Skip to content

[Ops] Use block matrix inversion to speed up solve_tril for the Ascend NPU backend #1145

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
OsirisDuan wants to merge 2 commits into
base: main
Choose a base branch
from
+341 −265
Open

[Ops] Use block matrix inversion to speed up solve_tril for the Ascend NPU backend #1145

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
2 commits
Select commit Hold shift + click to select a range
f1661af
perf(triton_ascend): Use block matrix inversion to speed up solve_tri…
OsirisDuan Aug 17, 2026
eae2682
perf(triton_ascend): Bind solve_tril kernels to NPU physical core grid
OsirisDuan Aug 19, 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
13 changes: 12 additions & 1 deletion fla/ops/utils/backends/triton_ascend/__init__.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -9,6 +9,8 @@

from __future__ import annotations

import torch

from fla.ops.backends import BaseBackend


Expand All @@ -23,7 +25,16 @@ def is_available(cls) -> bool:
from fla.utils import IS_NPU
return IS_NPU

def solve_tril_verifier(self, *args, **kwargs):
def solve_tril_verifier(self, A: torch.Tensor, *args, **kwargs) -> tuple[bool, str | None]:
from fla.utils import IS_NPU
if not IS_NPU:
return False, "not running on NPU"
if A.device.type != 'npu':
return False, f"input device is not NPU, got {A.device.type}"
if A.dtype not in (torch.float16, torch.bfloat16, torch.float32):
return False, f"unsupported dtype for NPU solve_tril: {A.dtype}"
if A.shape[-1] not in (16, 32, 64):
return False, f"solve_tril requires BT in (16, 32, 64), got {A.shape[-1]}"
return True, None

def solve_tril(self, *args, **kwargs):
Expand Down
Loading
Loading