot[bot] in #1031
• [Benchmark CI] Set welford num_inputs to 6 to avoid timeout by @yf225 in #1032
• Default config: reduce block_size and num_stages to avoid shared mem OOM by @yf225 in #1033
• Default config: reduce block_size further to avoid shared mem OOM by @yf225 in #1034
• Disable autotuner progress bar in fbcode unit test by @yf225 in #1035
• Always print cached config by @oulgen in #1036
• Fix dtype mismatch error in se_block example by @yf225 in #1040
• Upgrade clang version by @oulgen in #1043
• Fix missing static_shapes=False in deployment_autotuning.md by @jansel in #1042
• Fix matmul output dtype to match PyTorch eager behavior by @yf225 in #1044
• Fix layernorm bwd unit test by @yf225 in #1047
• Fix FlattenedTileStrategy to handle unit-sized block dimensions by @yf225 in #1048
• [CI] Fix debug_str() to be compatible with latest PyTorch nightly by @yf225 in #1050
• [Fix upcoming CI error] Set current node in inductor lowering by @yf225 in #1052
• Remove Section Navigation pane from Deployment and Autotuning page. by @choijon5 in #1051
• Add settings.autotune_baseline_fn to allow passing in custom baseline function to autotuner by @yf225 in #1054
• Add HELION_PRINT_REPRO=1 to print Helion kernel repro script to console by @yf225 in #1049
• Fix caching for CPUs by @oulgen in #1055
• Add get_num_sm for cpu by @oulgen in #1056
• Support torch.rand / torch.rand_like with dynamic tile sizes by @yf225 in #1057
• Remove line numbers from expected files by @oulgen in #1061
• Ignore passed in config when force autotune is turned on by @oulgen in #1060
• Update Watch Talk link to Triton conf talk. by @choijon5 in #1058
• Helion Puzzle docs bug fixes by @Athe-kunal in #1062
• Update test_persistent_kernels.expected by @jansel in #1070
• Make HELION_PRINT_REPRO=1 take effect in more error cases by @yf225 in #1066
• add geglu backward by @parsshar-RH in #1069
• [Unblock internal] Fix log capture issue on internal tests by @yf225 in #1076
• Add best effort triton-cpu support by @oulgen in #1037
• Update test_debug_utils.py by @oulgen in #1077
• Raise user error if device-loop is empty after DCE by @yf225 in #1074
• Add GRPO loss example by @ighoshsubho in #1063
• Use HELION_PRINT_REPRO=1 to print repro when device IR lowering or Triton codegen error by @yf225 in #1078
• add AMD demo link by @vivienfanghuagood in #1068
• Update test.yml by @oulgen in #1083
• Fix GRPO loss example unit tests by @yf225 in #1079
• Remove requirements.txt by @oulgen in #1088
• Relax requirements for inline_triton output_like=None by @jansel in #1087
• feat(autotuner): Make autotune cache class configurable via env var by @fulvius31 in #1071
• Add support for while and pass by @jansel in #1090
• Update sphinxtheme to pull from pypi package by @sekyondaMeta in #1091
• [Autotuner] Better error message for default config error by @yf225 in #1092
• Ignore illegal instruction errors by @jansel in #1093
• Update talk links to PTC version by @jansel in #1094
• Add autotuning log by @jansel in #1095
• Fix builtin min / max handling in device loop by @yf225 in #1085
• Add skipIfRocm to failing test on main by @jansel in #1101
• Fix lint in newer triton by @jansel in #1098
• Add AGENTS.md by @jansel in #1100
• Refactor _decorators.codegen to allow multiple backends by @jansel in #1099
• Add extra line before repro log; update repro log tests by @yf225 in #1102
• Refactor inductor_lowering.py into two files by @jansel in #1103
• Use CPU machine for triton-cpu by @oulgen in #1105
• Fix no libdw.so issue on AMD CI by @yf225 in #1107
• Fixes in helion puzzles by @Athe-kunal in #1104
• Add distributed CI job (4xH100) and example unit tests by @yf225 in #1106
• Generalize aten_lowering.py for multiple backends by @jansel in #1108
• Support tensor.T for transpose by @yf225 in #1110
• Add warning to discourage use of acc += lhs @ rhs pattern by @yf225 in #1111
• Remove @helion.jit usage and advise use of @helion.kernel by @yf225 in #1116

