Introduction

The TVM community has worked since the v0.13.0 release to deliver the following new exciting improvements! The main tags are below (bold text is with lots of progress):

• Community, RFC
• Arith, MetaSchedule
• Adreno, ArmComputeLibrary, Hexagon, Metal, OpenCL & CLML, ROCm, Vulkan, cuda & cutlass & tensorrt, micoNPU, web
• Runtime, TVMC, AOT, LLVM, microTVM, CMSIS-NN
• Frontend, Relay, BYOC
• TOPI, TIR, TVMScript
• Docs, CI, Docker
• Misc, , BugFix

Please visit the full listing of commits for a complete view: v0.13.0...v0.14.0.

Community

• #15307 - Qingchao Shen -> Reviewer
• #15619 - community strategy decision process

RFC

• #102 - [Process RFC] Clarify Community Strategy Decision Process

AOT

• #15301 - Avoid call_extern() with incorrect argument count
• #15181 - Remove workaround to help resolve test flakiness

Adreno

• #15830 - Minor changes for Adreno docs and help scripts
• #15671 - [VM]Fix using buffers for weights in VM
• #15391 - Small fixes in Adreno schedules

Arith

• #15881 - Simplify the result of non-divisible floordiv
• #15665 - Fix detect non-divisible iteration form like (x % 255) // 16
• #15638 - MLIR PresburgerSet compile fix mlir >= 160
• #15628 - Added simplification rule for multiple equality compares
• #15558 - Fix detect linear equation with uint var
• #14690 - Add tvm::arith::PresburgerSetNode to work with Presburger Set in MLIR
• #15555 - Fix handling of overlapping predicates
• #15471 - Enhance Canonical Simplify for LE
• #15228 - Enhance buffer shape bound deduction to include offset

ArmComputeLibrary

• #15600 - [ACL] Update Compute Library to v23.05.1
• #15344 - [ACL] Update Compute Library to v23.05

BugFix

• #15891 - [Relay]fix axis parsing of repeat converter in the MXNet frontend
• #15873 - [Fix] Remove duplicated words from comments, NFC
• #15868 - [Relay]Fix conv transpose with default strides in ONNX frontend
• #15773 - [CPP] Fix cpp deploy bug
• #15778 - [Hotfix] Fix Windows Pipe
• #15748 - Move symbols that are relevant to the runtime from libtvm to…
• #15752 - [Relay]fix the wrong calculate logic of operator flip in PyTorch frontend
• #15715 - [Relay]Fix the wrong implementation about operator Threshold in oneflow
• #15711 - [Strategy] Fix arm_cpu int8 conv2d strategy for dotprod and i8mm targets
• #15717 - [Relay]fix the wrong implementation of Softplus in OneFlow
• #15677 - [Arith] IterMapRewriter abort rewriting once failure
• #15629 - [VTA] tvm.tir.Call has no name attribute
• #15584 - [Relay][Strategy] Enable compile time transformation of weights matrix for arm_cpu NHWC quantized conv2d
• #15542 - [Fix] Fix the typo in compile flag
• #15484 - [TOPI] Fix a bug in arm_cpu int8 conv2d i8mm schedule
• #15473 - [Relay] Fix some bugs of dominator pattern
• #15478 - [TIR] ThreadSync with shared.dyn awareness
• #15406 - [TIR]Ensure the Var's scope is correct
• #15399 - [TIR] Fix multi-grouped multi-warp allreduce
• #15350 - [Relay] fix a bug of printing dataflow pattern
• #15385 - Work around "Internal Compiler Error" in MSVC
• #15294 - [Bug][Relay] fix relay frontend pytorch op addmm bug
• #15323 - [Fix][TIR] LowerThreadAllreduce with correct thread mask
• #15291 - [Relay][GraphExecutor] Fix set_input_zero_copy() precision bug
• #15225 - Fix function to read all file

CI

• #15903 - [Target]Add LLVM functions for current system info
• #15897 - [ADRENO] Few updates to Adreno docker setup
• #15836 - Update ci-gpu image
• #15668 - Allow Limit CPUs in Docker
• #15568 - [Testing] Allow Capitalized name in CompareBeforeAfter
• #15519 - [TEST] Run tests/python/relay/aot tests in ci-cortexm
• #15485 - Remove cython version pin
• #15421 - Bump Flax and Jaxlib versions to fix Jaxlib install error
• #15226 - Add ml_dypes dependency for all docker images
• #15353 - Pin cython version to fix cython compilation
• #15352 - Make Graviton3 default AArch64 job runner node
• #15339 - Update test to include unique attribute
• #15277 - [Testing] Return BenchmarkResult in local_run and rpc_run
• #15268 - [Testing] Add tvm.testing.local_run
• #15136 - [UnitTest][NVPTX] Avoid cascading failures from CUDA postproc

