# Introduction The TVM community has worked since the last release to deliver the following new exciting improvements!
The main tags are below (**bold text is with lots of progress**): Relax etc. Please visit the full listing of commits for a complete view: [v0.24.dev0...v0.24.0.rc0](https://github.com/apache/tvm/compare/v0.24.dev0...v0.24.0.rc0). ### Community None. ### RFCs None. ### Adreno * [#18867](https://github.com/apache/tvm/pull/18867) - Revive and consolicate Adreno features ### Arith * [#19417](https://github.com/apache/tvm/pull/19417) - Expose allow_override parameter in Python Analyzer.bind() ### BugFix * [#19432](https://github.com/apache/tvm/pull/19432) - [Fix][CUDA] Version compatibility of CUDA symbols * [#19427](https://github.com/apache/tvm/pull/19427) - [FIX] Skip metal target tag registration for unsupported LLVM CPUs * [#19390](https://github.com/apache/tvm/pull/19390) - [LLVM] Fix `insertDeclare` API mismatch for ROCm-bundled LLVM 20 * [#19410](https://github.com/apache/tvm/pull/19410) - [Fix][Runtime][RPC] Fix remote tensor handle cleanup for RPC return values * [#19385](https://github.com/apache/tvm/pull/19385) - [MetaSchedule] Fix `compile_relax` to apply `MetaScheduleApplyDatabase` after `FuseOps` * [#19383](https://github.com/apache/tvm/pull/19383) - [TIRx] Fix bad-optional-access in BF16/FP8 legalize passes for target-less PrimFuncs * [#19382](https://github.com/apache/tvm/pull/19382) - [TIRx] Fix VerifyMemory crash for PrimFuncs without target attribute * [#19380](https://github.com/apache/tvm/pull/19380) - [TOPI] Fix get_const_tuple hanging indefinitely when passed a te.Tensor * [#19368](https://github.com/apache/tvm/pull/19368) - Align `tir.round` to ties-to-even across all backends * [#19367](https://github.com/apache/tvm/pull/19367) - [ONNX] Fix Round op to use ties-to-even * [#19362](https://github.com/apache/tvm/pull/19362) - [TVMScript] Fix invalid f-string format spec causing TypeError on Python 3.14 * [#19352](https://github.com/apache/tvm/pull/19352) - [TVMScript] Add `doc.keyword` handling for `ExprEvaluator._visit` * [#18957](https://github.com/apache/tvm/pull/18957) - [FIX] Inline ceil_log2 in gpu_2d_continuous_cumsum to fix MakePackedAPI error * [#18940](https://github.com/apache/tvm/pull/18940) - [Fix] Fix tvm.tir references in Tflite frontend * [#18887](https://github.com/apache/tvm/pull/18887) - [FIX] Fix cumsum kernel sblock_alloc_buffer for non-sblock buffer * [#18881](https://github.com/apache/tvm/pull/18881) - [FIX][Adreno] Replace AllocBuffer with Bind in texture alloc injection * [#18838](https://github.com/apache/tvm/pull/18838) - [TOPI] Fix resize accuracy issue with non-floor rounding * [#18782](https://github.com/apache/tvm/pull/18782) - [S-TIR][FIX] Remove redundant std::move() to itself * [#18742](https://github.com/apache/tvm/pull/18742) - [Fix] Handle empty variable name in NameSupply::FreshName * [#18694](https://github.com/apache/tvm/pull/18694) - [TIR] Fix incorrect optimization when lowering floordiv and f… * [#18695](https://github.com/apache/tvm/pull/18695) - [FIX] Fix T.sblock due to concurrent merge ### CI * [#19445](https://github.com/apache/tvm/pull/19445) - [REFACTOR] Decouple data.py from Jenkins script and docker images * [#18827](https://github.com/apache/tvm/pull/18827) - Update images to `20260301-134651-63f099ad` * [#18863](https://github.com/apache/tvm/pull/18863) - [S-TIR][Test] Mark meta_schedule tuning tests as skip * [#18851](https://github.com/apache/tvm/pull/18851) - Remove stale test scripts (i386, hexagon, mypy) * [#18850](https://github.com/apache/tvm/pull/18850) - [TEST] Remove stale URL mappings from request_hook * [#18848](https://github.com/apache/tvm/pull/18848) - Remove legacy lint scripts and Apache RAT * [#18817](https://github.com/apache/tvm/pull/18817) - [REFACTOR]Further cleanup docker images * [#18812](https://github.com/apache/tvm/pull/18812) - [REFACTOR]Modernize Python dependency management with uv * [#18809](https://github.com/apache/tvm/pull/18809) - Add GitHub Actions lint workflow * [#18805](https://github.com/apache/tvm/pull/18805) - [REFACTOR][TEST] Migrate tir-transform tests from TE to TVMScript * [#18804](https://github.com/apache/tvm/pull/18804) - [REFACTOR][TEST] Remove unused te imports from test files * [#18800](https://github.com/apache/tvm/pull/18800) - Update images to `20260219-160550-72f51851` * [#18796](https://github.com/apache/tvm/pull/18796) - Refactor Dockerfiles and installation scripts * [#18775](https://github.com/apache/tvm/pull/18775) - Update images