New Contributors

• @Athe-kunal made their first contribution in #1062
• @parsshar-RH made their first contribution in #1069
• @ighoshsubho made their first contribution in #1063
• @vivienfanghuagood made their first contribution in #1068
• @fulvius31 made their first contribution in #1071

Full Changelog: v0.2.1...v0.2.2

v0.2.1

What's Changed

• No autotuning on block_ptr if tma is available by @PaulZhang12 in #997
• Add reps for benchmarking stability by @PaulZhang12 in #999
• Prioritize outermost loop for warp spec by @PaulZhang12 in #1000
• Add backward pass for softmax kernel by @karthickai in #744
• Fix linter in softmax by @oulgen in #1003
• Fix test_examples.expected by @oulgen in #1002
• Beef up caching tests by @oulgen in #1001
• Add HELION_ASSERT_CACHE_HIT to debug/explain cache miss by @oulgen in #1006
• Better error message for calling Helion kernel from another kernel by @yf225 in #1008
• Assert that we are cache hitting on the CI by @oulgen in #1007
• Always raise FailedToUnpackTile when for tile_m, tile_d in hl.tile(m, d) is used by @yf225 in #1009
• Adding demo for running softmax kernel on Google colab by @choijon5 in #944
• int4 gemm accurate baselines by @PaulZhang12 in #1010
• Add sitemap xml by @sekyondaMeta in #1013
• [helion] backward support for swiglu by @shunting314 in #756
• Raise informative error when hl.dot with 3D inputs have batch dim mismatch by @yf225 in #1012
• [CI] Fix AMD journal check errors by @yf225 in #1016
• Support breakpoint() in device code when interpret mode is on by @yf225 in #1020
• Sort requirements file by @oulgen in #1021
• Better type checking for eviction policies by @oulgen in #1024
• Bump linter versions by @jansel in #1018
• Garbage collect expected results by @jansel in #1017
• Make indexing choice a list by @oulgen in #1025
• [Docs] Add list of indexing autotuning docs by @oulgen in #1027
• Make store indexing also individually tunable by @oulgen in #1028

New Contributors

• @shunting314 made their first contribution in #756

Full Changelog: v0.2.0...v0.2.1

v0.2.0

What's Changed

• Verify compiled kernels in subprocess by @jansel in #914
• Auto-shrink autotune_precompile_jobs based on free memory by @jansel in #940
• Make HELION_FORCE_AUTOTUNE or kernel.autotune() skip the cache by @jansel in #930
• Support warp specialization on B200 by @oulgen in #935
• Update README.md by @oulgen in #943
• Register tile symbol origin, to support tile + offset use case in blackwell attention by @yf225 in #939
• [CI] Print failed tests by @oulgen in #942
• Update examples to use run_example by @jansel in #941
• blackwell attn with triton attr set by @v0i0 in #918
• Set static_shapes=True by @oulgen in #937
• run.py env var to skip exception logging by @v0i0 in #946
• Fix bug with unit sized dims and block_sizes by @jansel in #932
• Update static_shapes docs by @jansel in #951
• Add tile.count by @oulgen in #955
• Auto detect low vram by @oulgen in #956
• [CI] Use official PyTorch 2.9 by @oulgen in #962
• Use interleaved_bench for run_example by @jansel in #945
• Generalize tile_with_offset pass by @jansel in #949
• Docstring updates by @jansel in #952
• Import updates by @jansel in #953
• Add missing environment variables to docs by @jansel in #957
• Print out errors vs timeouts in autotuning status by @jansel in #960
• Add HELION_AUTOTUNE_IGNORE_ERRORS by @jansel in #961
• Exit autotuning faster on KeyboardInterrupt by @jansel in #963
• Remove default settings by @jansel in #964
• Add missing settings environment variables by @jansel in #965
• Skip test_differential_evolution_search due to slowness by @jansel in #968
• [Benchmark CI] Give nightly job permissions by @oulgen in #970
• [Benchmark CI] Allow kicking off workflow dispatch by @oulgen in #971
• [Benchmark CI] Allow specifying custom env vars via UI by @yf225 in #972
• [blackwell attn example] qk scale as param by @v0i0 in #969
• [Benchmark CI] Allow specifying custom args to benchmark runner via UI by @yf225 in #974
• Add initial backwards compatibility tests by @oulgen in #958
• Remove unrolling + warp spec by @PaulZhang12 in #967
• [Benchmark CI] Set atol and rtol to 1e-2 by @yf225 in #976
• [Benchmark] Fix tritonbench auto-installation by @yf225 in #980
• [Autotuner] Fix fork-based autotuner to avoid re-initializing CUDA context in subprocess by @yf225 in #981
• Make fork default precompilation strategy by @oulgen in #979
• [benchmarks] change tritonbench path by @xuzhao9 in #966
• Add skipIfA10G decorator by @yf225 in #982
• Suggest HELION_AUTOTUNE_PRECOMPILE=spawn when IMA happens by @jansel in #984
• Layer Norm bwd kernel to support large B*M case used by internal by @yf225 in #973
• Fix timeouts in autotuning by @jansel in #985
• Log generated triton code at the DEBUG level rather than INFO by @jansel in #986
• Remove extra debug log for timeouts by @jansel in #987
• Add squeeze_and_excitation_net kernel by @mengluy0125 in #870
• Generalize test cases to support XPU by @EikanWang in #983
• Updated README with News section of upcoming events. Added link to GPU mode talk. by @choijon5 in #991
• Update README.md by @oulgen in #992
• Update README.md by @oulgen in #993
• Mamba2 Chunk Scan & State by @v0i0 in #950
• Remove unrolling with tma + pipelining by @PaulZhang12 in #994
• Add provenance annotations to output code by @jansel in #988