CMSIS-NN

• #15747 - Move CMSIS_5 from SHA to release based upgrade
• #15407 - Support for Softmax Int16 operator

Docker

• #15799 - Add LLVM 17 to the LLVM install script
• #15862 - Upgrade oneflow to v0.8.0
• #15819 - Install oneflow from PyPi
• #15310 - Update ci-cortexm docker image
• #15293 - tensorflow_aarch64 package upgrade

Docs

• #15619 - community strategy decision process
• #15508 - Add v0.13.0 docs to site
• #15213 - [#15157][Rust][Doc] Re-enable the Rust documentation build

Frontend

• #15821 - [TFLite]Support quantized ELU
• #15844 - [TFLite]Fix test failures caused by div-by-zero
• #15798 - [TFLite]Support quantized Pow
• #15829 - [Relay][Keras][Bugfix] fix the converters of GRU and SimpleRNN about the go_backwards attribute
• #15838 - Fix unnecessary pylint errors
• #15802 - [SkipCI][Hotfix][TFLite] Disable test of quantized floor mod
• #15790 - [TFLite]Support quantized LESS_EQUAL
• #15775 - [TFLite]Support quantized GREATER_EQUAL
• #15769 - [TFLite]Support quantized NOT_EQUAL
• #15768 - [TFLite]Support quantized div
• #15746 - [TFLite]Support quantized LESS
• #15733 - [TFLite]Support quantized floor_mod
• #15724 - [TFLite]Support quantized floor_div
• #15602 - [ONNX][BugFix] Support If body with free variable from graph input
• #15472 - [Relay][TFLite] Fix in qnn.conv2d when parameter groups not equal to 1
• #15117 - [TFLITE] Add support for TFLite's regular NMS operator
• #15415 - [ONNX] add onnx Mish operator
• #15422 - [Keras] Add support for swish actiivation
• #15370 - [Relay][Pytorch] Add aten::view_as
• #15335 - [Bugfix][Keras] Add a check to reject the invalid input shape
• #15334 - [Bugfix][Relay][Keras] Add a assertion to reject a invalid value for attribute units in RNN layers
• #15337 - [Bugfix][Keras]Fix a corner case bug in softmax converter of keras frontend
• #15259 - [TFLITE][BugFix] Fix variable typo in batchmatmul converting func
• #15261 - [bugfix][keras] Fix go_backwards attribute of LSTM in keras frontend

Hexagon

• #15788 - Properly handle RPC server shutdown
• #15599 - F2qi avgpool bug fix
• #15414 - Add default vtcm capacity for targets
• #15367 - Simplify Mul->Sub->Conv to Conv->Add when possible
• #15258 - Propagate QNN Concat Quantization Params to Inputs

LLVM

• #15921 - Fix for llvm CodeGenOpt API change

MetaSchedule

• #15792 - Allow generating uint random data
• #15574 - Fix metaschedule flop estimation for non-integer loop dimensions
• #15532 - Enable subprocess to stdout for DEBUG level
• #15437 - Fix mma default rule and disable tuning abort
• #15133 - [XGBoost,MetaSchedule] Support xgb set tree method

Metal

• #15756 - [Unittest]Add minimal metal functionality test to CI
• #15749 - [UnitTest]Parametrize allreduce GPU tests
• #15401 - [Codegen]Support metal warp-level primitive

OpenCL & CLML

• #15745 - [OpenCL] Don't initialize OpenCL runtime on host
• #15400 - [VM][OpenCL] Introduce textures allocation to VM memory manager

ROCm

• #15777 - [Codegen]Mismatched Dtype of Workgroup/Workitem
• #15464 - fma intrin
• #15454 - Fix some ROCm codegen bugs

Relay

• #15889 - fix the conflicted documentation description
• #15648 - [TOPI] Remove input padding for arm_cpu conv2d int8 native schedule in Legalize pass
• #15386 - Fix an adaptive_max_pool1d operator conversion bug
• #15533 - Disable exception for ADT in mixed precision pass
• #15506 - [Strategy] Use x86 pool schedules for arm_cpu
• #15470 - [Strategy] Use x86 dense schedules for arm_cpu
• #15392 - add redirecting operation to dataflow pattern graph
• #15468 - [Strategy] Fix arm_cpu int8 conv2d schedule selection for 32-bit targets
• #15461 - Stop ToMixedPrecision when constant is out of dtype range
• #15362 - improve SimplifyClipAndConsecutiveCast pass
• #15137 - Introduce arguments limit to FuseOps pass
• #15211 - Fix bug in MergeCompilerRegions pass
• #15237 - ExprMutator Return Origin Expr When All Fields Isn't Changed
• #15235 - [QNN] Support Dequantize to "float16" and Quantize to "uint16"

