v1.2.0

What's Changed

• [Pallas] Add Pallas interpret CI job (revival of #1938) by @norx1991 in #2491
• Disable rope-bwd from nightly perf CI for now by @choijon5 in #2494
• torch_tpu: update pin to f863c2bacd52aaef02599a0e80f53a24fe9431c0 by @cota in #2528
• Improve CuTe RMS norm reduction path by @oulgen in #2540
• [pallas] support dynamic shapes in indirect gather lowering by @cota in #2493
• Enable >32-thread persistent reduction on CuTe for RMS norm perf by @oulgen in #2541
• Cache CuTe launcher imports and dtype mapping in hot path by @oulgen in #2542
• Add CuTe reduction-kernel autotuner seed by @oulgen in #2543
• Fuse two-pass loads of x in CuTe reduction kernels by @oulgen in #2544
• Add wide-chunk reduction seed and tidy fuse-two-pass-loads helper by @oulgen in #2545
• [Pallas] correct RNG seed buffer argument ordering in launcher by @AmesingFlank in #2539
• [cutedsl] move CuTe state out of DeviceFunction by @jansel in #2497
• [cutedsl] split CuTe tcgen05 config logic by @jansel in #2498
• [cutedsl] replace CuTe golden codegen tests by @jansel in #2499
• [cutedsl] handle CuTe edge epilogue autotune by @jansel in #2500
• [cutedsl] support flat edge TMA stores with scalar fallback by @jansel in #2501
• [cutedsl] enable clustered edge TMA and seed configs by @jansel in #2502
• [cutedsl] add hybrid TMA store pipeline for edge tiles by @jansel in #2503
• [cutedsl] tune edge K-tail epilogue placements by @jansel in #2504
• [cutedsl] split scheduler edge epilogues by @jansel in #2505
• [cutedsl] add clc scheduler diagnostics and lifecycle fixes by @jansel in #2506
• [cutedsl] add aux TMA load mode by @jansel in #2507
• [cutedsl] seed aux TMA and clc persistence candidates by @jansel in #2508
• [cutedsl] retile and pack aux epilogue by @jansel in #2509
• [cutedsl] vectorize and stage aux epilogue stores by @jansel in #2510
• [cutedsl] add staged scheduler mailbox diagnostic by @jansel in #2511
• [cutedsl] skip static full tma fallback by @jansel in #2512
• [cutedsl] add tcgen05 lifecycle context and kloop ownership by @jansel in #2513
• [cutedsl] introduce pure matmul object model for tcgen05 by @jansel in #2514
• [cutedsl] add explicit tcgen05 epilogue tile and two-cta kloop by @jansel in #2515
• [cutedsl] cache cute launch arguments by @jansel in #2516
• [cutedsl] add guarded flat tcgen05 role coordinates by @jansel in #2517
• [cutedsl] add tcgen05 tvm ffi launch with seed configs by @jansel in #2518
• [cutedsl] introduce target1 clc scheduler object by @jansel in #2519
• [cutedsl] add target1 direct entry path by @jansel in #2520
• [cutedsl] Update benchmarking code by @jansel in #2521
• [Pallas] int32 gather lowering by @thcmbs in #2550
• [Autotuner] Cap inner-tile block size by outer tile by @thcmbs in #2478
• Use fixed seed for nightly benchmarks by @choijon5 in #2561
• torch_tpu: update pin to 5281ef191211f27bde9144cb17755c4f6a89f8fa by @cota in #2562
• Emit vectorized 128-bit loads for CuTe inner reductions by @oulgen in #2546
• [Pallas] Fix mask broadcast on size-1 tensor dimensions by @AmesingFlank in #2556
• Enable vec loads for bf16/fp16 CuTe reductions via Uint16 bitcast by @oulgen in #2547
• [cutedsl] extend tcgen05 direct-entry path to additional targets by @jansel in #2551
• [cutedsl] add tcgen05_aux_stages and tcgen05_consumer_regs autotune knobs by @jansel in #2552
• [cutedsl] stage dynamic_persistent scheduler object infrastructure by @jansel in #2553
• [cutedsl] Expand hillclimber benchmarking to 32 targets by @jansel in #2554
• Fix XPU CI SYCL runtime setup by @choijon5 in #2566
• (Onboarding task) examples: add XSA (exclusive self-attention) kernel by @calebmkim in #2564
• Vec-aware two-pass load fusion for CuTe reductions by @oulgen in #2548
• [autoreview] Add --codex-no-sandbox flag for hosts without bubblewrap by @choijon5 in #2568
• Support tensor_descriptor + epilogue subtile by @ethche in #2557
• [CI] Bump Pyrefly, ruff, cute by @oulgen in #2569
• [cute] skip test_xsa on cute by @oulgen in #2581
• [cute] Softmax perf: vec hoist, fuser alias, hoist warp reduce, heuristics by @oulgen in #2570
• [cute] V=8 fp16/bf16 via 4+4 split; multi-row investigation tests by @oulgen in #2571
• [cute] Merge sibling constexpr V-loops + elide warp-reduce dtype wraps by @oulgen in #2572
• [cute] Warp-per-row layout for softmax-shape reductions by @oulgen in #2573
• Reject tensor_descriptor indexing when block size exceeds tensor dim (#2555) by @mengluy0125 in #2555
• [cute] LICM for reciprocals: hoist 1/divisor out of inner loops by @oulgen in #2574
• [Pallas] Fix synchronize_device skipping sync for tuple returns on TPU by @thcmbs in #2580
• [Pallas] remove jax_export_ignore_forward_compatibility setting by @cota in #2452
• [cute] Cleaner LICM: alias DCE + FMA-friendly scale hoist by @oulgen in #2575
• [Pallas] Add support for non zero dim in gather by @thcmbs in #2558
• [cute] Software-pipeline inner vec loads to hide HBM latency by @oulgen in #2576
• [cute] Optional carried-only gate for the load pipeline pass by @oulgen in #2577
• [cute] Rewrite online softmax_two_pass → equivalent 3-pass form by @oulgen in #2578
• [test] Split cute compiler-pass tests by pass, drop kernel name from filenames by @oulgen in #2579
• [CI] bump cutlass by @oulgen in #2610
• A jagged_hstu_attention example that works on Pallas TPU by @AmesingFlank in #2218
• torch_tpu: update pin to adbe52d4a6082d23bc40f630b335d37379b604d2 by @cota in #2602
• test_examples: skip test_jagged_dense_add on Pallas by @cota in #2607
• [metal] Auto-cap threadgroup size via CuteND/CuteFlattenedTileStrategy by @aditvenk in #1855
• [cosmetic/UI] Quieter console output and nicer progress formatting by @choijon5 in #2598
• [Pallas] Propagate inner tile alignment min_size to bounding outer tiles by @thcmbs in #2559
• [Pallas] Suppress Pallas RNG x64 warning spam in random tests by @thcmbs in #2617
• [metal] Add DeviceIR MPPGraph lowering for aten matmul by @aditvenk in #2472
• [Pallas] Avoid bool reshape lowering by @thcmbs in #2618
• ci: bump TPU jax/jaxlib pin to 0.10.1 by @thcmbs in #2621
• [Pallas] Fix indirect gather by @thcmbs in #2619
• [Pallas] Add generic fallback scatter by @thcmbs in #2620
• Migrate H100 CI to OSDC runners (pytorch/pytorch cu128 base image) by @huydhn in #2625
• torch_tpu: update pin to dd5b8dd76b1eb0175db812c35b0d2f52ca1f4e3b by @cota in #2633
• [cute] Fix benchmark errors on CI for cute by @oulgen in #2622
• [cute] Require cuda13+ by @oulgen in #2623
• [compile] add ast backend customization hook by @hinriksnaer in #2624
• torch_tpu: update pin to e2250c4179c5022f7d0a4007dcc746f60fde64bb by @cota in https://github.com/pytorch...