Full Changelog: v0.1.8...v0.2.0

v0.1.8

What's Changed

• fix rmsnorm fwd tritonbench by @v0i0 in #840
• Update input shapes for example kernels by @yf225 in #845
• Extend eviction policy tests to all indexing types by @oulgen in #833
• [Docs] Remove early development warning by @oulgen in #846
• [Docs] Add link to gpumode discord by @oulgen in #847
• [Docs] Add PTC promotional material by @oulgen in #848
• [Benchmark] Add low mem dropout example by @karthickai in #641
• Update lint.yml by @oulgen in #854
• Remove hl.register_reduction_dim API by @yf225 in #834
• Error message for boolean masking or torch.nonzero by @yf225 in #687
• Remove hardcoded block_size=1 usage in attention kernel example by @yf225 in #843
• Revert "Update to use the new attribute setting for tf32." by @choijon5 in #856
• Decrease num_stages default from 3 to 2, to avoid shared memory OOM by @yf225 in #841
• Allow user-defined specialization key by @jansel in #853
• [Benchmark CI] Use fewer num_inputs for flash_attention to avoid timeout by @yf225 in #857
• Remove legacy register_inductor_lowering code by @yf225 in #864
• Set setstate/getstate methods to Config by @jansel in #868
• [doc] Add deployment/autotuning guide by @jansel in #869
• [Benchmark CI] Use equally-spaced-k mode to sample input shapes by @yf225 in #861
• Fix sphinx warnings by @jansel in #871
• Normalize tl.sqrt and libdevice.sqrt for tests by @oulgen in #866
• [CI] Pin py3.10 and one py3.12 on pytorch2.9 by @oulgen in #858
• [Docs] Suggest PyTorch 2.9 or above by @oulgen in #859
• [Benchmark] Pin benchmarks to PyTorch 2.9 by @oulgen in #860
• Print Triton code when error for easier debugging by @yf225 in #874
• Terminate autotuning faster if progress is minimal by @oulgen in #855
• Update README.md by @oulgen in #877
• [CI] pin b200 to pytorch2.9 by @oulgen in #878
• [Autotuner] Run CUDA synchronize before / after candidate func call, to surface CUDA errors sooner by @yf225 in #872
• [Benchmark] bf16 x int16 helion kernel by @karthickai in #794
• Install git for benchmarks by @oulgen in #882
• Pin AMD to 6.4.4 by @oulgen in #883
• Faster int4 gemm by @PaulZhang12 in #751
• Pin AMD to 6.4.4 by @oulgen in #881
• Remove PyTorch requirement from deps so that it is easier to install arbitrary version of pytorch by @oulgen in #879
• [Benchmark CI] Use regular matmul instead of split-k by @yf225 in #884
• [Benchmark] Use bespoke setup-python action by @oulgen in #885
• [Benchmark] Drop memory bound kernels and replace them with gemms by @oulgen in #887
• Add dependabot by @oulgen in #888
• Update dependabot.yml by @oulgen in #891
• chore: Bump actions/setup-python from 5 to 6 by @dependabot[bot] in #893
• chore: Bump actions/download-artifact from 4 to 5 by @dependabot[bot] in #895
• chore: Bump actions/upload-pages-artifact from 3 to 4 by @dependabot[bot] in #894
• chore: Bump actions/checkout from 4 to 5 by @dependabot[bot] in #892
• Upgrade ruff==0.14.0 by @jansel in #889
• [Benchmark CI] grouped_gemm: include input preproc in timing measurement; update gemm backend name mapping by @yf225 in #898
• chore: Bump astral-sh/setup-uv from 6 to 7 by @dependabot[bot] in #896
• [Benchmark] use logger.exception for process errors by @oulgen in #902
• [Benchmark CI] Reduce num_inputs for grouped_gemm and gemm benchmarks by @yf225 in #903
• Query minimum dot size for XPU by @EikanWang in #900
• Add matmul/addmm bwd examples and add test coverage by @tianrengao in #748
• [CI] Pin amd to rocm7.0 by @oulgen in #907
• [Benchmark] Move benchmark kernel sharding to dispatch by @oulgen in #905
• [Benchmark] Provide a way to pass custom list of kernels by @oulgen in #906
• [Benchmark CI] Use triton_tutorial_matmul for triton matmul baseline by @yf225 in #911
• Remove cache around set_triton_allocator by @oulgen in #912
• Add int4_gemm by @oulgen in #917
• chore: Bump actions/github-script from 7 to 8 by @dependabot[bot] in #916
• Catch missing cudnn error by @jansel in #873
• Add progress bar for precompiling by @jansel in #919
• Adding new setting, autotune_effort=[none/quick/full] by @choijon5 in #913
• Print error message for torch.chunk / torch.unbind to redirect users to hl.split by @yf225 in #921
• Avoid setting default --input-sample-mode to equally-spaced-k by @yf225 in #922
• Remove triton_helpers.* usage in lifted device function arguments by @yf225 in #849
• Set HELION_DEV_LOW_VRAM=1 on a10g CI machines by @yf225 in #923
• Suggest use of @helion.kernel(index_dtype=torch.int64) if index offset is out of bound for int32 by @yf225 in #850
• Deprecate use_default_config and replace all its uses with autotune_effort by @choijon5 in #924
• Support hl.arange() with non-power-of-2 input by @yf225 in #862
• Setting up RunLLm AI Chatbot by @sekyondaMeta in #925
• Generalize examples with the DEVICE variable by @adam-smnk in #915
• Fix lint error by @jansel in #926
• Add lint to make sure examples and tests use device=DEVICE by @oulgen in #929
• Support tile+offset and tensor descriptors by @jansel in #928
• Fix triton/torch.compile compability issue by @jansel in #927
• Fix CUDA IMA from combination of unrolling + pipelining by @PaulZhang12 in #920
• Update the Agent ID by @sekyondaMeta in #931
• [Benchmark CI] Use --non-square flag for gemm by @yf225 in #938

New Contributors

• @dependabot[bot] made their first contribution in #893
• @tianrengao made their first contribution in #748