to `20260214-152058-2a448ce4` * [#18783](https://github.com/apache/tvm/pull/18783) - Update system cuda version 12.4->12.8 * [#18780](https://github.com/apache/tvm/pull/18780) - Remove unity from tvm-bot * [#18777](https://github.com/apache/tvm/pull/18777) - Update Pillow, pytest-rerunfailures, junitparser, xgboost, onnx and pytorch * [#18647](https://github.com/apache/tvm/pull/18647) - Upgrade Python to 3.10 in CI * [#18749](https://github.com/apache/tvm/pull/18749) - Remove i386 and Hexagon from CI pipeline (2) * [#18757](https://github.com/apache/tvm/pull/18757) - Further cleanup CI after merging unity to main test * [#18456](https://github.com/apache/tvm/pull/18456) - Move conda config files to tests/conda and remove unused conda build infrastructure * [#18755](https://github.com/apache/tvm/pull/18755) - [TEST] Cleanup legacy tests and migrate unity tests to main one * [#18737](https://github.com/apache/tvm/pull/18737) - Remove i386 and Hexagon from CI pipeline (1) * [#18748](https://github.com/apache/tvm/pull/18748) - Remove i386 and hexagon from `.asf.yaml` * [#18719](https://github.com/apache/tvm/pull/18719) - [REFACTOR][TEST] Migrate all codegen test to tvmscript * [#18717](https://github.com/apache/tvm/pull/18717) - Fix double newlines in nightly docker update * [#18711](https://github.com/apache/tvm/pull/18711) - [REFACTOR][TEST] Replace CompareBeforeAfter for pytest compact * [#18692](https://github.com/apache/tvm/pull/18692) - Fix NameError in nightly docker update workflow ### Docker * [#18854](https://github.com/apache/tvm/pull/18854) - Refactor bash.sh: auto-detect rootless, add --shell, TVM_DEV_MOUNTS * [#18710](https://github.com/apache/tvm/pull/18710) - [ci]Nightly Docker image update ### Docs * [#19439](https://github.com/apache/tvm/pull/19439) - Refactor BYOC example NPU tutorial * [#19414](https://github.com/apache/tvm/pull/19414) - Fix stale tvm.tirx exclude list and add missing legalize_ops.unary entry * [#19409](https://github.com/apache/tvm/pull/19409) - Fix outdated source install and API reference docs * [#19407](https://github.com/apache/tvm/pull/19407) - Fix #18714: python -c "import tvm; print(tvm.file)" fail * [#19396](https://github.com/apache/tvm/pull/19396) - Add code generation architecture documentation * [#19398](https://github.com/apache/tvm/pull/19398) - Add TVMScript architecture documentation * [#19397](https://github.com/apache/tvm/pull/19397) - Add PyModule tutorial to How-To toctree * [#19399](https://github.com/apache/tvm/pull/19399) - Clean up architecture docs: remove duplicates, fix stale content * [#19389](https://github.com/apache/tvm/pull/19389) - Add Relax VM architecture documentation * [#19394](https://github.com/apache/tvm/pull/19394) - Add operator fusion architecture documentation * [#19395](https://github.com/apache/tvm/pull/19395) - Add BYOC external library dispatch architecture documentation * [#19387](https://github.com/apache/tvm/pull/19387) - Add docstrings for nn.Module classes and core APIs in relax.frontend.nn * [#19386](https://github.com/apache/tvm/pull/19386) - Add tvm.s_tir.tensor_intrin API reference and remove empty legacy tvm/tir directory * [#19379](https://github.com/apache/tvm/pull/19379) - Add API reference for tvm.arith, tvm.testing, tvm.exec, tvm.tirx.backend and extend topi/contrib/ir/target docs * [#19369](https://github.com/apache/tvm/pull/19369) - Add API reference for tvm.s_tir submodules: dlight, meta_schedule, backend * [#19366](https://github.com/apache/tvm/pull/19366) - Add API reference documentation for tvm.script module * [#19356](https://github.com/apache/tvm/pull/19356) - Add DLight and MetaSchedule deep-dive instructions * [#19364](https://github.com/apache/tvm/pull/19364) - TFLite tests requiring Python 3.10 and specific package versions to avoid core dumps * [#19354](https://github.com/apache/tvm/pull/19354) - Add tutorial for importing models from PyTorch, ONNX, and TFLite * [#19358](https://github.com/apache/tvm/pull/19358) - Add Dataflow Pattern Language (DPL) documentation for Relax * [#19357](https://github.com/apache/tvm/pull/19357) - Add Disco distributed runtime architecture overview * [#19351](https://github.com/apache/tvm/pull/19351) - Fix outdated paths, links, and add missing API references across documentation(3) * [#19353](https://github.com/apache/tvm/pull/19353) - Add tvm.s_tir.analysis API reference page * [#19350](https://github.com/apache/tvm/pull/19350) - Add Relax VM architecture overview in documentation * [#19344](https://github.com/apache/tvm/pull/19344) - Fix