v1.1.0

What's Changed

• [Pallas] Use TPU-aware synchronize_device across autotuner and testing by @norx1991 in #1944
• Refactor if-else branches, unblock test_if_new_variable_in_static_range for Pallas TPU by @AmesingFlank in #1935
• [Pallas] Register pallas_loop_type only when inner loops exist by @norx1991 in #1915
• [Pallas] Fix emit_pipeline program_id mapping with loop_order reordering by @norx1991 in #1916
• [Pallas] Add expected-failure test for non-power-of-2 RDIM size by @norx1991 in #1945
• Update docs: add missing API entries by @choijon5 in #1948
• [Benchmark] Add compile time measurement to CI benchmarks by @choijon5 in #1952
• use is_symm_mem_tensor if the API is available by @shunting314 in #1933
• [Pallas] When not doing tiling for entire kernel, use explicit BlockSpecs instead of None by @AmesingFlank in #1960
• Fix FROM_BEST_AVAILABLE matching with hl.specialize() after #1883 by @fulvius31 in #1940
• [Pallas] Fix indexing scalars using SMEM memory space by @AmesingFlank in #1955
• [Pallas] Add xfail tests for scalar .begin index not collapsing dims by @norx1991 in #1971
• Skip test_hl_rand_mixed_argument_order on MTIA due to unaligned address crash by @karthickai in #1977
• Fix _supports_maxnreg() to guard against non-CUDA backends by @karthickai in #1981
• Skip TestRandomPhiloxParity class on MTIA (#1979) by @karthickai in #1982
• [Pallas] Fix acccesing tensors using index from hl.grid() by @AmesingFlank in #1956
• [Pallas] Support using traced size-1 tensor as condition predicate, unblocking test_if_arg_indexed_scalar by @AmesingFlank in #1957
• [Pallas] Fix atomic_add dtype cast and VMEM preload for fori/pipeline launchers by @thcmbs in #1966
• [Pallas] Use exact RDIM size instead of next-power-of-2 by @norx1991 in #1954
• [Pallas] Add support for accessing tensors with the pattern of tile.index + offset /.id/.begin/.end by @AmesingFlank in #1968
• [metal] Reuse Inductor's MetalOverrides for MSL expression emission by @aditvenk in #1853
• [metal] Add Metal codegen handlers for load, store, and mask_to by @aditvenk in #1854
• [Pallas] Enable a subset of test_grid tests for Pallas by @AmesingFlank in #1985
• [Pallas] Add a test for accessing tensors with hl.grid() index + offset by @AmesingFlank in #1988
• Relax rms_norm example tolerance for Pallas bf16 by @thcmbs in #1983
• [autotuner] Introduce BenchmarkProvider abstraction for kernel benchmarking by @hinriksnaer in #1928
• [Pallas] Use HBM BlockSpecs for output-only tensors to save VMEM by @norx1991 in #1984
• Emit autotune failure summary warnings by @allgather in #1994
• [Pallas] Remove unused is_device_loop variable in _pallas_index_str by @AmesingFlank in #1990
• [Pallas] Fix BlockSpecs for 2D tl.grid([m, n]), unblocking test_scalar_access_hl_grid_2d by @AmesingFlank in #1986
• Reduce some autotuner overhead without changing kernel behavior by @svdrecbd in #1885
• [compile] Add pre_codegen hook to Backend ABC by @hinriksnaer in #1976
• Adding AMD Mi350x machines to CI with new labels. by @umechand-amd in #1835
• [Pallas] Fix accessing tensor via hl.grid() index within a device loop, unblocking test_scalar_access_hl_grid_2d_nested by @AmesingFlank in #1989
• Support mtia in LocalAutotuneCache by @Hamlin-Li in #1996
• removed stale comments by @hinriksnaer in #1997
• Fix invalid default config for kernels with large tensor numel by @fulvius31 in #1839
• chore: Bump actions/github-script from 8 to 9 by @dependabot[bot] in #2000
• Fix _benchmark dropping configs that fail compilation by @fulvius31 in #1942