Full Changelog: v0.1.7...v0.1.8

v0.1.7

What's Changed

• Generalize the cuda-bias test cases by replacing hardcoded "cuda" literal with the DEVICE variable by @EikanWang in #775
• Make progress bar prettier by @oulgen in #786
• Upgrade ruff==0.13.3 pyright==1.1.406 by @jansel in #790
• Add hl.split and hl.join by @jansel in #791
• Generalize test_print and test_tensor_descriptor to support different accelerators by @EikanWang in #801
• Limit rebench to 1000 iterations by @jansel in #789
• Turn down autotuner defaults by @jansel in #788
• Enable torch.xpu._XpuDeviceProperties in Helion kernel by @EikanWang in #798
• Better error message for augmented assignment (e.g. +=) on host tensor without subscript by @yf225 in #807
• Add Pattern Search autotuning algorithm to docs. by @choijon5 in #810
• Support 0dim tensor in output code printing by @oulgen in #806
• Set range_num_stages <= 1 if using tensor_descriptor, to avoid CUDA misaligned address error by @yf225 in #792
• Add hl.inline_triton API by @jansel in #811
• Add out_dtype arg to hl.dot by @jansel in #813
• Add autotune_config_overrides by @jansel in #814
• Reduce initial_population to 100 by @jansel in #800
• Disable range_num_stages for kernels with aliasing by @jansel in #812
• Adding new setting, autotune_max_generations, that allows user to set the maximum number of generations for autotuning by @choijon5 in #796
• Increase tolerance for test_matmul_reshape_m_2 by @jansel in #816
• Update docs by @jansel in #815
• Fix torch version check by @adam-smnk in #818
• [Benchmark] Keep going when a single benchmark fails by @oulgen in #820
• Faster Helion JSD by @PaulZhang12 in #733
• Faster KL Div by @PaulZhang12 in #822
• Normalize device name and decorate cuda-only test cases by @EikanWang in #819
• Improved log messages for autotuning by @choijon5 in #817
• Apply simplification to range indexing in order to reuse block size symbols by @yf225 in #809
• Fix hl.rand to use tile specific offsets instead of fixed offsets, ensure unique random num per tile by @karthickai in #685
• Match cuda versions for benchmark by @oulgen in #828
• Print nvidia-smi/rocminfo by @oulgen in #827
• Dump nvidia-smi/rocminfo on benchmarks by @oulgen in #829
• Add 3.14 support by @oulgen in #830
• Remove py312 vanilla test by @oulgen in #831
• Pad to next power of 2 for hl.specialize'ed shape value used in device tensor creation by @yf225 in #804
• Autotune eviction policy by @oulgen in #823
• [Docs] Consistent pre-commit/lint by @oulgen in #836
• [Docs] Recommend venv instead of conda by @oulgen in #837
• [Docs] Helion works on 3.10 through 3.14 by @oulgen in #838
• [Docs] Add eviction policy by @oulgen in #839
• Update to use the new attribute setting for tf32. by @choijon5 in #835

Full Changelog: v0.1.6...v0.1.7

v0.1.6

What's Changed

• ci: Always auth for benchmarking workflows by @seemethere in #719
• [Benchmark] jagged_sum kernel and test by @Sibylau in #676
• Skip default config printing if in ref eager mode by @yf225 in #721
• [Benchmark CI] Make benchmark runner respect custom CLI args by @yf225 in #723
• Upgrade rocm CI to 7.0 by @oulgen in #720
• Add eviction policy argument to tl.load by @oulgen in #714
• [CI] use complete rocm docker images by @oulgen in #724
• More inconsistent naming by @oulgen in #725
• [Benchmark] jagged_layer_norm kernel and test by @Sibylau in #704
• [Bug fix] Preserve masks on reduction inputs that depend on reduction outputs; fix layer_norm accuracy check failure by @yf225 in #722
• Support torch.matmul with 3D inputs by @yf225 in #715
• Slightly improve logs by @angelayi in #740
• Autotuning Progress Bar by @msaroufim in #739
• make tritonbench optional in run.py so install works again by @v0i0 in #746
• fix new factory when size comes from kwargs by @v0i0 in #750
• Add linting instructions to README by @msaroufim in #763
• Add backward kernel for exp by @aditvenk in #736
• fix roll reduction meta when for ops with none output (like wait), cl… by @v0i0 in #767
• Move upload benchmark results to a separate workflows by @huydhn in #758
• Add flash_attention to benchmarks by @oulgen in #769
• Fix jagged_layer_norm linter error by @yf225 in #770
• Add SIGINT handler for clean interrupt of autotuning background processes by @msaroufim in #766
• Enable tensor descriptor for XPU by @EikanWang in #765
• Fix the issue that the XPU kernels cannot be cached well by @EikanWang in #761
• Print Helion kernel source line in symbolic shape debugging by @yf225 in #771
• ci: Set fail-fast to false by @seemethere in #776
• Add XPU support for RNG operations by @EikanWang in #774
• Enable test_dot for XPU by @EikanWang in #773
• Handle XPU compilation error by @adam-smnk in #779
• Fix type prop for and/or by @oulgen in #781
• Make print output code more robust by @oulgen in #780
• Revert "Add SIGINT handler for clean interrupt of autotuning background processes" by @oulgen in #784
• Add torch compile unit test to helion by @oulgen in #782

New Contributors

• @seemethere made their first contribution in #719
• @angelayi made their first contribution in #740
• @msaroufim made their first contribution in #739
• @aditvenk made their first contribution in #736
• @EikanWang made their first contribution in #765
• @adam-smnk made their first contribution in #779