outdated code examples, typos, and missing API reference in documentation(2) * [#18965](https://github.com/apache/tvm/pull/18965) - Fix outdated code examples, types, and missing references across documentation * [#18966](https://github.com/apache/tvm/pull/18966) - [DOC] Fix various issues * [#18953](https://github.com/apache/tvm/pull/18953) - Align documentation with tirx/s_tir namespace split * [#18947](https://github.com/apache/tvm/pull/18947) - Add tutorial for mixing Python/PyTorch with TVM using BasePyModule * [#18939](https://github.com/apache/tvm/pull/18939) - [DOC] Fix inconsistent code comments * [#18941](https://github.com/apache/tvm/pull/18941) - Fix duplicate license headers and incorrect module paths after tirx rename * [#18908](https://github.com/apache/tvm/pull/18908) - Clean up stale references from recent refactors * [#18906](https://github.com/apache/tvm/pull/18906) - Update outdated references from recent refactors * [#18860](https://github.com/apache/tvm/pull/18860) - [CI]Update Sphinx dependencies * [#18855](https://github.com/apache/tvm/pull/18855) - Fix RPC tutorial to use set_input + invoke_stateful API * [#18808](https://github.com/apache/tvm/pull/18808) - [DOC] Update installation docs with missing dependencies (#18194) * [#18799](https://github.com/apache/tvm/pull/18799) - [DOC] Fix docstring, unify CMake, nvidia-docker deprecation * [#18797](https://github.com/apache/tvm/pull/18797) - [DOC] Unify CUDA naming * [#18794](https://github.com/apache/tvm/pull/18794) - [DOC] Unify GitHub naming * [#18770](https://github.com/apache/tvm/pull/18770) - [DOC] Fix PYTHONPATH in "Install from Source" * [#18732](https://github.com/apache/tvm/pull/18732) - [DOC] Fix RST syntax * [#18753](https://github.com/apache/tvm/pull/18753) - [DOC] Fix the loop length in a loop tiling example * [#18731](https://github.com/apache/tvm/pull/18731) - [DOC] Fix grammar * [#18718](https://github.com/apache/tvm/pull/18718) - Clarify trusted usage ### Frontend * [#19401](https://github.com/apache/tvm/pull/19401) - [TFLite] Add test coverage for SHAPE and RANGE operators * [#19402](https://github.com/apache/tvm/pull/19402) - [Test][TFLite] Add unit tests for `PRELU` * [#19400](https://github.com/apache/tvm/pull/19400) - [TFLite] Add TILE operator tests and edge cases * [#19388](https://github.com/apache/tvm/pull/19388) - [Test][TFLite] Add unit tests for `LEAKY_RELU`, `HARD_SWISH` `ReLU_N1_to_1` and `LOG_SOFTMAX` * [#19365](https://github.com/apache/tvm/pull/19365) - [Test][TFLite] Add unit tests for RESIZE_BILINEAR and RESIZE_NEAREST_NEIGHBOR ops * [#18970](https://github.com/apache/tvm/pull/18970) - [TFLite]Add expected IRModule checks for conv2d, pool2d, and batch_matmul tests * [#19341](https://github.com/apache/tvm/pull/19341) - [ONNX] Fix SplitToSequence keepdims=0 and uneven last chunk * [#18969](https://github.com/apache/tvm/pull/18969) - [ONNX] Support select_last_index for ArgMax and ArgMin * [#18951](https://github.com/apache/tvm/pull/18951) - [ONNX] Add MatMulInteger support to Relax ONNX frontend * [#18946](https://github.com/apache/tvm/pull/18946) - [ONNX] Add If operator support to Relax ONNX frontend * [#18929](https://github.com/apache/tvm/pull/18929) - [TFLite] Fix undefined symbols and Relay API remnants in TFLite frontend * [#18773](https://github.com/apache/tvm/pull/18773) - [ONNX] Handle Gelu approximate attribute from Opset 20 ### LLVM * [#18909](https://github.com/apache/tvm/pull/18909) - [Target]Fix -mcpu validation compatibility across LLVM versions * [#18853](https://github.com/apache/tvm/pull/18853) - Bump minimum LLVM version to 15 * [#18818](https://github.com/apache/tvm/pull/18818) - Fix build failures when building with llvm>=22 * [#18772](https://github.com/apache/tvm/pull/18772) - [Codegen] Cast NaN to bool gives true * [#18706](https://github.com/apache/tvm/pull/18706) - Fix insertDbgValueIntrinsic for Metal backend ### MetaSchedule * [#19438](https://github.com/apache/tvm/pull/19438) - [S-TIR]Make evolutionary search resilient to trace replay failures ### Metal * [#19493](https://github.com/apache/tvm/pull/19493) - Include logging headers for metal * [#18877](https://github.com/apache/tvm/pull/18877) - Batched command dispatch and staging buffer pool * [#18819](https://github.com/apache/tvm/pull/18819) - [REFACTOR]Update CHECK_LE to TVM_FFI_ICHECK_LE in Metal runtime * [#18811](https://github.com/apache/tvm/pull/18811) - [Refactor]Update ICHECK to TVM_FFI_ICHECK in Metal runtime ### ROCm * [#15518](https://github.com/apache/tvm/pull/15518) - Fix some