• [metal] Add MSL AST walker for Python-to-C++ translation by @aditvenk in #1794
• Enable tensor_descriptor based atomic ops by @ethche in #1953
• [metal] Add @metal_jit decorator for AST-to-MSL compilation by @aditvenk in #1991
• Epilogue subtiling: store indexing fix, example, and tuple output support in run_example by @choijon5 in #1907
• [Compiler] Added backend registry by @hinriksnaer in #1967
• [metal] Wire @metal_jit into MetalBackend and simplify launcher by @aditvenk in #1992
• [Pallas] Skip trivial reduction mask when RDIM size equals actual dim by @norx1991 in #1993
• [Pallas] Fix TPU min_dot_size for matmul autotuning by @norx1991 in #1999
• chore: Bump actions/upload-pages-artifact from 4 to 5 by @dependabot[bot] in #2015
• Make _get_tile_with_offset_info accept a torch.fx.Node as arg, instead of the entire CodegenState by @AmesingFlank in #2005
• [metal] Skip tests better to pass in internal CI by @aditvenk in #2016
• Add to_code() as backend-agnostic alias for to_triton_code() by @norx1991 in #2012
• [Pallas] Fix fori_loop multi-dim index decomposition with nested loops by @norx1991 in #1917
• custom config filter by @shunting314 in #1847
• [Pallas] Reject 64-bit input tensors and fix tiling ZeroDivisionError by @norx1991 in #1950
• Bump CI libtpu to 0.0.40 by @AmesingFlank in #2017
• [Pallas] Make test_tensor_access_tile_index_offset more meaningful and reveal codegen issue in Pallas backend by @AmesingFlank in #2006
• [Pallas] [NFC] Dedup _jnp_dtype_map helper by @thcmbs in #2021
• [Pallas] Add a plan_tiling pre_codegen pass to make more consistent tiling and indexing decisions by @AmesingFlank in #2007
• [Compiler] Add reserved_launch_param_names to Backend ABC by @hinriksnaer in #1970
• [Pallas] Remove non-tiled fallback paths which are no longer used after #2007 by @AmesingFlank in #2026
• [TPU][Pallas] Add lower bound analytical VMEM estimation and OOM guard for Pallas launchers by @yarongmu-google in #2024
• [Pallas] Fix codegen for slice indexing when there are squeezed dimensions by @AmesingFlank in #2027
• [CI] Pin Triton-to-tile-IR to last known-good commit by @norx1991 in #2028
• [Pallas] Add non-DMA fori_loop fallback for DMA-unaligned inner blocks by @thcmbs in #1969
• Skip test_matmul_smaller_than_min_dot_size on MTIA (#2030) by @will-cromar in #2030
• docs: add missing autofunction directives for 9 API functions by @Bodlux in #2036
• [AMD ROCM] matrix_instr_nonkdim restricted to 16 by @umechand-amd in #2032
• [Autotuner] Move benchmarking implementation into BenchmarkProvider by @hinriksnaer in #2029
• [Bug][Autotuner] Autotuner failed to clone mutated input arg by @xiaohongchen1991 in #2042
• [TPU][Pallas] Fix OOM on large reductions by delegating chunking to Mosaic by @yarongmu-google in #2033
• Remove ref baseline kernel count for test_clone_with_multiple_views_one_mutated as it depends on PyTorch version by @choijon5 in #2037
• docs: remove remaining mentions of HELION_USE_AUTOTUNE by @cota in #2038
• Fix test broken by #2033: use env.backend_name by @norx1991 in #2046
• Simplify CI: update PyTorch releases to 2.11, use bundled Triton by @choijon5 in #2047
• [Pallas] Under interpret mode, Use float16 as HALF_DTYPE because bfloat16 is not supported on CPU by @AmesingFlank in #2050
• [AMD ROCm] Use AMD Triton backend for min_dot_size instead of NVIDIA by @choijon5 in #2048
• [Pallas] Adjust block size constraints by analyzing subscript exprs on tensors, unblock rms_nor...