Runtime

• #15693 - Make CSourceModule and StaticLibraryModule Binary Serializable
• #15658 - Make export_library parameters after file_name keyword-only
• #15637 - [Backport]Fix ICE from Clang
• #15244 - Serialization/Deserialization of runtime module
• #15630 - Utils to Stringify Device
• #15623 - Expose ModuleGetFunction as PackedFunc
• #15595 - Enhance PackedFunc Metaprogramming with PackArgs
• #15543 - [Minor] Suppress verbose logging in Metal device API
• #15305 - Flush L2 cache in time eval
• #15332 - Device API to query L2 cache size

TIR

• #15913 - Fix offset_factor in cuda tensor core intrins
• #15906 - Fix the error example in the documentation for pad_einsum
• #15816 - Revert "[TensorIR][Visitor] Visit buffer members in match_buffer's in block visitor functions (#15153)
• #15763 - Do not drop 4th argument to tir.max
• #15646 - Output DeclBuffer in LowerThreadAllreduce
• #15493 - Output DeclBuffer in SplitHostDevice
• #15517 - Shuffle in PointerValueTypeRewrite for scalar reads
• #15263 - Output DeclBuffer in MakePackedAPI
• #15465 - [TIR, Schedule] Fix decompose reduction with thread binding loops
• #15432 - Generalize implementation of T.macro to work with other dialects
• #15413 - Fix Primitive Rfactor DType
• #15404 - Allow starred expressions in TIR script
• #15374 - Finer predicate handling in cross-thread reduction
• #15373 - Allreduce broadcast result to each thread in multi-warp case
• #15214 - [UX] Implement privacy annotations in TIR
• #15241 - Return error code from kernels in SplitHostDevice
• #15327 - ThreadAllreduce warp-level primitive support with multi-warp
• #15260 - Implement TIR macros
• #15253 - Call TVMBackendFreeWorkspace inside LetStmt
• #15264 - Allow symbolic bounds in IndexMap analysis
• #15243 - Output DeclBuffer in LowerTVMBuiltin
• #15236 - [Schedule] Scoped CacheRead/Write producing compact region
• #15242 - Preserve AllocateNode::annotations
• #15247 - Allow VerifyWellFormed to accept IRModule
• #15192 - Support cross-threaad reduction lowering with thread-broadcasting rewrite
• #15210 - [Schedule] Derive Nonnegative Bounds from Shape Var
• #15207 - [Transform] Add LiftThreadBinding Pass

TOPI

• #15685 - [Target]Use LLVM for x86 CPU feature lookup
• #15710 - Ensure vectorization of input padding in arm_cpu int8 conv2d interleaved schedule
• #15513 - check empty array of x86 injective's iters
• #15371 - Revert "Add arm_cpu specific pooling schedules"
• #15311 - Add arm_cpu specific pooling schedules
• #15286 - Revert "Add arm_cpu specific pooling schedules"
• #14855 - Add arm_cpu specific pooling schedules

TVMC

• #15779 - enable dumping imported modules too
• #15349 - Add tvmc flag to print compilation time per pass

TVMScript

• #15824 - Preserve traceback across TVMScript parsing
• #15762 - Use environment variable TVM_BLACK_FORMAT for .show()
• #15706 - Disable black_format by default
• #15705 - [FIX] Disable show_object_address in printing by default
• #15579 - Optionally output the address as part of variable names
• #15564 - Use triple-quoted python strings for metadata
• #15547 - Create loop var with min_val dtype in for frame
• #15492 - Allow use of Python builtins in script
• #15442 - Support starred indices in for-loop
• #15249 - Ensure completed root block has no read/write
• #15239 - Handle parsing of PrimFunc calls with non-void return

cuda & cutlass & tensorrt

• #15573 - [CUTLASS][Cherry-pick] Introduce several features of cutlass profiler
• #15480 - [Bugfix][CUTLASS] CUTLASS path finding

micoNPU

• #15780 - [microNPU][ETHOSU] MatMul legalization support
• #15428 - [microNPU][ETHOSU] Fix concatenation with reused buffers
• #14909 - [ETHOSU][MicroNPU][Pass] Add a pass to replicate pads
• #15186 - [microNPU][ETHOSU] Add Vela's logic to select configuration block