ROCm codegen bugs ### Relax * [#19492](https://github.com/apache/tvm/pull/19492) - [BugFix]Add legalize for isnan, isinf, isfinite * [#19489](https://github.com/apache/tvm/pull/19489) - [Frontend][TFLite] Add BROADCAST_TO, EMBEDDING_LOOKUP, and SELECT_V2 * [#19490](https://github.com/apache/tvm/pull/19490) - [Frontend][TFLite] Add SCATTER_ND operator for Relax TFLite * [#19467](https://github.com/apache/tvm/pull/19467) - [ONNX] Fix CumSum axis handling: support runtime axis tensor, error on multi-element axis * [#19473](https://github.com/apache/tvm/pull/19473) - [Frontend][TFLite] Add `RANDOM_UNIFORM`, `RANDOM_STANDARD_NORMAL`, and `MULTINOMIAL` * [#19487](https://github.com/apache/tvm/pull/19487) - [Frontend][TFLite] Add BROADCAST_ARGS operator mapping * [#19481](https://github.com/apache/tvm/pull/19481) - [Frontend][TFLite] Add DILATE operator mapping * [#19485](https://github.com/apache/tvm/pull/19485) - [Frontend][TFLite] Add ATAN2 op and TFLite mapping * [#19480](https://github.com/apache/tvm/pull/19480) - [BugFix][ONNX] Fix ConstantOfShape converter when value attr is absent * [#19468](https://github.com/apache/tvm/pull/19468) - [Frontend][TFLite] Fix `STRIDED_SLICE` negative stride and add `STRIDED_SLICE/SPLIT_V` tests * [#19421](https://github.com/apache/tvm/pull/19421) - [Frontend][TFLite] Add DENSIFY operator test and fix prefetched handling * [#19464](https://github.com/apache/tvm/pull/19464) - [Frontend][TFLite] Add NON_MAX_SUPPRESSION_V4 converter * [#19466](https://github.com/apache/tvm/pull/19466) - [Frontend][TFLite] Add BITCAST operator mapping * [#19433](https://github.com/apache/tvm/pull/19433) - [Frontend][TFLite] Fix dynamic FILL/SPLIT_V partial implementations * [#19426](https://github.com/apache/tvm/pull/19426) - [Frontend][TFLite] Add soft-NMS support for TFLite NON_MAX_SUPPRESSION_V5 * [#19450](https://github.com/apache/tvm/pull/19450) - [BugFix][ONNX] Honor auto_pad in ConvTranspose converter * [#19431](https://github.com/apache/tvm/pull/19431) - [Frontend][KVCache] Extend masked sequence prefill to causal left-padding * [#19434](https://github.com/apache/tvm/pull/19434) - [Frontend][TFLite] Add CUMSUM operator mapping * [#19430](https://github.com/apache/tvm/pull/19430) - [NN] Use int64 for RoPE apply flag * [#19428](https://github.com/apache/tvm/pull/19428) - [FRONTEND][ONNX] Support Softmax, LogSoftmax and Hardmax when opset version ≤12 * [#19425](https://github.com/apache/tvm/pull/19425) - [Backend]Add NPU BYOC backend example * [#19424](https://github.com/apache/tvm/pull/19424) - Fix deprecation warning * [#19416](https://github.com/apache/tvm/pull/19416) - [TVMScript] Print ExternFunc struct_info when non-default * [#19415](https://github.com/apache/tvm/pull/19415) - [Frontend][TFLite] Fix bool `REDUCE_ANY`/`REDUCE_ALL` compile failure * [#19413](https://github.com/apache/tvm/pull/19413) - [Frontend][TFLite] Add `REDUCE_ANY` and `REDUCE_ALL` * [#19411](https://github.com/apache/tvm/pull/19411) - fix 'occured' -> 'occurred' in transform.h doc comment * [#19408](https://github.com/apache/tvm/pull/19408) - [Frontend][TFLite] Fix and test `MATRIX_DIAG`, `MATRIX_SET_DIAG`, `SPARSE_TO_DENSE` * [#19405](https://github.com/apache/tvm/pull/19405) - [Frontend][KVCache] Restructure kv_cache kernels * [#19404](https://github.com/apache/tvm/pull/19404) - [tflite] Add PRELU/LRN/SQUARED_DIFFERENCE tests (partial #18971) * [#19392](https://github.com/apache/tvm/pull/19392) - [Frontend][KVCache] Add masked sequence prefill helper for encoder valid lengths * [#19372](https://github.com/apache/tvm/pull/19372) - [frontend][tflite] Add tests for fully_connected/depthwise_conv2d/transpose_conv/l2_pool2d * [#19391](https://github.com/apache/tvm/pull/19391) - [ONNX] Add frontend support for QuantizeLinear, DequantizeLinear, and DynamicQuantizeLinear * [#19384](https://github.com/apache/tvm/pull/19384) - [BugFix]Select target-specific pipeline in tvm.compile when GPU target is provided * [#19371](https://github.com/apache/tvm/pull/19371) - [frontend][tflite] Add tests for l2_normalization/slice/reverse_v2 * [#19345](https://github.com/apache/tvm/pull/19345) - [Frontend][TFLite] Implement DETECTION_POSTPROCESS tflite operator * [#19381](https://github.com/apache/tvm/pull/19381) - [TFLite] Fix and test DEPTH_TO_SPACE/SPACE_TO_DEPTH, SELECT ops * [#19373](https://github.com/apache/tvm/pull/19373) - [TFLite] Fix `MIRROR_PAD`/`ONE_HOT` converters and add tests for `PAD`, `PADV2`, `MIRROR_PAD`, `TOPK_V2`, `ONE_HOT` * [#19370](https://github.com/apache/tvm/pull/19370) - [TFLite] Add