Skip to content

[release/2.11] Add gfx1100, gfx1101, gfx1151 to hipBLASLt preferred archs #3317

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Open
nemanjaudovic wants to merge 3 commits into
base: release/2.11
Choose a base branch
from
+15 −4
Open

[release/2.11] Add gfx1100, gfx1101, gfx1151 to hipBLASLt preferred archs #3317

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
3 commits
Select commit Hold shift + click to select a range
af33880
Add gfx1100, gfx1101, gfx1151 to hipBLASLt preferred archs
nemanjaudovic Jun 17, 2026
81fdbba
Merge branch 'release/2.11' into hipblaslt-preferred-gfx1100-gfx1151-…
nemanjaudovic Jul 1, 2026
196fc37
Add test for hipBLASLt preferred archs (gfx1100/gfx1101/gfx1151)
nemanjaudovic Jul 1, 2026
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
3 changes: 3 additions & 0 deletions aten/src/ATen/cuda/detail/CUDAHooks.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -560,6 +560,9 @@ const std::vector<std::string>& CUDAHooks::getHipblasltPreferredArchs() const {
#if ROCM_VERSION >= 70000
"gfx950",
#endif
#if ROCM_VERSION >= 71300
"gfx1100", "gfx1101", "gfx1151",
#endif
#if ROCM_VERSION >= 70200
"gfx1250"
#endif
Expand Down
16 changes: 12 additions & 4 deletions test/test_cuda.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -681,10 +681,18 @@ def _check_default():
self.assertTrue(default == torch._C._BlasBackend.Cublas)
else:
# ROCm logic is less so, it's cublaslt for some Instinct, cublas for all else
gcn_arch = str(
torch.cuda.get_device_properties(0).gcnArchName.split(":", 1)[0]
)
if gcn_arch in ["gfx90a", "gfx942", "gfx950", "gfx1200", "gfx1201", "gfx1250"]:
# Mirror CUDAHooks::getHipblasltPreferredArchs in CUDAHooks.cpp
ROCM_VERSION = tuple(int(v) for v in torch.version.hip.split(".")[:2])
archs = ["gfx90a", "gfx942"]
if ROCM_VERSION >= (6, 4):
archs.extend(["gfx1200", "gfx1201"])
if ROCM_VERSION >= (7, 0):
archs.append("gfx950")
if ROCM_VERSION >= (7, 13):
archs.extend(["gfx1100", "gfx1101", "gfx1151"])
gcn_arch_name = torch.cuda.get_device_properties(0).gcnArchName
hipblaslt_preferred = any(arch in gcn_arch_name for arch in archs)
if hipblaslt_preferred:
self.assertTrue(default == torch._C._BlasBackend.Cublaslt)
else:
self.assertTrue(default == torch._C._BlasBackend.Cublas)
Expand Down