v1.0.0

What's Changed

• Add PyTorch Conference Europe 2026 events by @choijon5 in #1812
• Fix LFBO surrogate training data duplication by @fulvius31 in #1806
• Move imports to top of file by @jansel in #1804
• [cutedsl] Generalize shape-chain lowering and fix nested matmul argreduce by @jansel in #1787
• Fix docs build failure by @jansel in #1814
• Capping the number of warps to avoid non-performing configs and reducing the autotuning time. by @umechand-amd in #1674
• Add pytorch-probot config to enable retryBot for Helion by @huydhn in #1808
• Update torch_tpu pin by @oulgen in #1816
• avoid nvshmem symm-mem backend by @shunting314 in #1750
• increase signal pad size for dist matmul kernels by @shunting314 in #1753
• move distributed runtime utils out of examples/ folder by @shunting314 in #1771
• Increasing block size dimensions to avoid configs which are slow and poor candidates. by @umechand-amd in #1677
• Cleanup the tests skips on ROCM. by @umechand-amd in #1820
• Fix Pallas buffer donation: only donate in-place mutated tensors by @thcmbs in #1802
• Gracefully handle the error, "failed to translate module to LLVM IR" by @umechand-amd in #1793
• fix to prevent incorrect output in test_print by @umechand-amd in #1821
• HELION_FORCE_AUTOTUNE: skip cache read but write back result by @fulvius31 in #1815
• Support block_ptr/TensorDescriptor with extra_mask for loads by @hinriksnaer in #1768
• [autotuner] Add benchmark_batch as unified entry point for autotuner benchmarking by @hinriksnaer in #1810
• restrict to persistent pid for dist kernels by @shunting314 in #1772
• remove hl.signal/wait by @shunting314 in #1791
• cleanup the autotuner factory function by @shunting314 in #1770
• properly codegen hl.triton_kernel by @shunting314 in #1797
• move compile time dist APIs to _dist_utils.py by @shunting314 in #1799
• Use a larger timeout setting to autotune distributed kernels by @shunting314 in #1800
• [autotuner] removed unused duplicate function check_config_consistancy by @hinriksnaer in #1829
• chore: Bump actions/deploy-pages from 4 to 5 by @dependabot[bot] in #1826
• [pallas-tpu] Add tests for previously untested examples by @v0i0 in #1665
• [cutedsl] Enable more tests by @jansel in #1822
• Unskipping tests on ROCM by @umechand-amd in #1831
• [Helion + torch.compile] Fix dynamic shapes in HOP path and enable test_dynamic_shapes_basic by @yf225 in #1836
• fix regress caused by large block size in dist autotuner by @shunting314 in #1834
• Revert "Add scheduled workflow to rerun GPU health check failures (#1683)" by @huydhn in #1833
• rendezvous only once for matmul reduce scatter by @shunting314 in #1824
• Refactor: use backend *_expr methods in PointerIndexingStrategy by @aditvenk in #1850
• Fix #1081: Skip failed configs in parallel benchmark by @tianrengao in #1673
• Fix pre-commit pyrefly on macOS by sharing ignore-missing-imports logic by @aditvenk in #1858
• Revert "Capping the number of warps to avoid non-performing configs and reducing the autotuning time. (#1674)" by @jansel in #1860
• [autotuning] Refactor PrecompileFuture construction out of BaseSearch by @hinriksnaer in #1817
• [cutedsl] Fix launch axis mapping and serialized reductions by @jansel in #1840
• [Helion + torch.compile] Always use HOP path, gate fusion with torch_compile_fusion kernel setting by @yf225 in #1837
• [cutedsl] Fix CuTe shape/view lowering and loop-carried matmul accumulation by @jansel in #1856
• [Pallas] Fix block_spec_info grid/loop_order mismatch by @norx1991 in #1788
• add script to make tpu wheels by @v0i0 in #1818
• chore: Bump actions/configure-pages from 5 to 6 by @dependabot[bot] in #1882
• Fix Pallas BMM block size alignment via new dim_matches method by @thcmbs in #1828
• Remove FROM_DEFAULT, unify initial population on FROM_BEST_AVAILABLE with pad control by @fulvius31 in #1809
• [Helion + torch.compile] Reject static_shapes=True with dynamic=True to prevent wrong results by @yf225 in #1864
• [Helion + torch.compile] Add bench_compile_config protocol for autotuning by @yf225 in #1865
• Add pallas_attention kernel and default-loop-type test by @v0i0 in #1866
• Skip Triton-specific batch dim constraint for Pallas matmul by @v0i0 in #1867
• [autotune] Migrate precompile future to a dedicated file location by @hinriksnaer in #1872
• [Pallas] Enable test_use_block_size_var_without_hl_tile Pallas test which already passes, removing triton-specific code checks by @AmesingFlank in #1880
• torch_tpu: point to new repo location by @cota in #1886
• [pallas-tpu] treat sublane constraint like lane constraint by @v0i0 in #1887
• Add epilogue subtiling pass by @choijon5 in #1838
• [Pallas] Enable test_batch_softmax for Pallas backend by @norx1991 in #1889
• fix ref eager subtile tests by @v0i0 in #1899
• use post unfiication symbols during indexing by @v0i0 in #1896
• Fix _output_indices to recognize views of inputs by @v0i0 in #1868
• Add emit_pipeline loop-carried state for Pallas attention by @v0i0 in #1869
• Add fori_loop loop-carried state and refactor shared helpers by @v0i0 in #1870
• [metal] Add MetalBackend method overrides for codegen by @aditvenk in #1852
• add CodegenDict to better handle common codegen function by @shunting314 in #1902
• [Helion + torch.compile] Add extra_cache_key() and is_cacheable() to _AutotunableKernel protocol by @yf225 in #1903
• Skip overridden config indices during autotuner neighbor generation by @fulvius31 in #1884
• [Pallas] Fix if branches which modify nonlocal variables, unblocking test_grpo_loss_fwd by @AmesingFlank in #1892
• [Pallas] Add HelionPallasPrinter to avoid triton_helpers in codegen by @norx1991 in #1900
• [pallas] Add interpret=True support for CPU reference mode by @v0i0 in #1841
• [Helion + torch.compile] Move extra_cache_key to its own field in LooseAutotuneCacheKey by @yf225 in #1906
• [Helion + torch.compile] Add fusion-aware autotuning tests (temporarily skipped) by @yf225 in #1904
• [Autotuner] Handle CUDA OOM errors gracefully during autotuning by @yf225 in #1912
• [Helion + torch.compile] Switch to fusion-aware autotuning by @yf225 in #1801
• Cap maxnreg to fit register file budget for num_warps by @v0i0 in #1918
• use the user provide process group for autotuning by @shunting314 in #1823
• disable the flex attention reference in the attention example on tpus by @v0i0 in #1929
• Refactor RNG op handling to be backend-independent by @jansel in #1871
• [cutedsl] Fix tile-origin indexing and atomic/matmul correctness by @jansel in #1873
• Speed up tests by @jansel in #1897
• add max_mismatch_pct argument to run_example by @shunting314 in #1909
• scaled matmul reduce scatter by @shunting314 in #1842
• Use warps_to_threads() instead of hardcoded warp size in maxnreg cap by @v0i0 in #1930
• [Helion + torch.co...

