v0.25.0

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, Frontend, TIR, Runtime, etc.

Please visit the full listing of commits for a complete view: v0.24.0...v0.25.0.

Community

None.

RFCs

None.

Arith

• #19604 - [REFACTOR][TIR]Phase out ControlFlowGraph, NarrowPredicateExpression, and rename Simplify to StmtSimplify
• #19638 - [REFACTOR]Phase out arith/scalable_expression; arith no longer proves over scalable vectors
• #19670 - Memoize IntervalSet variable relaxation to avoid exponential blowup
• #19669 - Gate canonical-simplify LT Case 2 on extra scale == +1
• #19675 - Make Analyzer a tvm-ffi Object

BugFix

• #19502 - [TIR] Skip bool-typed expressions in CSE
• #19497 - [Relax] Fix scatter_elements and scatter_nd CUDA compilation
• #19498 - [Relax][ONNX] Resolve param Vars in Concat to handle mixed Shape/Tensor inputs
• #19511 - [Relax][Torch] Honor multi-axis dims in torch.flip converter
• #19512 - [Relax][Torch] Honor correction in std/var converter
• #19514 - [S-TIR] Wrap bare scalar bodies in DefaultGPUSchedule to avoid root-block crash
• #19527 - [Relax]: handle ONNX ScatterElements reduction
• #19535 - [Fix][Relax]: ONNX Clip NaN bounds and preserve input NaN (ORT parity)
• #19554 - [Fix][CI]: remove astral-sh/setup-uv from lint workflow
• #19557 - [Fix][Relax] Lower bool prod as logical all
• #19567 - [Target][LLVM] Use libm for asin/acos instead of buggy inline Taylor
• #19568 - [Target][LLVM] Route sinh/cosh/atan/asinh/erf through libm extern
• #19619 - [Vulkan][CodeGen] Change OpControlBarrier to AcquireRelease
• #19643 - [Fix] Stabilize layer_norm variance computation with two-pass reduction
• #19650 - [Fix][Relax] Support ND batched matmul chains in AdjustMatmulOrder pass
• #19683 - [Fix] CommReduce could handle 0-dim data
• #19779 - [Fix] nn.attention support dynamic batch_size
• #19808 - [Fix] Revert C++20-only lambda captures for C++17 build

CI

• #19629 - Remove tvm-lint from tvm-bot
• #19656 - Add cibw-based wheel publishing to PyPI
• #19659 - Wheel publishing follow-ups
• #19665 - Derive the version from Git tags via setuptools_scm
• #19664 - Reformat the macOS repair-wheel-command as a multiline script
• #19697 - Target apache-tvm for PyPI wheel publishing
• #19775 - Merge PR against its target branch instead of main (#19712)
• #19685 - Remove PyPI-only tag ref guard from wheel publishing
• #19703 - Pin actions by version tag, trim wheel perms
• #19706 - [Tests] Fix s_tir tests using removed T.block API in TIRx script
• #19700 - Fix release verification script
• #19704 - [Tests] Skip test modules cleanly when optional deps are missing
• #19713 - Fix CI script test subprocess environment
• #19724 - [Tests][Disco] Skip CCL tests when runtime support is absent
• #19725 - [Tests][Relax] Gate multi-GPU VM test on three devices
• #19726 - [Tests][Hexagon] Lazily import pytest plugin dependencies
• #19730 - [Tests][NNAPI] Skip tests cleanly when remote environment is unavailable
• #19729 - [Tests][S-TIR] Fix stale MetaSchedule sketch expectations and migrate let binds to T.let
• #19715 - [Tests] Remove test_runtime_ndarray (covered by tvm-ffi)
• #19731 - [Script][Tests] Fix dialect redirect module re-execution and stray category-less tirx.intrin_test op
• #19735 - [S-TIR][Tests] Fix transform test failures after TIRx bringup
• #19740 - [Tests] Check WebGPU volatile allreduce annotation structurally
• #19746 - [Tests] Fix flaky popen pool executor test
• #19738 - Align cuda-python with PyTorch cuda-bindings
• #19745 - [Tests][LLVM] Gate stepvector intrinsic rename on LLVM 20
• #19751 - [S-TIR][Tests] Mark test_cp_async_in_if_then_else as xfail
• #19737 - Run s_tir/transform tests in the python-unittest stage
• #19754 - Updated cibw to 4.1.0
• #19752 - [Tests][AArch64] Make SVE codegen assertions robust across LLVM versions
• #19761 - Drop redundant cmake/ninja install from the Linux wheel CUDA sidecar
• #19777 - [Tests] Modernize test gating
• #19786 - [Tests] Make TargetCreation.DeduplicateKeys host-agnostic on AArch64
• #19787 - [Tests] Replace remaining requires_* helpers with standard pytest
• #19793 - Pin GitHub Actions to SHA for ASF INFRA compliance
• #19798 - Remove Jenkins PR linter step
• #19800 - [Tests][Refactor] Remove unused testing helpers

