This is an automated email from the ASF dual-hosted git repository.
tqchen pushed a commit to branch main
in repository https://gitbox.apache.org/repos/asf/tvm-ffi.git
The following commit(s) were added to refs/heads/main by this push:
new 0932956f [CI] Enable Windows ARM64 wheels (#691)
0932956f is described below
commit 0932956fbd5595125e51606698c1bf9864e39541
Author: vortex-captain <[email protected]>
AuthorDate: Mon Oct 12 20:16:41 2026 +0800
[CI] Enable Windows ARM64 wheels (#691)
## Summary
- add a native Windows ARM64 job to the `tvm-ffi` wheel publishing
matrix
- enable the ARM64 target in cibuildwheel on Windows
- select the MSVC developer environment from the running Python
architecture so JIT extensions compile for ARM64 rather than x64
- install the Windows ARM64 PyTorch nightly from NVIDIA before running
wheel tests
- enable prerelease resolution and bypass the upload-date cutoff for
that target because the NVIDIA nightly index does not expose upload
timestamps
Following maintainer feedback, this PR enables and tests the core
`tvm-ffi` Windows ARM64 wheel without changing the separate
`torch-c-dlpack-ext` wheel workflow.
## Validation
- Windows ARM64 publish job: 465/465 C++ tests passed
- CPython 3.12 ARM64: 2414 passed, 44 skipped, 3 xfailed
- CPython 3.13 ARM64: 2416 passed, 42 skipped, 3 xfailed
- CPython 3.14t ARM64: 2418 passed, 40 skipped, 3 xfailed
- Windows AMD64, Linux x86_64/ARM64, and macOS ARM64 wheel jobs passed
- synchronized the branch with current `main`
- installed `torch 2.15.0.dev20260904+cu134` from NVIDIA’s nightly index
on native Windows ARM64
- Torch tensor, DLPack, and optional-extension tests: 27 passed, 1
skipped
- Windows ARM64 inline C++ build test passed
- applicable changed-file lint hooks passed
## Follow-up
Replace the NVIDIA nightly override with the official PyTorch
distribution source once stable Windows ARM64 wheels are available.
---------
Co-authored-by: Chao Zhang <[email protected]>
Co-authored-by: Charles Zhang <[email protected]>
---
.github/workflows/publish_wheel.yml | 1 +
pyproject.toml | 14 ++++++++++++--
python/tvm_ffi/cpp/extension.py | 13 +++++++++++--
3 files changed, 24 insertions(+), 4 deletions(-)
diff --git a/.github/workflows/publish_wheel.yml
b/.github/workflows/publish_wheel.yml
index 804e396c..8b2085d3 100644
--- a/.github/workflows/publish_wheel.yml
+++ b/.github/workflows/publish_wheel.yml
@@ -34,6 +34,7 @@ jobs:
- {os: ubuntu-latest, arch: x86_64, linux_image: manylinux_2_28,
build_sdist: "true"}
- {os: ubuntu-24.04-arm, arch: aarch64, linux_image: manylinux_2_28,
build_sdist: "false"}
- {os: windows-latest, arch: AMD64, linux_image: "", build_sdist:
"false"}
+ - {os: windows-11-arm, arch: ARM64, linux_image: "", build_sdist:
"false"}
- {os: macos-14, arch: arm64, linux_image: "", build_sdist: "false"}
steps:
diff --git a/pyproject.toml b/pyproject.toml
index dddadc2a..4fe77c73 100644
--- a/pyproject.toml
+++ b/pyproject.toml
@@ -43,7 +43,8 @@ cpp = ["ninja"]
[dependency-groups]
torch = [
- "torch; python_version < '3.14'", # pytorch does not yet ship with 3.14t
+ # Windows ARM64 tests install PyTorch from NVIDIA's nightly index below.
+ "torch; python_version < '3.14' and (platform_system != 'Windows' or
platform_machine != 'ARM64')",
"setuptools", # setuptools is needed by torch jit for
best perf
"ninja",
"numpy",
@@ -268,9 +269,18 @@ archs = ["x86_64", "arm64"]
environment = { MACOSX_DEPLOYMENT_TARGET = "10.14" }
[tool.cibuildwheel.windows]
-archs = ["AMD64"]
+archs = ["AMD64", "ARM64"]
repair-wheel-command = "delvewheel repair --ignore-existing -w {dest_dir} -v
{wheel}"
+[[tool.cibuildwheel.overrides]]
+select = "*-win_arm64"
+before-test = "uv pip install torch"
+
+[tool.cibuildwheel.overrides.environment]
+UV_INDEX = "https://pypi.nvidia.com/nvtorch_oot_nightly/";
+UV_NO_CONFIG = "1"
+UV_PRERELEASE = "allow"
+
[tool.ty.environment]
python-version = "3.9"
extra-paths = ["python", "examples", "tests/python"]
diff --git a/python/tvm_ffi/cpp/extension.py b/python/tvm_ffi/cpp/extension.py
index a575b767..0bf06f07 100644
--- a/python/tvm_ffi/cpp/extension.py
+++ b/python/tvm_ffi/cpp/extension.py
@@ -22,6 +22,7 @@ import functools
import hashlib
import logging
import os
+import platform
import shutil
import subprocess
import sys
@@ -344,11 +345,19 @@ def _run_command_in_dev_prompt(
if not Path(vsdevcmd_path).exists():
raise FileNotFoundError(f"VsDevCmd.bat not found at:
{vsdevcmd_path}")
+ machine = platform.machine().lower()
+ if machine in ("arm64", "aarch64"):
+ msvc_arch = "arm64"
+ elif machine in ("amd64", "x86_64"):
+ msvc_arch = "x64"
+ else:
+ raise RuntimeError(f"Unsupported Windows architecture:
{platform.machine()}")
+
# Use cmd.exe to run the batch file and then your command.
# The /k flag keeps the command prompt open after the batch file runs.
# The "&" symbol chains the commands.
- cmd_command = '"{vsdevcmd_path}" -arch=x64 & {command}'.format(
- vsdevcmd_path=vsdevcmd_path, command=" ".join(args)
+ cmd_command = '"{vsdevcmd_path}" -arch={arch} & {command}'.format(
+ vsdevcmd_path=vsdevcmd_path, arch=msvc_arch, command=" ".join(args)
)
# Execute the command in a new shell
