This is an automated email from the ASF dual-hosted git repository.
tlopex pushed a commit to branch main
in repository https://gitbox.apache.org/repos/asf/tvm.git
The following commit(s) were added to refs/heads/main by this push:
new e7a385a9ca [FIX][CUDA] Select NVRTC architecture for output format
(#20100)
e7a385a9ca is described below
commit e7a385a9cacdccfa4b75b7f6154fe7c7e05b8683
Author: Hongyi Jin <[email protected]>
AuthorDate: Thu Aug 6 17:51:14 2026 -0400
[FIX][CUDA] Select NVRTC architecture for output format (#20100)
NVRTC uses `--gpu-architecture` together with the requested output
format. A real `sm_*` target produces a cubin, while a virtual
`compute_*` target produces PTX.
The previous revision normalized every target to `compute_*`. With CUDA
13.2, compilation reported success, but `nvrtcGetCUBINSize` returned
zero bytes; loading that image then failed with
`CUDA_ERROR_INVALID_IMAGE`.
This change preserves architecture suffixes such as `100a` and `100f`,
while selecting the prefix from the requested output:
- cubin: normalize to `sm_*`;
- PTX: normalize to `compute_*`.
Validation:
- `python -m compileall -q python/tvm/support/nvcc.py`
- CUDA 13.2 / B200 empty-kernel check: `compute_100a` yielded a
zero-byte cubin and failed to load; `sm_100a` yielded a 4,888-byte cubin
and loaded successfully.
- A TIRx NumSim/GPU float-atomic microtest failed with
`CUDA_ERROR_INVALID_IMAGE` before this correction and passed through the
default NVRTC path afterward.
---
python/tvm/support/nvcc.py | 12 ++++++++++--
1 file changed, 10 insertions(+), 2 deletions(-)
diff --git a/python/tvm/support/nvcc.py b/python/tvm/support/nvcc.py
index f9b0ebcc89..de83e70468 100644
--- a/python/tvm/support/nvcc.py
+++ b/python/tvm/support/nvcc.py
@@ -304,7 +304,7 @@ def _compile_cuda_nvrtc(
target_format : str, optional
Output format: "cubin" or "ptx". Default: "cubin"
arch : str, optional
- Target architecture (e.g., "sm_80"). Auto-detected if None.
+ Target architecture (e.g., "compute_80" or "sm_80"). Auto-detected if
None.
options : str or list of str, optional
Additional NVRTC options.
path_target : str, optional
@@ -357,10 +357,18 @@ def _compile_cuda_nvrtc(
if options is not None and not isinstance(options, str | list):
raise ValueError("options must be str or list of str")
- # Auto-detect architecture
+ # NVRTC selects the output kind through both target_format and the
+ # architecture spelling. A virtual ``compute_*`` target produces PTX,
+ # while a real ``sm_*`` target produces a loadable cubin. Keep the suffix
+ # (including family/architecture qualifiers such as ``a`` and ``f``), but
+ # normalize the prefix to the requested output format.
if arch is None:
compute_version =
get_target_compute_version(Target.current(allow_none=True))
arch = f"sm_{''.join(compute_version.split('.'))}"
+ if target_format == "ptx" and arch.startswith("sm_"):
+ arch = f"compute_{arch.removeprefix('sm_')}"
+ elif target_format == "cubin" and arch.startswith("compute_"):
+ arch = f"sm_{arch.removeprefix('compute_')}"
# Get NVSHMEM paths if needed
nvshmem_include_path, nvshmem_lib_path = None, None