microTVM

• #15667 - Check the output of microNPU demos in CI

web

• #15218 - Increase default EMCC compilation total memory size

Misc

• #15934 - [Release] [Dont Squash] Modify version number to 0.14.0 and 0.15.0.dev on main branch
• #15934 - [Release] [Dont Squash] Modify version number to 0.14.0 and 0.15.0.dev on main branch
• #15847 - [release] Update version to 0.14.0 and 0.15.0.dev on main branch
• #15867 - Bump pillow from 9.3.0 to 10.0.1 in /apps/microtvm/ethosu
• #15866 - Bump pillow from 9.3.0 to 10.0.1 in /apps/microtvm/cmsisnn
• #15865 - Bump pillow from 9.2.0 to 10.0.1 in /apps/microtvm
• #15833 - [VM] Memory Manager moved up to runtime
• #15859 - [Script] Fix miscs of make_notes.py
• #15818 - [CLI TOOLS][RTVM] Improve rtvm tool with new options to measure native performance
• #15761 - [Target] LLVM helper functions for any target info
• #15672 - [IR] Implemented Variant<...> container
• #15714 - [Target][Device] Auto detect target and create device from str in torch style
• #15723 - fix _convert_simple_rnn
• #15725 - Revert "[CodeGenC] Handle GlobalVar callee as internal function call"
• #15684 - [Hopper TMA] Add intrinsic to create barriers for synchronization
• #15683 - Fix a bug caused by PyTorch instance_norm when the input shape is [1,1,1,2]
• #15596 - [FFI] Propagate Python errors across FFI boundaries
• #15666 - [Module] Implement custom imported modules serialization
• #15656 - [Hopper TMA] Add CUDA codegen support for bulk asynchronous copy
• #15664 - [IR] Use structural equal for Range equality
• #15649 - Add output_data_sec section in corstone300.ld
• #15639 - Do not link LLVM libraries into cpptest binary
• #15631 - [RPC] Enhance RPC Protocol to support TVM Object
• #15624 - [CMake] Add RCCL to TVM and TVM Runtime
• #15616 - [Hopper TMA] CUDA codegen for async copy with barrier synchronization
• #15537 - [CPP_RPC] export listdir for RPC
• #15605 - [CMake] Add NCCL to TVM and TVM Runtime
• #15580 - Fix "to" duplicate word in python and C header file
• #15581 - Remove duplicate load word inside .cc file
• #15582 - Remove duplicate 'from' word inside python script
• #15554 - Bump tornado from 6.1 to 6.3.3 in /apps/microtvm
• #15552 - Bump tornado from 6.1 to 6.3.3 in /apps/microtvm/ethosu
• #15553 - Bump tornado from 6.1 to 6.3.3 in /apps/microtvm/cmsisnn
• #15536 - fixed typo [TypoFix]
• #15529 - [quantize] fix bug of annotate for output of add op
• #15535 - Fixed search task comment
• #15530 - Remove duplicate msg word and condition inside the function doc
• #15511 - Remove IRModule Dependency from Target
• #15525 - Fix typo mistake and change whethe to whether
• #15524 - Remove duplicate the word
• #15103 - [CodeGenC] Handle GlobalVar callee as internal function call
• #15419 - [VM][Textures] Enable OpenCL textures for VM
• #15483 - [Script] Be more careful when generating ast.ExtSlice for Subscript
• #15469 - [CYTHON] Make cython compatible with 3.0
• #15423 - [Submodule] Add Flash attention v2
• #15380 - [Target] Add Jetson Orin Nano tag
• #15359 - [CMAKE] Conditionally link "clog" in NNPack install
• #15326 - [OP] Add rms_norm into TOPI
• #15312 - [skipci] Fix typo in docs/arch/index.rst
• #15298 - [Release] Extend PR tags and Format PR hyper-links in release report
• #15328 - [Package] Remove cutlass media/docs inside cutlass_fpA_intB_gemm
• #15321 - [JVM] Fix the Maven pom.xml for OS X arm64 tvm4j build
• #15265 - Fix keras version problem
• #15292 - [RPC] Fix socket bind errno on corner case
• #15287 - [Exec] Add a script to test GPU memory bandwidth
• #15234 - [Miscs] Enhance script about make release notes
• #15229 - [CMAKE] Add Vulkan header for Android
• #15215 - [Android] ndk static build
• #15208 - Update version to 0.14.dev0 on main branch