test coverage for Reduction operations (#18971) * [#19361](https://github.com/apache/tvm/pull/19361) - [ONNX] Support ConcatFromSequenc/SequenceInsert with new_axis=1 * [#19349](https://github.com/apache/tvm/pull/19349) - [TFLite] Add NON_MAX_SUPPRESSION_V5 support * [#18963](https://github.com/apache/tvm/pull/18963) - [ONNX] Support Resize dynamic ROI via TOPI * [#18955](https://github.com/apache/tvm/pull/18955) - [ONNX] Fix shape/dynamic restrictions for `Squeeze`/`Unsqueeze` and `Slice` * [#18956](https://github.com/apache/tvm/pull/18956) - [ONNX] Complete ShapeExpr reshape handling in ONNX frontend * [#18950](https://github.com/apache/tvm/pull/18950) - [ONNX] Add Optional and MatMulInteger16 frontend support * [#18952](https://github.com/apache/tvm/pull/18952) - [ONNX] Add roi_pool op and MaxRoiPool frontend support * [#18948](https://github.com/apache/tvm/pull/18948) - Add conv3d_transpose and ONNX ConvTranspose 3D support * [#18943](https://github.com/apache/tvm/pull/18943) - [Vision] Add get_valid_counts and classic NMS * [#18942](https://github.com/apache/tvm/pull/18942) - [TOPI] Add relax.vision.multibox_transform_loc for SSD/TFLite box decode * [#18933](https://github.com/apache/tvm/pull/18933) - Add affine_grid operator with PyTorch and ONNX frontend support * [#18937](https://github.com/apache/tvm/pull/18937) - [PyTorch] Add 3D interpolate support using resize3d * [#18936](https://github.com/apache/tvm/pull/18936) - [ONNX][Torch] Add roi_align support and frontend integration * [#18931](https://github.com/apache/tvm/pull/18931) - [ONNX] Add image.resize3d op and wire 5D Resize * [#18932](https://github.com/apache/tvm/pull/18932) - [ONNX] Add GridSample ONNX frontend integration * [#18868](https://github.com/apache/tvm/pull/18868) - [TFLite] Introduce TensorFlow Lite frontend * [#18869](https://github.com/apache/tvm/pull/18869) - [LAYOUT] Support multiple axis paching * [#18904](https://github.com/apache/tvm/pull/18904) - [PyTorch] Add torch.cond support to ExportedProgram frontend * [#18870](https://github.com/apache/tvm/pull/18870) - Add input type validation for make_shape and corresponding tests * [#18903](https://github.com/apache/tvm/pull/18903) - [PyTorch] Fix crash on dynamic shapes with identity slice in ExportedProgram importer * [#18878](https://github.com/apache/tvm/pull/18878) - [ONNX] Support dynamic repeats for Tile * [#18864](https://github.com/apache/tvm/pull/18864) - [Refactor] Phase out FewShotTuning * [#18814](https://github.com/apache/tvm/pull/18814) - Make ShapeType ndim parameter mandatory * [#18815](https://github.com/apache/tvm/pull/18815) - [PyTroch] Add randn.default and randn_like.default support * [#18520](https://github.com/apache/tvm/pull/18520) - Fix llama4_rope_with_position_map to support partial rotary factor * [#18764](https://github.com/apache/tvm/pull/18764) - Add size heuristic to skip folding large creation ops * [#18762](https://github.com/apache/tvm/pull/18762) - Remove TODO comment for moving code in fuse_tir.cc * [#18733](https://github.com/apache/tvm/pull/18733) - Migrate NN conv/pooling/grad attrs from Array<IntImm> to Array<int64_t> * [#18736](https://github.com/apache/tvm/pull/18736) - Support constant folding for call_tir with tuple outputs * [#18726](https://github.com/apache/tvm/pull/18726) - [PyTorch] Simplify tensor args conversion in Dynamo * [#18725](https://github.com/apache/tvm/pull/18725) - [PyTorch] Fix scalar parameter inputs in Dynamo * [#18670](https://github.com/apache/tvm/pull/18670) - [Torch] Avoid decomposition crash with sparse CSR buffers * [#18704](https://github.com/apache/tvm/pull/18704) - [Onnx][BatchNorm] Pass momentum and training_mode into BatchNorm Operator * [#18661](https://github.com/apache/tvm/pull/18661) - [Python]Fix YaRN correction dim calculation * [#18691](https://github.com/apache/tvm/pull/18691) - [Onnx][Resize] Fix ROI values when tensor ROI is Empty pass node Constant * [#18673](https://github.com/apache/tvm/pull/18673) - [Onnx] Support Multi Input Ops with Multidirectional Broadcasting * [#18677](https://github.com/apache/tvm/pull/18677) - [NN] Add batch_flatten operator * [#18690](https://github.com/apache/tvm/pull/18690) - Add NN operator attributes include to TensorRT codegen ### Runtime * [#19476](https://github.com/apache/tvm/pull/19476) - [REFACTOR]Phase out include/tvm/runtime/object.h * [#19465](https://github.com/apache/tvm/pull/19465) - [REFACTOR][CODEGEN] Backend specific target and runtime to enable cross-compile fallback * [#19471](https://github.com/apache/tvm/pull/19471) - [REFACTOR]Phase