v0.3.3

What's Changed

• bump cublas for b200 benchmarks by @v0i0 in #1666
• Remove need to pass device=... into torch functions by @hinriksnaer in #1657
• [docs] Updated docs to communicate fast math support by @hinriksnaer in #1669
• Fix #762: Replace internal assertion with user-facing error in set_pid by @tianrengao in #1670
• [Helion + torch.compile] Add expected_num_kernels validation to torch.compile tests by @yf225 in #1655
• [pallas-tpu] Fix segfault xfails: softmax tests now pass, reclassify others by @v0i0 in #1721
• Reject data-dependent output shapes in infer_output_spec by @gmagogsfm in #1722
• [pallas-tpu] Add test_long_sum_manual to verify range bound fix by @v0i0 in #1732
• [pallas-tpu] fix default configs for TPU examples by @v0i0 in #1731
• Move Triton-specific implementations from Backend base class to TritonBackend. by @norx1991 in #1728
• Fix lint in ref_mode.py by @jansel in #1681
• Add missing rebenchmark, finishing phase, and effort profile wiring to DESurrogateHybrid by @fulvius31 in #1680
• Faster expected generation in test_indexing.py by @jansel in #1682
• [cutedsl] Implement layout planning phase by @jansel in #1664
• Revert "[pallas-tpu] fix default configs for TPU examples (#1731)" by @norx1991 in #1740
• [pallas-tpu] Fix Pallas test_add by making non-contiguous inputs contiguous in pallas launcher. by @norx1991 in #1737
• [Helion + torch.compile] Refactor HelionTemplateBuffer to use TemplateBuffer base class by @yf225 in #1723
• precompile in the current process by @shunting314 in #1730
• [CI] Fix Pyrefly lint error in template_buffer.py by @yf225 in #1746
• [Autotuner] Add autotune_baseline_accuracy_check_fn for custom accuracy checks by @yf225 in #1733
• Add Dockerfile by @jansel in #1748
• Add scripts/runpod.py by @jansel in #1749
• [cutedls] Initial mma support by @jansel in #1742
• Fix doubled test output in non-distributed CI jobs by @norx1991 in #1741
• Add hl.jagged_tile by @nullplay in #1651
• Fix logging to be compatible with pytest by @bringlein in #1734
• Add autotune_initial_population_strategy kernel setting by @bringlein in #1735
• [CI] Increase atol for test_squeeze_and_excitation_net_fwd on B200 by @yf225 in #1752
• Unpin H100 nightly torch and Triton versions by @v0i0 in #1654
• Enable pyrefly on macOS with ignore-missing-imports by @aditvenk in #1760
• [metal] Register "metal" backend with minimal MetalBackend and launcher by @aditvenk in #1761
• [metal] Respect force_tile_mask() in NDTileStrategy mask generation by @aditvenk in #1762
• Remove past hackathon event from README by @choijon5 in #1780
• Remove deadcode _clone_tree and _assert_args_close by @choijon5 in #1781
• Fix PermutationFragment.encode() returning wrong value by @choijon5 in #1779
• Fix atomic_max ref to return previous value by @choijon5 in #1782
• Fix lints by @jansel in #1775
• Add runpod SKILL.md by @jansel in #1776
• [Helion + torch.compile] Add store/load transform hooks and prologue/epilogue fusion codegen by @yf225 in #1724
• [Helion + torch.compile] Enable torch.compile fusion tests by @yf225 in #1727
• [Helion + torch.compile] Simplify _remap_or_resolve for compound sympy expressions by @yf225 in #1785
• Add scheduled workflow to rerun GPU health check failures by @v0i0 in #1683
• Fix device sync in generic benchmarking functions for TPU/Pallas by @norx1991 in #1773
• Removing skips and in some cases adding skipIfNotCUDA for cuda only features. by @umechand-amd in #1790
• APIs to debug distributed kernel by @shunting314 in #1743
• [cutedsl] tcgen05 MMA support by @jansel in #1777
• [cutedsl] Fix broadcast and reshape-backed matmul lowering by @jansel in #1783
• [cutedsl] Fix packed-RHS lowering and add general stack/reshape views by @jansel in #1784
• Fix ROCm failures on main by @jansel in #1805
• add kernel-filter to select kernel for allreduce-rmsnorm by @shunting314 in #1744
• helion distributed kernel autotuning by @shunting314 in #1532
• [Helion + torch.compile] Add ref baseline kernel count checks by @yf225 in #1786

New Contributors

• @norx1991 made their first contribution in #1728
• @nullplay made their first contribution in #1651

Full Changelog: v0.3.2...v0.3.3

v0.3.2

What's Changed

• [pallas-tpu] Enable test_geglu, test_swiglu on Pallas; fix softmax/kl_div/gdn xfails by @v0i0 in #1637
• [pallas-tpu] make dot work: use jnp.matmul and f32 accumulator by @v0i0 in #1647
• [pallas-tpu] Enable broadcasting and control flow tests for Pallas backend by @v0i0 in #1643
• Use ephemeral Triton cache during autotuning to eliminate loser artifacts by @fulvius31 in #1649
• Try public bazel cache for torch_tpu build by @v0i0 in #1658
• fix cudagraph grouped gemm fail in benchmarks by @v0i0 in #1662
• Use matmul fwd direclty in autograd for performance by @tianrengao in #1045
• Recreate foreign SymInts in helion's own ShapeEnv by @gmagogsfm in #1671