Docs

• #19606 - Reorganize development guide content
• #19720 - Clarify loading serialized artifacts requires a trusted source
• #19782 - [CI] Bump tlcpack-sphinx-addon to restore search result summaries
• #19788 - Modernize test-gating documentation

Frontend

• #19590 - [ONNX] Add RMSNormalization converter for ONNX opset 23

Hexagon

• #19747 - [Tests] Clean up stale hexagon tests
• #19796 - [REFACTOR]Phase out Hexagon app and test wrappers

LLVM

• #19716 - [Codegen]Accept splat form in VLA broadcast test
• #19744 - [Codegen][Tests] Gate +v9a vscale_range expectation on LLVM version

Relax

• #19495 - [Frontend] Add ParameterList and ParameterDict containers
• #19491 - [Frontend][TFLite] Add segment operator mappings
• #19499 - [Frontend][TFLite] Add tests coverage for SPACE_TO_BATCH_ND and BATCH_TO_SPACE_ND
• #19516 - [TFLite] Add gather frontend expected IRModule tests
• #19488 - [PyTorch] Fix segfault in from_exported_program when model uses index_put_ with tuple output
• #19523 - [Frontend][TFLite] Add Conv3D support
• #19525 - [ONNX] Normalize negative indices before the take call for Gather operator
• #19530 - [Frontend] Add TFLite Frontend Support for CONV_3D_TRANSPOSE
• #19536 - [Frontend][TFLite] Add initial StableHLO builtin operator support
• #19547 - [ONNX] Set max_output_boxes_per_class default value to 0 for NonMaxSuppression
• #19515 - [ONNX] Add ONNX Backend Tests for systematic frontend coverage
• #19566 - [ONNX] Prevent Div divide-by-zero crashes
• #19573 - [ONNX] Fix TopK scalar K extraction in from_onnx
• #19587 - [Frontend][TFLite] Support StableHLO region-based ops and multi-subgraph models
• #19588 - Normalize negative concat axis in ReorderPermuteDimsAfterConcat
• #19603 - [REFACTOR]Fold CalleeCollector into relax DeadCodeElimination
• #19538 - [Frontend][TFLite] Support quantized TFLite import via QDQ decomposition
• #19616 - [Frontend][TFLite] Support control-flow multi-subgraph operators
• #19601 - [Frontend][TFLite] Add UNIDIRECTIONAL_SEQUENCE_RNN converter
• #19637 - [Frontend][TFLite] Add REDUCE_WINDOW support
• #19632 - [Frontend][TFLite] Add RNN converter
• #19633 - [Frontend][TFLite] Add LSTM and SVDF converter
• #19639 - [Frontend][TFLite] Add TFLite Resource Variable and Static Hashtable Import Support
• #19634 - [Frontend][TFLite] Support sequence LSTM and RNN operators
• #19646 - [Frontend][TFLite] Support STABLEHLO_WHILE
• #19644 - [IR] Skip in-place multiply when two operands are views of the same tensor
• #19649 - [Frontend][TFLite] Support STABLEHLO_CUSTOM_CALL
• #19654 - [Frontend][TFLite] Add HASHTABLE_LOOKUP converter
• #19651 - [Frontend][TFLite] Support STABLEHLO_RNG_BIT_GENERATOR
• #19645 - [PyTorch] Cast non-bool inputs to bool in logical_not converter
• #19652 - [Frontend][TFLite] Add EMBEDDING_LOOKUP_SPARSE converter
• #19660 - [PyTorch] Decompose integer pow into repeated multiplication
• #19626 - [ONNX] Fix Cast operator float->int NaN/Inf handling
• #19674 - [ONNX] Preserve NaN in Sign to align with ONNX Runtime
• #19679 - [PyTorch] Cast non-bool inputs to bool in logical_and converter
• #19711 - [CoreML] Fix CoreML partition pass
• #19732 - [PyTorch][DLight] Fix exported-program CUDA test failures
• #19756 - [PyTorch] Add logical_or and logical_xor converters
• #19772 - [ONNX] Fix LayerNormalization no-bias zero tensor shape and dtype
• #19773 - [ONNX] Support exclusive option in CumSum
• #19755 - [ONNX] Make ReduceMax/ReduceMin NaN propagation order-independent(numpy semantics)
• #19789 - [TensorRT] Update TensorRT runtime to 10
• #19763 - [Frontend][TFLite] Add support for FFT/complex operators: REAL, IMAG, COMPLEX_ABS

