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 a59b4266 [FIX] Build CUDA sources through tvm_ffi.cpp on Windows (#829)
a59b4266 is described below
commit a59b4266c0b692de311edb63ff7fd12ca7785d24
Author: YevheniiKotyrlo <[email protected]>
AuthorDate: Tue Sep 29 21:28:55 2026 +0300
[FIX] Build CUDA sources through tvm_ffi.cpp on Windows (#829)
## Problem
`_generate_ninja_build` has never built a CUDA source on Windows (#822).
Each failure is reached only once the one before it is fixed:
1. `default_cuda_cflags = ["-Xcompiler", "/std:c++17", "/O2"]`:
`-Xcompiler` forwards one argument, so nvcc reads `/O2` as a second
input file — `nvcc fatal : A single input file is required for a
non-link phase when an outputfile is specified`.
2. That `-Xcompiler /std:c++17` reaches cl after nvcc's own `-std`, so
it overrides a caller's `-std=c++20`.
3. No arch target is passed and `cudart` is never linked — `LNK2019:
unresolved external symbol __cudaRegisterFatBinary`.
4. `/LIBPATH:` is unquoted, and the CUDA toolkit installs under
`C:\Program Files` — `LNK1181: cannot open input file
'Files\NVIDIA.obj'`. The torch C-DLPack addon links Python's `libs` the
same way, and a Python installed for all users lives there too —
`LNK1181: cannot open input file 'Files\Python313\libs.obj'`.
5. **With 1-4 fixed, a module mixing C++ and CUDA sources fails to
link**: nvcc compiles host code against cl's default static runtime
while the C++ objects use `/MD` — `LNK2038: mismatch detected for
'RuntimeLibrary': value 'MD_DynamicRelease' doesn't match value
'MT_StaticRelease'`.
Separately, `default_cxxflags` on Windows has no `/O2`, so C++ sources
build unoptimized, although the `build` and `load` docstrings list `/O2`
and the Linux branch passes `-O2`.
## Solution
The Windows nvcc defaults become Linux's, `-std=c++17 -O2`, which nvcc
forwards to cl for the host pass (`nvcc --dryrun` shows `cl.exe
/std:c++17 ... -O2`), plus `-Xcompiler /MD`. A caller's `-std=c++20`
then applies the way it does on Linux: nvcc keeps the last value of a
repeated option. With CUDA, the arch target from `_get_cuda_target()`
and `cudart.lib` from `CUDA_HOME/lib/x64` are added. Every `/LIBPATH:`
goes through `subprocess.list2cmdline`, which quotes it only when it
holds a space, so a path without one is emitted as before. C++ gets
`/O2`.
## Why it isn't caught today
The Windows CI job has no GPU, so the `load_inline` CUDA tests skip, and
the Python it installs does not live under a path with a space.
## Tests
No new test: the existing CUDA tests cover this build line. On Windows,
`test_load_inline_cuda`, `test_load_inline_both` and
`test_cuda_memory_alloc_noleak` fail on `main` and pass with this
change, and so do `test_build_torch_c_dlpack_extension` and
`test_parallel_build` with Python under a path with a space.
## Verification
Windows 11, MSVC 19.51, CUDA 13.3, RTX 3090 Ti, Python 3.13 under
`C:\Program Files`, `pytest tests/python` inside `VsDevCmd` as CI runs
it: on `main` 5 failed (`test_load_inline_cuda`,
`test_load_inline_both`, `test_cuda_memory_alloc_noleak`,
`test_build_torch_c_dlpack_extension`, `test_parallel_build`) and 2586
passed · on this branch all of them pass · `ruff check`, `ruff format
--check` and `ty check --error-on-warning` clean. With #830, FreeToken's
82 JIT kernels compile, link and load on Windows through this build
line.
## Known limits
A space elsewhere in a generated build still splits it: an include path,
or a build directory under a home directory whose name has one. Those
sit on build statements too, which need ninja's `$ ` escaping as well as
quoting, on every platform, so I kept them out of this PR.
## Base
`c96f9e3`, and it merges cleanly with `main`.
Fixes #822
---
python/tvm_ffi/cpp/extension.py | 26 +++++++++++++++++-----
.../utils/_build_optional_torch_c_dlpack.py | 7 ++++--
2 files changed, 26 insertions(+), 7 deletions(-)
diff --git a/python/tvm_ffi/cpp/extension.py b/python/tvm_ffi/cpp/extension.py
index f11b3284..a575b767 100644
--- a/python/tvm_ffi/cpp/extension.py
+++ b/python/tvm_ffi/cpp/extension.py
@@ -364,6 +364,14 @@ def _run_command_in_dev_prompt(
) from e
+def _windows_argument(argument: str) -> str:
+ """Quote one argument of a Windows command line if it would otherwise
split on a space.