Full Changelog: v0.3.1...v0.3.2

v0.3.1

What's Changed

• [Autotuner] Use chunked comparison in autotuner accuracy checks to reduce peak memory by @yf225 in #1538
• one_shot_allreduce_bias_rmsnorm: alloc&rendezvous symm memory only once by @shunting314 in #1525
• Add Horace as Core Maintainer by @oulgen in #1552
• add two-shot all redcue rms norm kernel by @shunting314 in #1526
• Update bazel build command to use remote caching by @malfet in #1550
• Add BlockSpec support and proper indexing to Pallas launcher by @oulgen in #1548
• Add Pallas codegen for prims.iota by @oulgen in #1556
• [CI] Fix AMD error on test_default_block_sizes_high_dim_with_reduction by @yf225 in #1546
• [Autotuner] Eliminate self._original_args clone to reduce peak memory by @yf225 in #1547
• sync bench/rebench result across ranks by @shunting314 in #1542
• Skip failing test on MTIA by @Myrthan in #1558
• [Cache] Defer some initialization in autotuner to skip unnecessary work on cache hits by @fulvius31 in #1557
• Enable Pallas backend for test_control_flow with lax.cond by @oulgen in #1554
• Add Pallas atomic ops -- everything is atomic by default by @oulgen in #1560
• Add Pallas emit_pipeline by @oulgen in #1561
• Move do_bench/sync_object from _testing to autotuner/benchmarking to fix docs build by @oulgen in #1562
• Improve launcher overhead by @oulgen in #1563
• fix mock in debug utils - hopefully make ci less flaky by @v0i0 in #1568
• fix fp8 autotuner fail for ci by @v0i0 in #1569
• fix some indexing by @v0i0 in #1453
• chore: Bump actions/upload-artifact from 6 to 7 by @dependabot[bot] in #1579
• chore: Bump actions/download-artifact from 7 to 8 by @dependabot[bot] in #1580
• Build TorchTPU in opt mode by @oulgen in #1573
• Add Pallas fori_loop with async DMA and pallas_loop_type config by @oulgen in #1574
• Enable waves_per_eu tunable on RDNA GPUs by @fulvius31 in #1571
• [cutedsl] Cross-warp reductions by @jansel in #1543
• sync seeds across ranks by @shunting314 in #1555
• [cutedsl] Add more reduction types by @jansel in #1565
• [cutedsl] Tuple reductions by @jansel in #1566
• Fix CPU tests by @jansel in #1582
• Add Pallas reduction support and enable test_reductions for TPU by @oulgen in #1549
• Tell AGENTS to not run git push by @oulgen in #1592
• partition aot tuning by config validity by @v0i0 in #1581
• bump pytorch to 2.10 in benchmark nightly to unblock grouped_gemm tritonbench by @v0i0 in #1597
• fix benchmark parsing by @v0i0 in #1596
• make sure persistent setup code uses index type for total pids by @v0i0 in #1595
• Fix issue with branching and static_range by @jansel in #1586
• Update AGENTS.md assertExpectedJournal reference by @jansel in #1587
• [cutedsl] Enable more test files by @jansel in #1583
• [cutedsl] Fix issue with sympy printer by @jansel in #1585
• [Cache] Fix cross-backend cache poisoning by adding backend to cache key by @fulvius31 in #1593
• [Helion + torch.compile] Add additional inductor fusion tests by @yf225 in #1594
• [Autotuner] Add FROM_BEST_AVAILABLE initial population strategy by @fulvius31 in #1365
• Unwrap single-element list in hl.tile() to match scalar behavior by @blake-snc in #1559
• Enable test_examples by @oulgen in #1598
• Generalize autotuning infrastructure to support Pallas/TPU backend by @oulgen in #1591
• Fix symbolic variable specialization when indexing tensor in host block by @yf225 in #1575
• Add CLAUDE.md as symlink to AGENTS.md by @oulgen in #1610
• update grouped gemm tb signature by @v0i0 in #1608
• Update logo and add events for hackathon and PLDI tutorial by @choijon5 in #1611
• Add claude skill for TPU development by @oulgen in #1612
• Enable test random and rnd by @oulgen in #1607
• [Helion + torch.compile] Temporarily skip torch.compile fusion test cases to wait for cross-repo changes by @yf225 in #1613
• add ci health check (hopefully catches cudaDeviceUnavailable) by @v0i0 in #1614
• [Helion + torch.compile] Skip test_symint_return_from_tensor_shape temporarily by @yf225 in #1615
• update rocm in benchmark now that we bumped pytorch by @v0i0 in #1609
• Remove Triton CPU backend and all CPU references by @oulgen in #1616
• [Internal CI] Use MTIA-aligned tensor shape in test_tile_single_element_list by @yf225 in #1618
• Refactor device_ir passes by @jansel in #1606
• Add @skipIfCudaSharedMemoryLessThan for failing RTX5090 tests by @jansel in #1617
• another benchmark parsing fix by @v0i0 in #1621
• Disable forking to fix failing test by @jansel in #1619
• Support for Advanced Control Files in Autotuner and Configs by @ptorru in #1576
• Fix scalar tensor indexing crash by @hinriksnaer in #1620
• [Docs] Update autotuner and TileIR backend documentation by @fulvius31 in #1600
• benchmarking all_gather_matmul by @shunting314 in #1605
• Pin H100 nightly and skip flakey pallas test by @jansel in #1624
• [CI] Relax tolerances in test_hl_arange_non_power_of_2 by @yf225 in #1627
• Decrease sizes for test_batch_softmax_block_ptr by @jansel in #1628
• fix dynamic shapes handling tensor descriptors conservatively by @v0i0 in #1604
• [Docs] Fix Helion repo link by @choijon5 in #1629
• Rename Helion Puzzles to Tutorials; fix broken tutorial examples; move to Markdown by @yf225 in #1631
• fix size hint lint from nightly by @v0i0 in #1630
• Introduce MTIA autotuning knobs by @kile01 in #1572
• change to gfx942.1 runner like pytorch, hopefully shorter queue by @v0i0 in #1632
• Set TRITON_STORE_BINARY_ONLY=1 during autotuning to reduce cache size by @fulvius31 in #1590
• Enable fast triton sigmoid by @hinriksnaer in #1564
• Fix fake impl inference with unbacked SymInts by @gmagogsfm in #1626
• [pallas-tpu] Add HALF_DTYPE constant for backend-portable half-precision tests by @v0i0 in #1636
• add gpu health check to tests as well by @v0i0 in #1634
• Properly generate floor division using // for constexpr by @PaulZhang12 in #1625
• Add --file-header option for custom headers in AOT generated files by @v0i0 in #1645
• [pallas-tpu] Bump libtpu to 0.0.37, jax/jaxlib to 0.9.1 by @v0i0 in #1641
• [pallas-tpu] Enable pallas backend for view/reshape tests in test_views.py by @v0i0 in #1638
• fix a10g lerp nightly error by @v0i0 in #1652
• [pallas-tpu] remove shape based block specs by @v0i0 in #1633
• Fix lerp decomp in _make_fx for PyTorch nightly by @v0i0 in #1653
• [pallas-tpu] Fix BlockSpec regressions from codegen BlockSpecs PR by @v0i0 in #1656
• aot tuning standalone output by @v0i0 in #1650
• [pallas-tpu] ...