Full Changelog: v0.1.5...v0.1.6

v0.1.5

What's Changed

• [Benchmark CI] Print config that causes tritonbench accuracy check failure by @yf225 in #716
• Add AMD to benchmarks by @oulgen in #717
• [Docs] Move docs requirements to docs/requirements.txt to make compatible with pypi by @oulgen in #718

Full Changelog: v0.1.4...v0.1.5

v0.1.4

What's Changed

• Update benchmark.yml by @oulgen in #570
• Update benchmark.yml by @oulgen in #571
• [Benchmark] Use custom kernel metric mappings list to accomodate for inconsistent namings by @oulgen in #567
• Add rms norm and cross entropy by @oulgen in #568
• Update benchmark_dispatch.yml by @oulgen in #573
• Update linters by @oulgen in #569
• Print config for PassManager::run triton errors by @jansel in #565
• Error when invalid loop reduction number config is generated by @oulgen in #572
• Add skipIfLowVRAM or use_default_config=False to specific unit tests to enable local testing by @yf225 in #574
• Fix bug with block_size smaller than minimum by @jansel in #575
• Better shape errors for mismatched tile sizes by @jansel in #566
• Print warning if block_size is specified in interpret mode. by @choijon5 in #576
• Run all shapes for benchmarks by @oulgen in #578
• [Benchmarks] Cooldown the GPU before recording results by @oulgen in #579
• [Benchmark] Fix layer_norm accuracy issue by @yf225 in #580
• [Benchmark] Remove hardcoded num_inputs for rms_norm kernel by @yf225 in #581
• Do not benchmark twice by @oulgen in #583
• Add missing functions to docs by @jansel in #586
• hl.atomic_add: support 1D tensor as index by @yf225 in #587
• Add atomic and/or/min/max/cas/xchg by @jansel in #589
• Add test shard with HELION_DEBUG_DTYPE_ASSERTS=1, only run one ref-eager shard by @jansel in #590
• Add link to github to docs by @jansel in #591
• Support layernorm without bias by @mengluy0125 in #585
• Allow passing tritonbench operator instance into kernel benchmark wrapper; Always return lambda for timing measurement by @yf225 in #596
• Add layer_norm backward kernels by @yf225 in #588
• Fix tf32 warning by @jansel in #592
• [Benchmark] geglu example and test by @Sibylau in #582
• Print default config when running with it by @oulgen in #599
• [Benchmark] swiglu example and test by @Sibylau in #584
• Login to Docker from the workflows by @huydhn in #601
• Add rms_norm backward kernels by @mengluy0125 in #597
• Revert "Login to Docker from the workflows" by @oulgen in #604
• Fix static shape typo by @oulgen in #609
• Add small dim size (<16) support to hl.dot and torch.addmm; Always prefer using tl.dot(acc=...) for addmm / baddbmm by @yf225 in #564
• Fix rms_norm and layer_norm by @mengluy0125 in #603
• [Benchmark] jsd kernel and test by @Sibylau in #611
• Refactor autotune error handling by @jansel in #595
• Possible fix for CI failures by @jansel in #617
• [Benchmark] Welford kernel and example by @karthickai in #614
• [Benchmark] kl_div kernel and test by @Sibylau in #615
• Ignore TServiceRouterException errors while autotuning by @jansel in #618
• [Example] int4_gemm kernel example and tritonbench integration by @yf225 in #613
• Set requires_grad=True for rms_norm backward inputs by @yf225 in #629
• Adjust tolerance for test_rms_norm_bwd_dx by @yf225 in #628
• Add more kernels to benchmarking by @oulgen in #632
• Reorder benchmarks by @oulgen in #633