Runtime

• #19617 - [CMAKE]Link tvm_rpc with all backend runtime libraries
• #19620 - [REFACTOR]Phase out tvm::runtime::regex_match
• #19622 - [REFACTOR]Remove leftover microTVM/CRT crumbs
• #19621 - [REFACTOR]Relocate nvtx.h to tvm/support/cuda and make it header-only
• #19628 - [REFACTOR]Structural reorganization: locality moves for thread_map, texture, minrpc, disco, contrib
• #19714 - [Tests] Fix contrib wheel tests
• #19736 - [Disco] Fix session attribute storage, NVSHMEM build, and test gating
• #19748 - [Tests] Drop int4 from random_fill test, fix dtype error message
• #19762 - [CoreML] Fix FFI casts in CoreML runtime

TIR

• #19581 - [TIRx] Bringup TIRx Infrastructure
• #19642 - [TIRx] Fix stale Simplify import in lowering test
• #19657 - [TIRx] Post-bringup op-dispatch / codegen / TVMScript follow-ups
• #19663 - [REFACTOR][TIRX] Consolidate split host device stages
• #19677 - [TIRx] Update scoped ops and CUDA launch bounds
• #19728 - [TIRx] Preserve Triton call_kernel compile options
• #19739 - [TIRx] Use canonical PTX async script API in s_tir test
• #19753 - [TIRX][Tests] Fix LLVM version gate for vectorized lround
• #19757 - [TIRx] Post-bringup follow-ups: op-dispatch, namespaces, launch bounds, gemm-async, backend reorg
• #19785 - [TIRX][CUDA] Framework support for FA4, CLC intrinsics, and nvfp4 tcgen05 GEMM
• #19776 - [TIRx][RISC-V] Use scalable RVV loops for fixed vectorize
• #19797 - [REFACTOR][TIRX] Add IntImm common scalar ctor and streamline MakeConst

TVMScript

• #19583 - Handle undefined functions when dumping IRModule

cuda & cutlass & tensorrt

• #19565 - [RFC][CodeGen][CUDA]: Gate fast math intrinsic lowering behind target option
• #19596 - [CodeGen][CUDA] Move fast math intrinsic lowering option to PassContext
• #19741 - [S-TIR][CUDA] Fix legacy predicated cp.async zero fill
• #19768 - [REFACTOR][CUDA] Phase out l2 cache flush preproc test
• #19770 - [REFACTOR][CUDA] Phase out cuda_common.h
• #19784 - [CUDA] Narrow the cuda extra from cuda-python to cuda-bindings

web

• #19494 - Add support for OPFS
• #19569 - [COS] Persist URL→hash mapping across page loads
• #19673 - Add support for OPFS synchronous access handles and committed records
• #19687 - Bump tvmjs version to 0.25.0-dev1
• #19790 - Destroy GPUDevice once on buffer creation error
• #19780 - use singular requestFileHandle() instead of requestFileHandles()

