Skip to content

[NO CP][ROCm][inductor][release/2.9] Fix tanh for FP64 #2834

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
wants to merge 4 commits into
base: release/2.9
Choose a base branch
from
+13 −1
Open

[NO CP][ROCm][inductor][release/2.9] Fix tanh for FP64 #2834

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
4 commits
Select commit Hold shift + click to select a range
a53d6cf
Fallback to tanh when dtype is FP64.
naromero77amd Nov 20, 2025
ad780aa
Handle case where there is no dtype.
naromero77amd Nov 21, 2025
495a6de
Fix UT.
naromero77amd Nov 27, 2025
5a6d97e
Correct Triton backport version.
naromero77amd Nov 27, 2025
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 test/inductor/test_op_dtype_prop.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -205,6 +205,9 @@ def test_dtype_aware_codegen(self, op_name: str, load_upcast_to_fp32, input_dtyp
triton_op_name_overrides = {
"round": "nearbyint",
}
# ROCm uses fast_tahnf for everything input types that are not float64
if torch.version.hip and input_dtype != torch.float64:
triton_op_name_overrides["tanh"] = "fast_tanhf"
override = triton_op_name_overrides.get(op_name)
triton_op_name = override if override is not None else torch_op_name

Expand Down
11 changes: 10 additions & 1 deletion torch/_inductor/codegen/triton.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -1315,7 +1315,16 @@ def tan(x):
@staticmethod
@maybe_upcast_float32()
def tanh(x):
if torch.version.hip and get_triton_version() > (3, 2):
cse_var = V.kernel.cse.varname_map.get(x)
if cse_var and hasattr(cse_var, "dtype"):
dtype = cse_var.dtype
else:
dtype = None
if (
torch.version.hip
and get_triton_version() > (3, 2)
and dtype != torch.float64
):
# On ROCm, use fast_tanhf depending on Triton version
# Requires ROCm fork of Triton 3.3, 3.4, 3.5 or upstream Triton 3.6+
return f"libdevice.fast_tanhf({x})"
Expand Down