• [Ref Mode] Fix hl.store for complex mask pattern by @yf225 in #621
• Support using block size var outside of hl.tile loop by @yf225 in #619
• [Benchmark CI] Print input shapes and surface problematic Helion config by @yf225 in #626
• Fix ValueError: numel (2097152) exceeds triton maximum tensor numel (1048576) by @mengluy0125 in #625
• Always clear inductor cache before benchmark by @yf225 in #608
• Make hl.specialize work on sequences by @jansel in #636
• Better error for passing Tile to hl.tile by @jansel in #640
• [Example] grouped_gemm kernel example and tritonbench integration by @yf225 in #620
• int4_gemm: remove use_default_config=True by @yf225 in #639
• [Easy][Benchmark CI] Exit job on any exception, for easier error catching by @yf225 in #643
• Avoid skipping CUDA errors that crashes the CUDA context by @yf225 in #645
• Add HELION_AUTOTUNE_RANDOM_SEED env var and autotune_random_seed setting by @yf225 in #644
• Bump linter by @oulgen in #647
• Skip test_autotune_random_seed_from_env_var on rocm by @oulgen in #648
• Fix lint related to welford and also local_cache by @yf225 in #646
• Skip test_autotune_random_seed_from_settings on rocm by @yf225 in #651
• PT Sphinx Theme Test by @sekyondaMeta in #600
• Print static_shapes settings value along with config for accurate repro by @yf225 in #649
• [Benchmark] gather_gemv kernel and test by @Sibylau in #635
• Add HELION_SKIP_CACHE env by @jansel in #653
• [lint] Remove UP038 reference by @jansel in #650
• Fix register_block_size codegen by @yf225 in #659
• Raise better error when hl.atomic_* is used on device tensor by @yf225 in #658
• [Autotune] Filter bad config with accuracy check by @yf225 in #655
• Add hl.rand op with seed arg lowering to tl.rand by @karthickai in #652
• Log autotune random seed for easier repro by @yf225 in #661
• Fix misaligned address error for matmul by @yf225 in #662
• skip gather_gemv code check for B200 and fb_code by @Sibylau in #666
• rms_norm: get weight from function args by @yf225 in #664
• skip full autotune if configs are provided by @xuanzhang816 in #670
• [example] fused_linear_jsd by @v0i0 in #494
• Fix CI by moving B200 to cuda13 and downgrade a100/h100 to cuda12.8 by @oulgen in #674
• No background image by @sekyondaMeta in #663
• Remove github link from index.md by @oulgen in #675
• [Autotune] Allow skipping Triton compilation error by @yf225 in #679
• [Benchmark CI] Run one kernel per gpu to maximize successful kernel reporting by @yf225 in #681
• Fix missing block size constexpr assignment in host code by @yf225 in #678
• [CI] Fix missing setuptools by @yf225 in #680
• faster rms norm backwards kernel by @v0i0 in #624
• [Benchmark CI] Use do_bench cudagraph to avoid profiler failure; select specific kernel impls to run by @yf225 in #682
• [Benchmark CI] use --op instead of --kernel for better tritonbench compat by @yf225 in #694
• Increase tolerance for _validate_against_baseline by @jansel in #691
• [Benchmark CI] Use equally-spaced K input shapes by @yf225 in #689
• Print bad default config if compute baseline fails by @yf225 in #688
• Support HELION_AUTOTUNE_ACCURACY_CHECK=0 by @jansel in #692
• rms norm: improve fwd perf by @v0i0 in #669
• Revert "Add hl.rand op with seed arg lowering to tl.rand (#652)" by @jansel in #698
• [Autotune] Skip Triton shared memory OOM by @yf225 in https://git...