out IntTuple alias; use ffi::Shape directly * [#19472](https://github.com/apache/tvm/pull/19472) - [REFACTOR]Phase out include/tvm/runtime/builtin_fp16.h * [#19469](https://github.com/apache/tvm/pull/19469) - [REFACTOR]Phase out include/tvm/runtime/threading_backend.h * [#19455](https://github.com/apache/tvm/pull/19455) - [REFACTOR]Phase out profiling.h heavy types, rename to timer.h * [#19457](https://github.com/apache/tvm/pull/19457) - [REFACTOR]Macro cleanup — TVM_DLL alignment, [[maybe_unused]], logging.h legacy macros * [#18837](https://github.com/apache/tvm/pull/18837) - [Builtin] Handle mismatched type on argument #0 when calling Builtin Runtime Operators * [#18813](https://github.com/apache/tvm/pull/18813) - [REFACTOR]Phase out legacy contrib runtime backends * [#18784](https://github.com/apache/tvm/pull/18784) - [REFACTOR]Transition metadata into ffi * [#18756](https://github.com/apache/tvm/pull/18756) - [COMPACT] Fix 32bit compact in vm ### TOPI * [#18880](https://github.com/apache/tvm/pull/18880) - Reject non-float inputs for inverse unary math ops ### TVMScript * [#18891](https://github.com/apache/tvm/pull/18891) - Remove T.Bind backward-compat alias * [#18889](https://github.com/apache/tvm/pull/18889) - Normalize T.Bind to T.bind for statement builder convention * [#18856](https://github.com/apache/tvm/pull/18856) - Fix PEP 563 closure variable resolution ### Vulkan * [#18914](https://github.com/apache/tvm/pull/18914) - Avoid explicit layout decoration on non-interface allocations ### web * [#18944](https://github.com/apache/tvm/pull/18944) - Update includes after FFI JSON refactor * [#18893](https://github.com/apache/tvm/pull/18893) - [Experimental] Add support for cross-origin storage caching * [#18680](https://github.com/apache/tvm/pull/18680) - [Version] Fix WebLLM vision model issues * [#18687](https://github.com/apache/tvm/pull/18687) - Handle LocalSession init in WASM RPC server ### Misc * [#19483](https://github.com/apache/tvm/pull/19483) - [REFACTOR][FFI] Cleanup ffi indirections in tvm headers + switch logging.h to ffi/error.h where only ICHECK/THROW are used * [#19484](https://github.com/apache/tvm/pull/19484) - [FFI][ABI] Bump tvm-ffi to 0.1.11rc2 * [#19477](https://github.com/apache/tvm/pull/19477) - [REFACTOR] Delete src/support/libinfo.cc; replace with runtime FFI-registry env query * [#19479](https://github.com/apache/tvm/pull/19479) - [REFACTOR][SCRIPT] TVMScript dialect-friendly refactor: per-dialect restructure + dialect registry * [#19475](https://github.com/apache/tvm/pull/19475) - [REFACTOR][S-TIR] Move tvm/support/random_engine.h → tvm/s_tir/random_engine.h * [#19474](https://github.com/apache/tvm/pull/19474) - [REFACTOR][IR] Move tvm/support/with.h → tvm/ir/with_context.h * [#19453](https://github.com/apache/tvm/pull/19453) - [S-TIR][Dlight] Add layered fall back strategy to handle missing attr `max_shared_memory_per_block` * [#19463](https://github.com/apache/tvm/pull/19463) - [REFACTOR][IR] Migrate include/tvm/node into include/tvm/ir * [#19462](https://github.com/apache/tvm/pull/19462) - [REFACTOR][NODE] Use fn_repr inside kRepr lambdas, not ffi::ReprPrint * [#19460](https://github.com/apache/tvm/pull/19460) - [REFACTOR][S-TIR] Minimize src/support/ by relocating s_tir-private headers * [#19459](https://github.com/apache/tvm/pull/19459) - [REFACTOR] Phase out src/support/ffi_testing.cc * [#19461](https://github.com/apache/tvm/pull/19461) - [REFACTOR][NODE] Migrate ReprPrinter to tvm-ffi __ffi_repr__ mechanism * [#19456](https://github.com/apache/tvm/pull/19456) - [REFACTOR] Move source_utils.h into runtime/opencl * [#19458](https://github.com/apache/tvm/pull/19458) - [REFACTOR] Phase out unreachable contrib/rust_extension.cc * [#19454](https://github.com/apache/tvm/pull/19454) - [REFACTOR][CODEGEN] Phase out tvm_global_barrier_state and tvm_prepare_global_barrier * [#19449](https://github.com/apache/tvm/pull/19449) - [REFACTOR] Use FFI types in runtime inline module-create wrapper signatures * [#18406](https://github.com/apache/tvm/pull/18406) - [TIR] Update symbolic index term order in loop fusion * [#19447](https://github.com/apache/tvm/pull/19447) - [REFACTOR] Isolate backend module creation via ffi.Module.create.