[VOTE] Release Apache TVM v0.25.0.rc0

[MasterJH5574]

Dear TVM community,

This is a call for vote to release Apache TVM version v0.25.0.

Release highlights:

• Deeper tvm-ffi integration: the arithmetic Analyzer is now a tvm-ffi Object, structural APIs and IRModule JSON serialization are routed directly through tvm-ffi, and related FFI/header cleanups across the codebase.
• TIRx infrastructure bring-up: initial TIRx infra with op-dispatch / codegen / TVMScript follow-ups and updated scoped ops and CUDA launch bounds.
• Major Relax TFLite frontend expansion: LSTM / RNN / SVDF, Conv3D and CONV_3D_TRANSPOSE, REDUCE_WINDOW, quantized import via QDQ decomposition, resource variables / static hashtables, embedding-lookup-sparse, segment ops, and StableHLO + control-flow multi-subgraph support.
• Relax ONNX correctness and ONNX Runtime parity: NaN/Inf handling for Sign / Cast / Clip / Div, RMSNormalization (opset 23), TopK scalar-K, Gather negative indices, ScatterElements reduction, NonMaxSuppression defaults, plus systematic ONNX backend tests.
• Relax PyTorch fixes: segfault in from_exported_program with index_put_, correction in std/var, multi-axis flip, logical_not, and integer pow decomposition.
• TIR / codegen: cooperative_tensor builtins with a metal.cooperative_tensor storage scope; CUDA fast-math intrinsic lowering gated behind a target/PassContext option; LLVM transcendental functions (sinh/cosh/atan/asin/acos/erf) routed through libm; Vulkan barrier fix.
• Arithmetic robustness: memoized IntervalSet variable relaxation to avoid exponential blowup, and a canonical-simplify correctness fix.
• Runtime / build: device runtime modularized into per-backend DSOs.
• WebGPU/WASM: OPFS synchronous access handles and committed records, and persistent URL→hash mapping across page loads.
• Security: fix missing exported flag in AndroidManifest; bound RPC tracker message size.

Link to release candidate:
https://dist.apache.org/repos/dist/dev/tvm/tvm-v0.25.0-rc0/

Link to GitHub pre-release page (full release notes):
https://github.com/apache/tvm/releases/tag/v0.25.0.rc0

Pre-release wheels are also published on PyPI for testing:
pip install --pre apache-tvm==0.25.0rc0

Everyone is welcomed to vote. Please vote by replying to this thread explicitly. According to release procedure, for the dev@ vote, there must be at least 3 binding +1 votes and more +1 votes than -1 votes.

+1 = approve
+0 = no opinion
-1 = disapprove (provide reason)

NOTE: this thread is being mirrored in dev@

This vote will remain open for at least 72 hours.

[MasterJH5574]

Verified the v0.25.0.rc0 candidate two ways — (A) building from the signed source release, and (B) installing the published wheel from PyPI. Kept separate since they are different installation paths.

A. Source build check — tests/scripts/release/test_release_package.sh

Run with the script as fixed in #19700. Exact script used:
https://github.com/MasterJH5574/tvm/blob/tvm-dev/2026-06-09-release-check/tests/scripts/release/test_release_package.sh

$ bash test_release_package.sh

+ version=v0.25.0
+ rc=rc0
[setup] Create an isolated virtualenv ...
+ python3 -m venv .venv
...
IGNORE
...
[1/10] Downloading from apache.org ...
...
IGNORE
...
[2/10] Downloading from github.com ...
...
IGNORE
...
[3/10] Check difference between github.com and apache.org ...
+ diff github/md5sum.txt ./apache/md5sum.txt
[4/10] Checking asc ...
+ gpg --verify ./apache-tvm-src-v0.25.0.rc0.tar.gz.asc ./apache-tvm-src-v0.25.0.rc0.tar.gz
gpg:                using RSA key 3EA0F45C3223CF0D81C8F2CCA9F8369106601185
gpg: Good signature from "Ruihang Lai <ruihangl@apache.org>" [unknown]
[5/10] Checking sha512 ...
+ sha512sum -c ./apache-tvm-src-v0.25.0.rc0.tar.gz.sha512
apache-tvm-src-v0.25.0.rc0.tar.gz: OK
[6/10] Unzip ...
+ tar -zxf apache-tvm-src-v0.25.0.rc0.tar.gz
[7/10] Checking whether binary in source code ...
[8/10] Compile and Python Import on Linux ...
+ cmake ..
+ make -j16
...
IGNORE
...
[100%] Built target tvm_objs
[100%] Linking CXX shared library lib/libtvm_compiler.so
[100%] Built target tvm_compiler
[9/10] Install Python runtime dependencies (declared in the source pyproject.toml) ...
Installing declared runtime deps: apache-tvm-ffi>=0.1.12 ml_dtypes numpy psutil typing_extensions
Successfully installed apache-tvm-ffi-0.1.12 ...
[10/10] Import TVM and print path ...
+ python3 -c 'import tvm; print(tvm.__path__)'
['.../github/apache-tvm-src-v0.25.0/python/tvm']

LGTM — md5 (github vs apache.org) identical, GPG signature good (verified against the dev KEYS), sha512 OK, no binaries in source, full compile, and import tvm succeeds.

B. Wheel install check — PyPI

All four platform wheels published for apache-tvm==0.25.0rc0 (macOS arm64, Linux x86_64 + aarch64 manylinux_2_28, Windows AMD64). Installed the Linux x86_64 wheel in a clean venv — dependencies (including apache-tvm-ffi) resolve automatically, no source build:

$ python3 -m venv venv && . venv/bin/activate
$ pip install apache-tvm==0.25.0rc0
...
Successfully installed apache-tvm-0.25.0rc0 apache-tvm-ffi-0.1.12 ...
$ python -c "import tvm; print(tvm.__version__)"
0.25.0rc0

LGTM — wheel installs with stable dependencies and imports at the expected version. (It is a pre-release, so a bare pip install apache-tvm still resolves the latest stable; testers opt into the RC by pinning ==0.25.0rc0 or using --pre.)

[tqchen]

+1

[tlopex]

+1

[yongwww]

+1

[Ubospica]

+1