v0.1.3

What's Changed

• Add torch compile to benchmark by @oulgen in #545
• Fix issues with wrong dtypes in generated code by @jansel in #542
• Limit concurrent precompile jobs while autotuning by @jansel in #543
• Create basic helion benchmark runner by @oulgen in #544
• Add multi selection radio buttons by @oulgen in #547
• Fix benchmark condition by @oulgen in #548
• Move to dispatcher model for benchmarking by @oulgen in #549
• Give permissions by @oulgen in #550
• Do not downgrade torch/triton by @oulgen in #551
• Use uv for pip freeze by @oulgen in #552
• Add jagged hstu attention example (i.e. ragged_attention) by @xuanzhang816 in #554
• Install quack/torchbench with no deps by @oulgen in #553
• Update test-reports dir by @oulgen in #556
• torch.rand_like and torch.randn_like support by @yf225 in #530
• [Benchmark] add addmm example and test by @Sibylau in #555
• Kick off benchmarks at midnight by @oulgen in #559
• Use profiler instead of inductor_benchmarker by @oulgen in #560
• Shard kernels by @oulgen in #561
• Add layer_norm and softmax by @oulgen in #562
• [Fix CI] Convert tiles to sizes for all torch.* functions by @yf225 in #563

Full Changelog: v0.1.2...v0.1.3