Skip to content

fix: remove redundant self-assignment out_ = out_ #3367

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
andrewwhitecdw wants to merge 2 commits into
base: main
Choose a base branch
from
+21 −2
Open

fix: remove redundant self-assignment out_ = out_ #3367

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
67154c9
fix: remove redundant self-assignment out_ = out_
andrewwhitecdw Aug 12, 2026
cb09267
[pre-commit.ci] auto fixes from pre-commit.com hooks
pre-commit-ci[bot] Aug 13, 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: 1 addition & 2 deletions tests/pytorch/attention/run_attention_with_cp.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -517,7 +517,7 @@ def run_dpa_with_cp(
torch.cuda.Stream(),
cp_comm_type,
)
if config.softmax_type != "vanilla":
if is_training and config.softmax_type != "vanilla":
core_attn.softmax_offset.grad.zero_()
if dtype == "fp8":
core_attn.fp8_initialized = False
Expand Down Expand Up @@ -690,7 +690,6 @@ def run_dpa_with_cp(
)
else:
out = out.index_select(0, seq_idx_q).contiguous()
out_ = out_

atol, rtol, rmse_tol = get_tols(config, dtype)
tensors_cp = [out_, dq_, dk_, dv_, dbias_, d_softmax_offset_, max_logit_]
Expand Down
20 changes: 20 additions & 0 deletions tests/pytorch/attention/test_softmax_offset_inference.py
View file
Open in desktop
Original file line number Diff line number Diff line change
@@ -0,0 +1,20 @@
import pytest
import torch
from transformer_engine.pytorch import DotProductAttention
Comment on lines +1 to +3

@greptile-apps greptile-apps Bot Aug 13, 2026

Copy link
Copy Markdown
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

P1 Missing source license header

The new Python file omits the required NVIDIA copyright and license notices, causing the repository's L0 license checker to reject it.

Suggested change
import pytest
import torch
from transformer_engine.pytorch import DotProductAttention
# Copyright (c) 2022-2026, NVIDIA CORPORATION & AFFILIATES. All rights reserved.
#
# See LICENSE for license information.
import pytest
import torch
from transformer_engine.pytorch import DotProductAttention

Knowledge Base Used: Tests and QA



@pytest.mark.skipif(not torch.cuda.is_available(), reason="CUDA not available")
def test_softmax_offset_grad_none_in_eval():
Comment on lines +6 to +7

@greptile-apps greptile-apps Bot Aug 13, 2026

Copy link
Copy Markdown
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

P2 Regression test omitted from CI

The QA jobs enumerate test files explicitly, but none includes this new test, so CI silently skips the intended inference regression coverage.

Knowledge Base Used: Tests and QA

Note: If this suggestion doesn't match your team's coding style, reply to this and let me know. I'll remember it for next time!

"""Regression test: eval mode leaves softmax_offset.grad as None.

The context-parallel test helper previously crashed here by calling
core_attn.softmax_offset.grad.zero_() unconditionally for non-vanilla
softmax. In eval mode requires_grad is False and no backward has run,
so .grad must stay None.
"""
core_attn = (
DotProductAttention(8, (64, 64), num_gqa_groups=4, softmax_type="softmax_offset")
.cuda()
.eval()
)
assert not core_attn.softmax_offset.requires_grad