Skip to content

fix: remove custom scaled bmm op on cpu and fix fp8 test #187

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

Merged
chichun-charlie-liu merged 3 commits into from
Oct 17, 2025
Merged

fix: remove custom scaled bmm op on cpu and fix fp8 test #187

Show file tree
Hide file tree
Changes from 1 commit
Commits
Show all changes
3 commits
Select commit Hold shift + click to select a range
eaad33b
Remove custom scaled bmm op on cpu and fix fp8 test
andrea-fasoli Oct 14, 2025
f61c561
Re-enable custom op for pt<=2.7
andrea-fasoli Oct 15, 2025
1c4f22e
Clean up versioning for int8 aiu op
andrea-fasoli Oct 15, 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
58 changes: 0 additions & 58 deletions fms_mo/aiu_addons/fp8/fp8_spyre_op.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -13,9 +13,6 @@
# limitations under the License.
"""Torch registration of FP8xFP8 operation for attention BMMs."""

# Standard
from typing import Optional

# Third Party
from torch import Tensor
import torch
Expand All @@ -29,61 +26,6 @@
# open issue in PyLint: https://github.com/pytorch/pytorch/issues/119482


def _scaled_mm_cpu_out(
mat1: Tensor,
mat2: Tensor,
scale1: Tensor,
scale2: Tensor,
bias: Optional[Tensor] = None,
scale_result: Optional[Tensor] = None,
out_dtype: Optional[torch.dtype] = None,
use_fast_accum: bool = False,
*,
out: Optional[Tensor] = None,
) -> Tensor:
if out_dtype is None:
out_dtype = torch.float32
mat1 = (mat1.to(dtype=out_dtype) * scale1).to(dtype=out_dtype)
mat2 = (mat2.to(dtype=out_dtype) * scale2).to(dtype=out_dtype)

if bias is not None:
ret = torch.addmm(bias, mat1, mat2).to(dtype=out_dtype)
else:
ret = torch.mm(mat1, mat2).to(dtype=out_dtype)

if out is not None:
out.copy_(ret)
return out
return ret


torch.library.register_kernel(torch.ops.aten._scaled_mm.out, "cpu", _scaled_mm_cpu_out)


@torch.library.register_kernel("aten::_scaled_mm", "cpu")
def _scaled_mm_cpu(
mat1: Tensor,
mat2: Tensor,
scale1: Tensor,
scale2: Tensor,
bias: Optional[Tensor] = None,
scale_result: Optional[Tensor] = None,
out_dtype: Optional[torch.dtype] = None,
use_fast_accum: bool = False,
) -> Tensor:
return _scaled_mm_cpu_out(
mat1,
mat2,
scale1,
scale2,
bias,
scale_result,
out_dtype,
use_fast_accum,
out=None,
)


@torch.library.custom_op("spyre::scaled_bmm", mutates_args=())
def spyre_scaled_bmm(
mat1: Tensor,
Expand Down
6 changes: 3 additions & 3 deletions tests/aiu_addons/test_fp8_addon.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -51,9 +51,9 @@ def test_fp8_op() -> None:
# Local
from fms_mo.aiu_addons.fp8.fp8_attn import _math_fp8_compute_op

query = torch.randn((1, 32, 64, 128), dtype=torch.bfloat16, device="cuda")
key = torch.randn((1, 32, 64, 128), dtype=torch.bfloat16, device="cuda")
value = torch.randn((1, 32, 64, 128), dtype=torch.bfloat16, device="cuda")
query = torch.randn((1, 64, 32, 128), dtype=torch.bfloat16, device="cuda")
key = torch.randn((1, 64, 32, 128), dtype=torch.bfloat16, device="cuda")
value = torch.randn((1, 64, 32, 128), dtype=torch.bfloat16, device="cuda")

out = _math_fp8_compute_op(query, key, value, 32, 32, 0.0, None)
assert out.size() == query.size()
Loading