+
+ The CUDA toolkit installs under ``Program Files`` by default.
+ """
+ return subprocess.list2cmdline([argument])
+
+
def _generate_ninja_build( # noqa: PLR0915, PLR0912
name: str,
extra_cflags: Sequence[str],
@@ -400,7 +408,7 @@ def _generate_ninja_build( # noqa: PLR0915, PLR0912
tvm_ffi_lib_name = tvm_ffi_lib.stem
if IS_WINDOWS:
default_cflags = ["/O2", "/MD"]
- default_cxxflags = ["/std:c++17", "/MD", "/EHsc"]
+ default_cxxflags = ["/std:c++17", "/O2", "/MD", "/EHsc"]
_win_warnings = [
"/wd4819",
"/wd4251",
@@ -415,12 +423,20 @@ def _generate_ninja_build( # noqa: PLR0915, PLR0912
]
default_cflags += _win_warnings
default_cxxflags += _win_warnings
- default_cuda_cflags = ["-Xcompiler", "/std:c++17", "/O2"]
+ # nvcc forwards its own -std and -O to cl for the host pass, as it
does to gcc.
+ # /MD matches the C++ objects: cl's default /MT fails a mixed link
(LNK2038).
+ default_cuda_cflags = ["-Xcompiler", "/MD", "-std=c++17", "-O2"]
default_ldflags = [
"/DLL",
- f"/LIBPATH:{tvm_ffi_lib_path}",
+ _windows_argument(f"/LIBPATH:{tvm_ffi_lib_path}"),
f"{tvm_ffi_lib_name}.lib",
]
+ if with_cuda:
+ default_cuda_cflags += [_get_cuda_target()]
+ default_ldflags += [
+ _windows_argument(f"/LIBPATH:{Path(_find_cuda_home()) / 'lib'
/ 'x64'}"),
+ "cudart.lib", # cuda runtime library
+ ]
else:
default_cflags = ["-fPIC", "-O2"]
default_cxxflags = ["-std=c++17", "-fPIC", "-O2"]
@@ -1266,7 +1282,7 @@ def build( # noqa: PLR0913
The default flags are:
- ['-Xcompiler', '-fPIC', '-std=c++17', '-O2'] (Linux/macOS)
- - ['-Xcompiler', '/std:c++17', '/O2'] (Windows)
+ - ['-Xcompiler', '/MD', '-std=c++17', '-O2'] (Windows)
extra_ldflags
The extra linker flags.
@@ -1422,7 +1438,7 @@ def load( # noqa: PLR0913
The default flags are:
- ['-Xcompiler', '-fPIC', '-std=c++17', '-O2'] (Linux/macOS)
- - ['-Xcompiler', '/std:c++17', '/O2'] (Windows)
+ - ['-Xcompiler', '/MD', '-std=c++17', '-O2'] (Windows)
extra_ldflags
The extra linker flags.
diff --git a/python/tvm_ffi/utils/_build_optional_torch_c_dlpack.py
b/python/tvm_ffi/utils/_build_optional_torch_c_dlpack.py
index 9e4f9a16..e8466d93 100644
--- a/python/tvm_ffi/utils/_build_optional_torch_c_dlpack.py
+++ b/python/tvm_ffi/utils/_build_optional_torch_c_dlpack.py
@@ -847,7 +847,7 @@ def main() -> None: # noqa: PLR0912, PLR0915
for lib_dir in torch.utils.cpp_extension.library_paths():
if IS_WINDOWS:
- ldflags.append(f"/LIBPATH:{lib_dir}")
+
ldflags.append(subprocess.list2cmdline([f"/LIBPATH:{lib_dir}"]))
else:
ldflags.extend(["-L", str(lib_dir)])
@@ -879,7 +879,10 @@ def main() -> None: # noqa: PLR0912, PLR0915
)
for python_libdir in python_libdir_list:
if python_libdir and (Path(python_libdir) /
python_lib).exists():
- ldflags.append(f"/LIBPATH:{python_libdir.replace(':',
'$:')}")
+ # A Python installed for all users lives under C:\Program
Files.
+ ldflags.append(
+
subprocess.list2cmdline([f"/LIBPATH:{python_libdir.replace(':', '$:')}"])
+ )
break
if IS_DARWIN:
