v0.1.1

What's Changed

• [Benchmark] Avoid using _run in TritonBench integration by @yf225 in #444
• Add H100 CI by @oulgen in #435
• Add B200 CI by @oulgen in #436
• Skip illegal memory access for autotuning by @oulgen in #453
• Re-enable associative_scan tests in ref eager mode by @yf225 in #443
• Fix tritonbench integration issue by @yf225 in #463
• [Benchmark] Allow passing kwargs; Set static_shape = True for better benchmark perf by @yf225 in #465
• [Example] One shot all reduce by @joydddd in #245
• Fix lint by @oulgen in #469
• Improve signal/wait doc by @joydddd in #478
• Cleanup ci by @oulgen in #449
• Run CI on mi325x by @oulgen in #441
• Improve Stacktensor Doc by @joydddd in #479
• Require tests to be faster than 30s by @oulgen in #471
• Improve error message when no good config is found by @oulgen in #455
• Add SequenceType Eq comparison by @oulgen in #482
• [Benchmark] Add try-catch for tritonbench import path by @yf225 in #487
• Add helion prefix to Triton kernel name by @yf225 in #486
• Support GraphModule inputs by @jansel in #488
• Improve stack trace for #457 by @jansel in #489
• [EZ] Replace pytorch-labs with meta-pytorch by @ZainRizvi in #490
• [generate_ast] providing AST args, and fall back to api._codegen when output is a tuple by @HanGuo97 in #481
• Support reshape with block_size expressions by @yf225 in #495
• [example] add jagged_softmax example by @pianpwk in #480
• Fix handling of fixed size reductions by @jansel in #499
• Improve error message for rank mismatch in control flow by @jansel in #502
• Fix reshape + sum case by @yf225 in #504
• Sort config keys alphabetically in __str__ by @yf225 in #505
• Fix issue with fp64 constants by @jansel in #506

New Contributors

• @ZainRizvi made their first contribution in #490
• @HanGuo97 made their first contribution in #481
• @pianpwk made their first contribution in #480

Full Changelog: v0.1.0...v0.1.1

v0.1.0

What's Changed

• Ref-eager and normal modes can share cache by @oulgen in #421
• Helion examples by @sekyondaMeta in #401
• Add extensive setter/getter unit tests for indexed tensor; fix bugs discovered by new tests by @yf225 in #422
• Always set triton allocator by @jansel in #416
• Add stacked tensor by @joydddd in #346
• Change references to pytorch-labs to pytorch by @oulgen in #430
• [BC breaking] Add StackTensor support to hl.signal & hl.wait (as_ptrs) by @joydddd in #261
• Fix: test/test_signal_pad by @joydddd in #432
• Fix test/test_stack_tensor.py by @oulgen in #431
• Skip associative_scan tests in ref eager mode by @yf225 in #433
• Fix scalar value assignment to tensor slices by @yf225 in #424
• Fix scalar tensor broadcasting in type propagation by @yf225 in #425
• Fix strided slice support for static slices (e.g., buf[::2]) by @yf225 in #426
• Better fix for triton allocator error by @jansel in #427
• Make bullet points clickable by @sekyondaMeta in #428

New Contributors

• @sekyondaMeta made their first contribution in #401

Full Changelog: v0.0.12...v0.1.0

v0.0.12

What's Changed

• Use autotuner's BoundKernel in caching by @oulgen in #388
• Temporarily move triton_key import to inner function to unblock older torch versions by @oulgen in #395
• [Examples] Add matmul variants with bias support and tests by @yf225 in #379
• [Benchmark] Support kernel variants; setup matmul tritonbench integration by @yf225 in #380
• Relax tolerance for test_input_float16_acc_float16_dynamic_shape by @yf225 in #399
• [Benchmark] Enable CSV output; clean up benchmark hot path by @yf225 in #398
• [Benchmark] Move per-operator settings from example file to benchmarks/run.py by @yf225 in #403
• [Ref Mode] PyTorch reference mode (eager only) by @yf225 in #339
• Clean up CI & fix caching by @oulgen in #408
• Add autotuner_fn argument to @helion.kernel for custom autotuners by @oulgen in #394
• Fix non-tuple indexing warning by @jansel in #411
• Add support for listcomp by @jansel in #412
• Fix allow_tf32 warning by @jansel in #413
• Update URL to helionlang.com by @jansel in #414
• Add metaclass [] syntax for cache classes by @jansel in #415
• [Ref Mode] Expand ref eager mode support to more hl.* APIs (e.g. load / store / scan / reduce) by @yf225 in #410
• [Benchmark] Fix arg parsing issue in tritonbench integration by @yf225 in #417

Full Changelog: v0.0.11...v0.0.12

v0.0.11

What's Changed

• Add tl._experimental_make_tensor_descriptor restrictions by @oulgen in #331
• Skip accuracy check for test_moe_matmul_ogs by @yf225 in #333
• Do not create a new variable for tile assignments since tiles are immutable by @oulgen in #334
• Clean pyright warning by @oulgen in #335
• Run lint with nightly (match test CI) by @oulgen in #349
• Refactor BoundKernel in memory caching by @oulgen in #351
• Allow string literal args by @jansel in #353
• Fix issue with integer in rolled reduction by @jansel in #354
• Fix test_fp8_attention expected by @jansel in #355
• Write test workflow without pt deps by @oulgen in #352
• Refresh the action cache once a month by @oulgen in #362
• Use bare nvidia cuda docker image by @oulgen in #363
• Inline install triton by @oulgen in #364
• Remove triton's conda deps by @oulgen in #365
• Make lint workflow leaner by @oulgen in #366
• Swap from conda to uv for lint workflow by @oulgen in #367
• Swap from conda to uv on test workflow by @oulgen in #368
• Fix fp16 var_mean multi-output issue by @jansel in #357
• Add fallbacks for unary ops that don't support fp16 by @jansel in #361
• Name the cache step so we can check its outputs by @oulgen in #369
• Fix pyright errors by @oulgen in #370
• Only use Tensor Descriptor indexing with appropriate shapes by @PaulZhang12 in #360
• Remerge LayerNorm (#348) by @PaulZhang12 in #373
• Do not crash autotuner on more triton/llvm/cuda errors seen on B200 by @oulgen in #374
• Set MAX_JOBS=4 for tritonbench build to avoid OOM by @yf225 in #376
• [Benchmark] Allow running a specific shard of input via --input-shard M/N cli arg by @yf225 in #377
• Use venv for pip install on lint by @oulgen in #381
• [RFC] Implement basic on disk caching by @oulgen in #336
• Add hl.dot() API; Use hl.dot instead of torch.matmul for FP8 GEMM ops in Helion kernel by @yf225 in #356
• Fix test_inline_asm_packed expected output due to upstream PyTorch change by @yf225 in #385