Misc

• #19446 - [release][Dont Squash] Update version to 0.24.0 and 0.25.0.dev on main branch
• #19528 - [REFACTOR][IR] Remove dead AttrFunctor template
• #19423 - [TIR] Add cooperative_tensor builtins and metal.cooperative_tensor storage scope
• #19539 - [Contrib] Fix CUDA contrib build after FFI/header cleanups
• #19594 - [BUILD] Modularize device runtime into per-backend DSOs
• #19586 - [RPC][Tracker] Bound msg_size to MAX_TRACKER_MSG_BYTES to prevent unbounded buffer growth
• #19597 - [IR] Add annotations to Call nodes
• #19602 - Fix PytestUnknownMarkWarning: Unknown pytest.mark.adreno_clml
• #19607 - [REFACTOR][IR] Cleanup attrs.h: drop NullValue, AttrsNodeReflAdapter, legacy BaseAttrsNode methods
• #19611 - [REFACTOR] Move src/ir/script_printer.cc to src/script/printer/
• #19613 - [REFACTOR][IR] Phase out src/ir/structural_{hash,equal}.cc to tvm-ffi
• #19612 - [REFACTOR][IR] Inline ApplyPassToFunction into relax decompose_ops, delete the util
• #19614 - [REFACTOR][IR] Phase out class Integer and class Bool in Attrs and PassConfig
• #19615 - [REFACTOR][IR] attrs.h follow-up cleanup: drop legacy vtable / rename / phase out AttrFieldInfo
• #19605 - [REFACTOR][TIR] Tie AnnotateDeviceRegions/SplitHostDevice/LowerDeviceKernelLaunch together
• #19618 - [IR] Rename Call annotations to attrs
• #19624 - [REFACTOR][PYTHON] Lift compiler/CLI/process modules from tvm.contrib to tvm.support
• #19627 - [REFACTOR][IR][FFI] Bump tvm-ffi (+ SEqHashDef migration) and phase out tvm/ir/repr.h
• #19625 - [REFACTOR][IR] Inline ReplaceGlobalVars into AttachGlobalSymbol
• #19630 - [REFACTOR][PYTHON] Consolidate derived_object into tvm.ir.utils
• #19631 - [REFACTOR][SCRIPT] tvmscript streamline: lift printer.h, restore one-way dep, migrate dialect config to extra_config
• #19636 - [REFACTOR][IR] Delete class Bool and class Integer boxed-type wrappers
• #19653 - [REFACTOR][PYTHON] Revisit lifted support modules from tvm.contrib
• #19648 - fix: Security Patch: Fix missing exported flag in AndroidManifest
• #19658 - [RPC] Import tvm.testing lazily in rpc.testing
• #19662 - [FFI][IR] Route JSON serialization through tvm-ffi
• #19661 - [FFI][REFACTOR] Direct structural APIs to tvm-ffi
• #19681 - [Bump] tvm-ffi to 59da4c0
• #19684 - [RELEASE] Bump web npm version to 0.25.0
• #19701 - [Python] Bump apache-tvm-ffi floor to >=0.1.12 on v0.25.0
• #19709 - [Refactor][Meta-schedule] Remove meta-schedule as_string mechanism in favor of default representation
• #19719 - [REFACTOR][PYTHON] Slim tvm.libinfo to info-only helpers
• #19717 - [Codegen][NVPTX] Skip runtime execution in Vulkan codegen tests
• #19721 - [REFACTOR][PYTHON] Remove tvm.ffi shim; import tvm_ffi directly
• #19722 - [REFACTOR][IR] Phase out diagnostic.h for visit-context-aware pass errors
• #19723 - [Python] Refactor pyproject.toml dependencies
• #19727 - [PYTHON] Autoload backends; simplify library loading; remove TVMError for native errors
• #19742 - [S-TIR] Fix software pipeline offsets for legacy MMA intrinsics
• #19758 - [REFACTOR][VM] Move CUDA graph VM builtin back under VM runtime
• #19760 - [REFACTOR][DataType] Phase out target custom datatype support
• #19759 - [REFACTOR][TARGET] Cleanup backend target registration
• #19767 - [MetaScheduler] Improve print info about builder/runner state
• #19778 - [CPP_RPC] Bugfix race conditions and enhance print infos
• #19734 - [CMAKE] Upgrade TVM build baseline to C++20
• #19769 - [REFACTOR][PYTHON] Consolidate backend autoload infra
• #19781 - [REFACTOR][IR] Cleanup IR naming utilities
• #19783 - [AGENT] Migrate agent instructions to vendor-neutral layout
• #19794 - [REFACTOR] Phase out unused queue and rang license entries
• #19799 - [REFACTOR][IR] Simplify CallingConv attribute access
• #19805 - [CMAKE] Revert build baseline to C++17