v0.3.0

What's Changed

• [xpu][test]Skip Failed xpu tests by @Stonepia in #1337
• [Autotuner] Introduce Triton-TileIR Backend Support by @qelk123 in #1250
• test: removing skip condition for ROCm since tests are passing locally. by @umechand-amd in #1340
• Removing skipIfRocm property from the tests which are passing. by @umechand-amd in #1341
• Fix softmax puzzle by @nicolo-domyn in #1344
• Diversity-aware batch selection for LFBO Pattern Search by @ethche in #1350
• [xpu][fix] fix errors related to maxnreg on xpu by @jianyizh in #1347
• AOT autotuning runner and cache and heuristics gener… by @v0i0 in #1278
• handle ToFloat for symints / from non-static shapes by @v0i0 in #1357
• [Benchmark CI] Set log level to INFO to reduce log volume by @yf225 in #1360
• Support List Type kernel cache by @mengluy0125 in #1359
• Improve autotuning log readability by @hinriksnaer in #1362
• Skip tensor_descriptor and eviction policy for AMD by @mengluy0125 in #1349
• Fix pyrefly type error in surrogate pattern search by @oulgen in #1364
• Add test for torch.nn.Parameter by @jansel in #1371
• Skip tests on lower memory cards by @jansel in #1370
• [TileIR][Doc] add doc for CUDA Tile IR backend by @qelk123 in #1352
• [Docs] Put API Reference on the docs earlier by @oulgen in #1374
• Increase Top Nav options to 7 by @sekyondaMeta in #1375
• v0i0/fix is none by @v0i0 in #1369
• disable tma with int64 by @v0i0 in #1372
• fix indexer expansion by @v0i0 in #1367
• [Helion-Inductor Fusion] Add extensive unit tests for mutation/aliasing edge cases by @yf225 in #1358
• [Helion + torch.compile] Fix Dynamo error when same Helion kernel is called twice in same compile region by @yf225 in #1378
• improve signal handling for the aot runner by @v0i0 in #1366
• Use unittest.skip instead of pytest.mark.xfail by @yf225 in #1380
• [CI] Speedup tests via parallelism by @oulgen in #1338
• Support List indexing with non-literal indices by @mengluy0125 in #1377
• Throw error if TRITON_INTERPRET=1 and HELION_INTERPRET=1 are used together by @parsshar-RH in #1379
• [autotuner] filter and clone only mutated tensors by @bringlein in #1252
• Update pyrefly to 0.51.1 by @oulgen in #1386
• Update ruff to 0.15.0 by @oulgen in #1388
• [CI] Increase atol for test_hl_arange_non_power_of_2 by @yf225 in #1390
• Set an adaptive compile timeout for autotuning by @ethche in #1384
• [Helion + torch.compile] Add allow_torch_compile_fusion flag to unit tests (no-op for now) by @yf225 in #1382
• [Helion + torch.compile] Add Helion kernel Dynamo variable and HOP by @yf225 in #1383
• AOT oneshot workflow fix by @tyler-romero in #1381
• [CI] Fix NCCL error in distributed CI by @yf225 in #1396
• Fix zero-size dim handling due to upstream PyTorch changes by @yf225 in #1398
• Fix hl.rand/hl.randint crash when dim is specialized to size 1 by @karthickai in #1401
• Update num_warps and num_stages range for AMD by @yvonne-lab in #1368
• [CI] Fix B200 flaky test errors by deferring CUDA initialization to test execution time for pytest-xdist compatibility by @yf225 in #1399
• Add generic kernel config adapter by @drisspg in #1385
• Disable autocast during kernel compilation to fix mixed-precision training by @yf225 in #1402
• Add nvfp4 gemm by @oulgen in #1403
• Fix unpacking of torch.sort/topk return types in Helion kernel by @yf225 in #1406
• Skip O(N^2) argsort in torch.sort lowering when indices is unused by @yf225 in #1408
• [CI] Adjust atol for squeeze_and_excitation_net_bwd unit tests by @yf225 in #1409
• [CI] Adjust atol for test_dtype_cast_preserved_before_second_dot test by @yf225 in #1411
• Fix torch.cumsum overwriting its input when used in Helion kernel by @yf225 in #1410
• Skip hl.rand/hl.randint specialize test on MTIA by @karthickai in #1414
• Skip "too many blocks in cooperative launch" error during autotuning by @yf225 in #1417
• [CI] Fix flaky OOM error for test_int32_offset_out_of_range_error on H100 by @yf225 in #1419
• [CI] Adjust atol for test_hl_arange_non_power_of_2 test by @yf225 in #1421
• [Helion + torch.compile] Add multi-output and scalar return support by @yf225 in #1395
• Improve DeviceIR.get_root() to handle multiple reduction dimensions by @yf225 in #1434
• Introduce Backend abstraction for multi-backend code generation by @oulgen in #1426
• Route codegen decorator, constexpr, and imports through Backend by @oulgen in #1427
• Route dtype string conversions through Backend abstraction by @oulgen in #1428
• Rename triton_index_type to backend-agnostic index_type by @oulgen in #1429
• dump triton failures by @v0i0 in #1407
• Fix PointwiseLowering rank mismatch for size-1 reductions by @yf225 in #1424
• [Helion + torch.compile] Fix aliasing, mutation detection, and nested return support by @yf225 in #1432
• for now rewrite gdn fwd h without hl.grid for fusion by @v0i0 in #1391
• Add PallasBackend class and make backend configurable via Settings by @oulgen in #1447
• sync compute_repeat output for distributed kernels by @shunting314 in #1435
• raise tolerance for test_hl_dot_out_dtype_argument by @shunting314 in #1457
• Skip too many resource error in autotuner by @oulgen in #1458
• Preserve symbolic shapes in HOP output_spec during tracing by @gmagogsfm in #1454
• Refactor "start_precompile_and_check_for_hangs" by @hinriksnaer in #1422
• Add finishing phase to autotuner for config minimization by @v0i0 in #1319
• [Helion + torch.compile] Add unit test for returning SymInt from Helion kernel by @yf225 in #1459
• Workaround var_to_val deprecated warning by @jansel in #1461
• Make triton an optional dependency so Pallas backend works without triton installed by @oulgen in #1452
• Add CuteBackend boilerplate by @jansel in #1462
• Allow 'common' as fallback backend implementation by @jansel in #1463
• Refactor backend hooks for launch, program IDs, and autotune by @jansel in #1464
• Add Pallas codegen for tile, load, store and runtime launcher by @oulgen in #1448
• Add Pallas ATen lowerings, backend-aware tile strategy, and creation ops by @oulgen in #1449
• Add Pallas backend tests for add kernel by @oulgen in #1450
• Add Pallas backend to CI test matrix by @oulgen in #1451
• Add Cute pointwise add codegen, launcher, and tests by @jansel in #1465
• Wire up inductor op_overrides for the cute backend by @jansel in #1467
• Update pyproject with cute-cu12/cute-cu13 by @jansel in #1468
• Add tests for cute scalar and kwargs by @jansel in #1469
• Ignore some linter warnings by @jansel in #1470
• Skip most tests when HELION_BACKEND!=triton and add .expected_{cute,pallas} files by @jansel in #1471
• [Cache] Add bac...