<kind> registry * [#19444](https://github.com/apache/tvm/pull/19444) - [CMAKE][REFACTOR] Split libtvm.so into libtvm_runtime.so and libtvm_compiler.so * [#19440](https://github.com/apache/tvm/pull/19440) - [REFACTOR] Remove runtime/object.py shim and route Object via tvm_ffi * [#19442](https://github.com/apache/tvm/pull/19442) - [REFACTOR] Remove tvm.runtime.packed_func and container shims; route via tvm_ffi * [#19441](https://github.com/apache/tvm/pull/19441) - [REFACTOR] Phase out include/tvm/runtime/module.h * [#19393](https://github.com/apache/tvm/pull/19393) - fix: use `is None` instead of `== None` in test files (PEP 8 E711) * [#19406](https://github.com/apache/tvm/pull/19406) - [S-TIR] Fix cache_read/cache_write region when inner block has T.whe… * [#19403](https://github.com/apache/tvm/pull/19403) - [S-TIR] Fix Segfault when applying Parallel during TIR schedule rewriting * [#18927](https://github.com/apache/tvm/pull/18927) - feat(meta_schedule): expand CUDA unroll steps for SM70 optimization * [#19347](https://github.com/apache/tvm/pull/19347) - fix: TFLite model retrieval with error handling * [#19343](https://github.com/apache/tvm/pull/19343) - test(relax): cover TFLite LOG and GREATER_EQUAL in test_frontend_tflite * [#18938](https://github.com/apache/tvm/pull/18938) - [FFI] Bump tvm-ffi to 63224e3 and fix regressions * [#18912](https://github.com/apache/tvm/pull/18912) - [TIR] Handle Bind in LowerDeviceKernelLaunch * [#18926](https://github.com/apache/tvm/pull/18926) - Revert "fix: add safety warning to pickle_memoize cache loading" * [#18925](https://github.com/apache/tvm/pull/18925) - fix: add safety warning to pickle_memoize cache loading * [#18913](https://github.com/apache/tvm/pull/18913) - [Refactor] Bring up tirx namespace * [#18240](https://github.com/apache/tvm/pull/18240) - [Optimization][Operator] Implement and enable Conv2d-Reshape-Add-ReLU fusion * [#18892](https://github.com/apache/tvm/pull/18892) - [Build] Fix version regex to anchor at line start in pyproject.toml * [#18879](https://github.com/apache/tvm/pull/18879) - [TIR] Reject non-floating inputs for trig unary ops * [#18886](https://github.com/apache/tvm/pull/18886) - [TIR][REFACTOR] Revamp Common Subexpression Elimination * [#18883](https://github.com/apache/tvm/pull/18883) - [TARGET] Fix round-trip reconstruction of targets with canonicalizer-generated `feature.*` attrs * [#18876](https://github.com/apache/tvm/pull/18876) - [REFACTOR][TIR] Remove body from AllocBuffer and DeclBuffer * [#18871](https://github.com/apache/tvm/pull/18871) - Batched GPU dispatch and object caching for WebGPU runtime * [#18875](https://github.com/apache/tvm/pull/18875) - [chore] Update docker/README.md documentation and fix links * [#18873](https://github.com/apache/tvm/pull/18873) - [TIR] Add VisitBufferDef/VisitBufferUse to base StmtVisitor/StmtMutator * [#18865](https://github.com/apache/tvm/pull/18865) - [REFACTOR][TIR] Introduce AllocBuffer and phase out Allocate+DeclBuffer * [#18862](https://github.com/apache/tvm/pull/18862) - [REFACTOR][TIR] Cleanup AttrStmt attributes * [#18857](https://github.com/apache/tvm/pull/18857) - [TIR][Refactor] Enhance error reporting with structured AssertStmt and TVMFFIABIBuilder * [#18861](https://github.com/apache/tvm/pull/18861) - fix: Complete CHECK update across contrib runtime * [#18859](https://github.com/apache/tvm/pull/18859) - fix: Use T.decl_buffer instead of T.Buffer for aliased buffers in LongRoPE * [#18858](https://github.com/apache/tvm/pull/18858) - fix: Complete ICHECK update across codebase * [#18845](https://github.com/apache/tvm/pull/18845) - [REFACTOR][CONTRIB] Remove MSC contrib module * [#18847](https://github.com/apache/tvm/pull/18847) - [PYTHON] Fix PEP 563 compat and remove args_converter * [#18843](https://github.com/apache/tvm/pull/18843) - [TIR][FEAT] Require DeclBuffer before use in verify_well_formed * [#18852](https://github.com/apache/tvm/pull/18852) - [REFACTOR] Remove unused mscclpp contrib module * [#18849](https://github.com/apache/tvm/pull/18849) - [CMAKE] Remove unused Libbacktrace.cmake * [#18844](https://github.com/apache/tvm/pull/18844) - [REFACTOR] Further cleanup node redirections * [#18830](https://github.com/apache/tvm/pull/18830) - [LINT][PYTHON] Modernize annotations with ruff UP rules * [#18832](https://github.com/apache/tvm/pull/18832) - [IR][TIR] Remove body from AssertStmt * [#18829](https://github.com/apache/tvm/pull/18829) - [REFACTOR][NODE] Remove node redirect headers * [#18825](https://github.com/apache/tvm/pull/18825) - [REFACTOR] Update CHECK and ICHECK_GE to TVM_FFI_ICHECK and TVM_FFI_ICHECK_GE in thrust.cu * [#18828](https://github.com/apache/tvm/pull/18828) - [REFACTOR] Phase out root Makefile * [#18821](https://github.com/apache/tvm/pull/18821) - fix: replace 6 bare except clauses with except Exception * [#18822](https://github.com/apache/tvm/pull/18822) - [TARGET] Specify correct `mcpu` for Metal target tags * [#18816](https://github.com/apache/tvm/pull/18816) - [REFACTOR][S-TIR] Lift STIR-only attributes out of tir::attr namespace * [#18810](https://github.com/apache/tvm/pull/18810) - [REFACTOR][LINT] Modernize ruff config * [#18801](https://github.com/apache/tvm/pull/18801) - Bump tvm-ffi to v0.1.9rc * [#18807](https://github.com/apache/tvm/pull/18807) - [LINT] Modernize lint to use pre-commit hooks * [#18803](https://github.com/apache/tvm/pull/18803) - [REFACTOR] Migrate CHECK macros to tvm-ffi ones * [#18768](https://github.com/apache/tvm/pull/18768) - support integer types in fast_tanh and fast_exp * [#18802](https://github.com/apache/tvm/pull/18802) - [FFI] Bring up latest tvm-ffi * [#18793](https://github.com/apache/tvm/pull/18793) - [REFACTOR][TARGET] Further cleanup target python api * [#18785](https://github.com/apache/tvm/pull/18785) - [REFACTOR][TARGET] Phase out legacy target string in favor of json * [#18786](https://github.com/apache/tvm/pull/18786) - [CONTRIB] Cache the shape and dtype array in json access * [#18781](https://github.com/apache/tvm/pull/18781) - [chore] cleanup unsed legacy backtrac code in logging * [#18779](https://github.com/apache/tvm/pull/18779) - [REFACTOR] Phase out dmlc dep * [#18776](https://github.com/apache/tvm/pull/18776) - [REFACTOR][S-TIR] More migrations to s-tir * [#18771](https://github.com/apache/tvm/pull/18771) - [REFACTOR][S-TIR] Migrate more transform to s_tir * [#18763](https://github.com/apache/tvm/pull/18763) - [REFACTOR][TIR] Phaseout BufferRealize * [#18759](https://github.com/apache/tvm/pull/18759) - [REFACTOR] Remove picojson dependency, replace with tvm::ffi::json API * [#18760](https://github.com/apache/tvm/pull/18760) - [chore] Cleanup stale dependencies * [#18761](https://github.com/apache/tvm/pull/18761) - [REFATOR][TIR] Phase out AllocConst * [#18758](https://github.com/apache/tvm/pull/18758) - [Cleanup] Remove redundant python/pyproject.toml and gen_requirements * [#18697](https://github.com/apache/tvm/pull/18697) - Fix Customize Optimization tutorial import error #18584 * [#18754](https://github.com/apache/tvm/pull/18754) - [REFACTOR][S-TIR] Cleanup items on block scope * [#18743](https://github.com/apache/tvm/pull/18743) - [REFACTOR][S-TIR] Move remaining data structures to s_tir * [#18739](https://github.com/apache/tvm/pull/18739) - fix: correct typos 'recieve' and 'occurence' * [#18744](https://github.com/apache/tvm/pull/18744) - fix: skip dsymutil for static tvm_runtime on Apple platforms * [#18740](https://github.com/apache/tvm/pull/18740) - fix: correct typo 'occuring' to 'occurring' * [#18738](https://github.com/apache/tvm/pull/18738) - [chore][TIR] reorganize src/tir/transforms to src/tir/transform * [#18734](https://github.com/apache/tvm/pull/18734) - [REFACTOR][S-TIR] Lift dlight into s_tir namespace * [#18735](https://github.com/apache/tvm/pull/18735) - [REFACTOR][S-TIR] Migrate meta_schedule into s_tir namespace * [#18705](https://github.com/apache/tvm/pull/18705) - Add Windows-specific build notes to installation guide * [#18727](https://github.com/apache/tvm/pull/18727) - fix: correct typos in Python docstrings * [#18728](https://github.com/apache/tvm/pull/18728) - [REFACTOR][S-TIR] Migrate tir/schedule to s_tir * [#18722](https://github.com/apache/tvm/pull/18722) - [REFACTOR][S-TIR] Lift transform passes to s_tir namespace * [#18724](https://github.com/apache/tvm/pull/18724) - Remove cron schedule from nightly Docker update workflow * [#18716](https://github.com/apache/tvm/pull/18716) - [REFACTOR] Migrate old tir.ir_builder to tvmscript or builder * [#18715](https://github.com/apache/tvm/pull/18715) - [SPIRV] Fix forloop codegen in vulkan * [#18712](https://github.com/apache/tvm/pull/18712) - [REFACTOR][S-TIR] Initialize the s_tir module * [#18636](https://github.com/apache/tvm/pull/18636) - [TIR][Schedule]Generalize fuseReductionEpilogue to support arbitrary epilogue expressions * [#18699](https://github.com/apache/tvm/pull/18699) - [TIR] Further robustify floordiv/mod intrin lowering to prevent overflow * [#18689](https://github.com/apache/tvm/pull/18689) - [REFACTOR][TIR] Rename tir.Block to SBlock * [#18671](https://github.com/apache/tvm/pull/18671) - [TIR] Fix InjectPTXLDG32 segfaults and skip non-CUDA targets -- View it on GitHub: https://github.com/apache/tvm/releases/tag/v0.24.0.rc0 You are receiving this because you are subscribed to this thread. Message ID: <apache/tvm/releases/[email protected]>