Full Changelog: v0.0.10...v0.0.11

v0.0.10

What's Changed

• [Benchmark] Add initial TritonBench integration and vector_add benchmark example by @yf225 in #247
• Add static_range by @joydddd in #235
• Cleanup/improve docstrings by @jansel in #250
• [Benchmark] Add embedding benchmark by @yf225 in #248
• [Benchmark] Add vector_exp benchmark by @yf225 in #249
• Add rms_norm example and test by @yf225 in #252
• [Benchmark] Add rms_norm benchmark by @yf225 in #253
• Strip extra newlines from *.expected files by @jansel in #255
• Fix issue with BLOCK_SIZE0.to(torch.int32) by @jansel in #254
• Add hl.wait & AllGather Matmul example (via hl_ext helper). by @joydddd in #189
• Add sum example and test by @yf225 in #256
• [Benchmark] Add sum to TritonBench integration by @yf225 in #257
• Rename benchmark folder by @yf225 in #258
• Add hl.signal by @joydddd in #233
• Add hl.wait for simultenous waiting for multiple gmem barriers by @joydddd in #243
• Swap to using pyright by @oulgen in #259
• Fix pyright errors in type_propagation.py by @yf225 in #266
• [BE] Add spellchecker by @oulgen in #265
• Remove pyre-ignore/pyre-fixme calls by @jansel in #274
• Improve typing for helion.kernel by @jansel in #270
• Add jagged_mean example by @yf225 in #263
• [Benchmark] Add jagged_mean tritonbench integration by @yf225 in #264
• Add fp8_gemm example and test by @yf225 in #267
• [Benchmark] Add fp8_gemm to TritonBench integration by @yf225 in #268
• Fix some pyright errors by @jansel in #276
• Remove unused exception types by @jansel in #271
• Fix docstring see also lists by @jansel in #272
• [benchmarks] Change tritonbench api by @xuzhao9 in #260
• Initial versison of documentation by @jansel in #273
• Deploy docs to github pages by @jansel in #277
• Fix lint error on main by @jansel in #281
• Add a link to the documentation by @jansel in #282
• [Benchmark] Fix tritonbench integration due to upstream changes by @yf225 in #278
• [Benchmark] Allow using 'python benchmarks/run.py' to run all kernels by @yf225 in #280
• Add implicit broadcasting tests by @jansel in #285
• Add additional tl.range choices to persistent loop by @jansel in #287
• Update autotuning example in docs by @jansel in #288
• Add host side dead code elimination by @oulgen in #289
• [Benchmark] Add attention tritonbench integration by @yf225 in #284
• Add helion.exc.CannotModifyHostVariableOnDevice and helion.exc.CannotReadDeviceVariableOnHost by @jansel in #290
• Fix unstable CI by @jansel in #299
• Make to_triton_code config arg optional by @jansel in #291
• Add helion.exc.DeviceTensorSubscriptAssignmentNotAllowed by @jansel in #292
• Remove default configs from examples by @jansel in #295
• Fix bug with tensor descriptor and small block size by @jansel in #296
• Relax typing for CombineFunction by @jansel in #297
• Add examples/segment_reduction.py by @jansel in #300
• Add error for using a host tensor directly by @jansel in #306
• Improve Tensor.item() handling by @jansel in #307
• Fix type_info null errors by @oulgen in #294
• Improve DCE by marking math functions as pure by @oulgen in #312
• [Benchmark] Add softmax tritonbench integration by @yf225 in #286
• Make imports relative by @jansel in #310
• Generalize l2_grouping to support 3+ dimensions by @jansel in #313
• Remove make_precompiler generated wrapper by @jansel in #314
• Enforce ANN/PGH lints by @jansel in #315
• Support dynamic fill value to hl.full by @jansel in #316
• Use tensor device reference in persistent kernels by @jansel in #317
• Add tl._experimental_make_tensor_descriptor support by @oulgen in #322
• Fix variable scoping in nested loops for multi-pass kernels by @yf225 in #324
• Add HELION_DEV_LOW_VRAM env var for low GPU memory machines by @yf225 in #325
• Add cross_entropy example and unit test by @yf225 in #320
• [Benchmark] Add cross_entropy to tritonbench integration by @yf225 in #321
• Add literal index into tuple by @joydddd in #327
• Improve naming for generated helper functions by @jansel in #323
• Add hl.inline_asm_elementwise by @jansel in #328
• Implement static tuple unrolling and hl.static_range by @jansel in #329
• Add fp8_attention example and unit test by @yf225 in #318
• [Benchmark] Add fp8_attention to tritonbench integration by @yf225 in #319

New Contributors

• @xuzhao9 made their first contribution in #260

Full Changelog: v0.0.9...v0.0.10

v0.0.9

What's Changed

• Add tl.range warp_specialize to autotuner by @jansel in #230
• Switch from TensorDescriptor to tl.make_tensor_descriptor by @jansel in #232
• Enable Test fixed by Fixed by #195 by @joydddd in #236
• Implement persistent kernels by @jansel in #238
• Add hl.associative_scan by @jansel in #239
• Fix failing tests on main by @jansel in #244
• Add hl.reduce by @jansel in #240
• Switch from expecttest/assertExpectedInline to assertExpectedJournal by @jansel in #241

Full Changelog: v0.0.8...v0.0.9

v0.0.8

What's Changed

• Improve loop end bound optimization for nested tiling by @jansel in #192
• Set default dot_precision to TRITON_F32_DEFAULT by @jansel in #197
• Use _disable_flatten_get_tile helper in tile_id by @jansel in #200
• Throw type errors immediately by @jansel in #202
• Fix typo in LiteralType.merge by @jansel in #201
• Add support for global statements in type propagation by @jansel in #203
• Remove ErrorReporting class and simplify warning handling by @jansel in #204
• Add InvalidDeviceForLoop exception type by @jansel in #205
• Fix bug with renamed variable flowing into phi() node by @jansel in #206
• Move hl.grid tests to their own file by @jansel in #208
• Remove NDGridTileStrategy by @jansel in #209
• Simplify codegen for hl.grid by @jansel in #210
• Add support for hl.grid(begin, end, step) by @jansel in #211
• Support range() loops (alias for hl.grid) by @jansel in #212
• Move yz_grid disabling logic to ConfigSpec by @jansel in #213
• Relax chebyshev kernel test tolerance by @jansel in #214
• [RFC] Add static loop unrolling by @oulgen in #216
• Add support for torch.arange by @jansel in #215
• Fix a performance issue with Helion-emitted Flash Attention by @manman-ren in #181
• Fix issue with phi nodes and aliasing by @jansel in #220
• Fix duplicate argument handling in inductor lowering by @jansel in #222
• x[i] returns scalar when i=scalar by @joydddd in #223
• Fix config flatten spec for tile.id by @joydddd in #224
• Fix failing tests on main by @jansel in #231
• Refactor examples to use run_example helper by @jansel in #225
• Add tl.range loop_unroll_factor to autotuner by @jansel in #226
• Add tl.range num_stages to autotuner by @jansel in #227
• Add tl.range disallow_acc_multi_buffer to autotuner by @jansel in #228
• Add tl.range flatten to autotuner by @jansel in #229