v0.2.10

What's Changed

• Add examples/split_k_barrier.py to docs by @oulgen in #1320
• Add list of all events (upcoming and previous) to docs. by @choijon5 in #1321
• Fix docs event page by @choijon5 in #1323
• Print output file path by @parsshar-RH in #1322
• [example] flex attention by @v0i0 in #764
• Add static shape support to hl.rand and hl.randint by @karthickai in #1328
• Fix test_jagged_hstu_attn for PyTorch nightly silu decomposition change by @karthickai in #1329
• [CI] Add normalization for tl.full by @oulgen in #1333
• Autotune persistent kernels for multi occupancy by @oulgen in #1307

Full Changelog: v0.2.9...v0.2.10

v0.2.9

What's Changed

• Update autotuner.md with LFBO Pattern Search by @ethche in #1293
• Pin pyrefly to 0.44.0 by @oulgen in #1313
• [CI] Bump pyrefly to 0.46 by @oulgen in #1314
• [CI] Bump ruff to 0.14.10 by @oulgen in #1315
• Update Events with Helion Tutorial at PLDI 2026 by @choijon5 in #1316
• Incorporate loop iteration variable for rand by @oulgen in #1310
• Add compile time breakdown by @oulgen in #1312
• Update pyrefly to 0.46.2 by @rchen152 in #1317
• Pin pyrefly in lint.yml to 0.46.2. by @rchen152 in #1318
• settings: normalize HELION_AUTOTUNER parsing by @fulvius31 in #1308
• Support hl.barrier for mega-kernels by @jansel in #1151

Full Changelog: v0.2.8...v0.2.9

v0.2.8

What's Changed

• Support passing Triton function object to hl.triton_kernel() by @yf225 in #1263
• chore: Bump actions/download-artifact from 6 to 7 by @dependabot[bot] in #1270
• chore: Bump actions/upload-artifact from 5 to 6 by @dependabot[bot] in #1271
• [Distributed] one_shot_allreduce_bias_rmsnorm example by @yf225 in #1266
• [Distributed] matmul_reduce_scatter example by @yf225 in #1269
• feat(benchmarks): add shapes to json output by @fulvius31 in #1273
• [Autotuner] Log the 'started' state to CSV, for easier user monitoring of kernel hanging at runtime by @yf225 in #1279
• default pattern search by @v0i0 in #1259
• Set LFBOPatternSearch as default by @ethche in #1280
• fix surrogate search for singleton population by @v0i0 in #1281
• Ignore bzl files in git. by @Myrthan in #1282
• chunk fused_linear_jsd by @v0i0 in #1277
• Fix buggy interation between XYZProgramIDs and L2GroupingProgramIDs by @jansel in #1288
• Fix bug with torch.rand_like compile error by @jansel in #1289
• [autotuner] print path to generated Triton code after selection of kernel by @bringlein in #1285
• Add support for torch.gather by @jansel in #1290
• [docs] Add more example to docs by @oulgen in #1301
• [lint] remove dead ignores by @oulgen in #1302
• Add proper error handling for torch.split and torch.tensor_split in device loops by @oulgen in #1297
• Fix BlockReductionStrategy to use existing index variables for argmax/argmin operations by @oulgen in #1298
• Skip more failing tests on cpu backend by @oulgen in #1304
• [CI] Fix broken notebook by @oulgen in #1305
• Fix shape inference for tile indexing on size-1 dimensions and use broadcast_to for block_ptr by @oulgen in #1299
• Fix codegen broadcasting for tile indexing on size-1 tensor dimensions by @oulgen in #1300
• Enable tests on py314 by @oulgen in #1306

New Contributors

• @bringlein made their first contribution in #1285

Full Changelog: v0.2.7...v0.2.8