New Contributors

• @manman-ren made their first contribution in #181

Full Changelog: v0.0.7...v0.0.8

v0.0.7

What's Changed

• Fix bug with computations based on hl.register_block_size by @jansel in #157
• Generalize workaround for unbacked size hints by @jansel in #159
• Don't hardcode cuda in test files by @jansel in #160
• Move register_block_size/register_reduction_dim to tunable_ops.py by @jansel in #161
• Unskip some previosly failing tests by @jansel in #162
• Use workflow matrix to deduplicate code by @oulgen in #168
• Rename TileIndexProxy to hl.Tile by @jansel in #171
• Fix block size variable handling and atomic operations with symints by @jansel in #177
• Codegen if tl.sum(one_elem_tensor): instead of if one_elem_tensor by @yf225 in #158
• Fix visitCall in deviceIR. Always visit argument nodes by @joydddd in #180
• Relax bounds on test_mask_dot by @oulgen in #182
• Add lowering for Constant assignment by @joydddd in #187
• Expose tile.id by @joydddd in #188
• Do not precompile set configs by @oulgen in #183
• Add option to ban/disallow autotuning by @oulgen in #184
• Recommend PyTorch nightly build in readme by @jansel in #193
• Fix issue with ConfigSpec mutation in codegen by @jansel in #195
• enable_python_dispatcher() in propagate_types by @laithsakka in #191

New Contributors

• @laithsakka made their first contribution in #191

Full Changelog: v0.0.6...v0.0.7

v0.0.6

What's Changed

• Fix ast read writes by @oulgen in #148
• Update pre-commit by @oulgen in #149
• Try enable test_moe_matmul_ogs on CI by @yf225 in #147
• [Ready for review] Add support for print(prefix_str, *tensors) by @yf225 in #140
• Support hl.tile_{begin,end,block_size} by @jansel in #150
• Rename TileStrategy.get_block_index to CompileEnvironment.get_block_id by @jansel in #151
• Fix bug in merging sequence types by @jansel in #152
• Increase atol for test_matmul_split_k by @jansel in #155
• Fix bug in test_matmul_split_k by @jansel in #156
• Add hl.register_tunable by @jansel in #154

Full Changelog: v0.0.5...v0.0.6

v0.0.5

What's Changed

• Rename linter/check_main.py -> scripts/lint_examples_main.py by @jansel in #124
• Improve error message for unpacking a tile by @jansel in #125
• Improve error message for overpacked tiles by @jansel in #126
• [BC breaking] Simplify block size configs by @jansel in #127
• Refactor reduction loop config spec by @jansel in #128
• Move BlockIdSequence to its own file by @jansel in #129
• Do not print output code durring autotuning by @jansel in #130
• Make helion.exc.TensorOperationInWrapper not fire on non-torch ops by @jansel in #131
• Add HELION_FORCE_AUTOTUNE=1 and update readme by @jansel in #132
• Correct units for time printouts by @jansel in #133
• Rename block_size_idx to block_id by @jansel in #134
• Rename block_indices to block_ids by @jansel in #135
• Add Pyre Pre-Commit Hook by @lolpack in #136
• Update .pre-commit-config.yaml by @oulgen in #137
• [Ready for review] Add hl.register_reduction_dim(); add support for matmul+layernorm example by @yf225 in #80
• Fix bug with errors on unreachable if branch by @jansel in #138
• [Error Message] Update block config size length mismatch by @drisspg in #139
• Increase atol/rtol for test_error_in_non_taken_branch by @jansel in #142
• Fix some typos by @jansel in #141
• More fair comparison by @drisspg in #146

New Contributors

• @lolpack made their first contribution in #136

Full Changelog: v0.0.4...v0.0.5