Skip to content

[ops] test: cover chunk CE with activation offload and checkpointing #1004

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
Coach257 merged 9 commits into from
Oct 29, 2026
Merged

[ops] test: cover chunk CE with activation offload and checkpointing #1004

Changes from all commits
Commits
Show all changes
9 commits
Select commit Hold shift + click to select a range
208292f
[ops, dist, config, docs, ci] fix: make chunk loss compatible with ac…
YLlllllllllll Jul 30, 2026
795577b
[ops, docs] fix: reconcile chunk loss offload with current main
YLlllllllllll Oct 9, 2026
e10bc1b
[ops, ci] fix: consolidate chunk-loss parity gates on current main
YLlllllllllll Oct 9, 2026
d3846dd
[test] fix: exercise native NPU chunk-loss offload transfers
YLlllllllllll Oct 9, 2026
29f26c1
[ops] fix: evaluate chunked loss without gradient precomputation
YLlllllllllll Oct 16, 2026
102943a
[ops] test: port chunk CE offload and checkpoint coverage to VeomniOp
YLlllllllllll Oct 20, 2026
2637106
[test] fix: make chunk logits eligible for offload audit
YLlllllllllll Oct 20, 2026
a5a3703
[ops] test: isolate deterministic chunk-loss offload parity
YLlllllllllll Oct 27, 2026
ccdf019
[integration] fix: preserve upstream tests and PR behavior
YLlllllllllll Oct 28, 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
16 changes: 9 additions & 7 deletions tests/ops/loss/test_chunk_loss_offload.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -33,19 +33,21 @@ def _capture_chunk_tokens(target):
chunk_loss_impl._CAPTURE.reset(token)


@pytest.fixture(scope="module", autouse=True)
def _deterministic_npu_comparisons():
"""Keep exact offload comparisons independent of NPU reduction scheduling."""
if not IS_NPU_AVAILABLE:
yield
return
@pytest.fixture(autouse=True)
def _deterministic_reductions(device):
"""Keep bitwise comparisons independent of parallel reduction scheduling."""
enabled = torch.are_deterministic_algorithms_enabled()
warn_only = torch.is_deterministic_algorithms_warn_only_enabled()
torch.use_deterministic_algorithms(True)
num_threads = torch.get_num_threads()
try:
if device == "cpu":
torch.set_num_threads(1)
if device in {"cpu", "npu"}:
torch.use_deterministic_algorithms(True, warn_only=False)
yield
finally:
torch.use_deterministic_algorithms(enabled, warn_only=warn_only)
torch.set_num_threads(num_threads)


def _run(device, dtype, *, offload, recompute, frozen, all_ignored, bias_kind=None, observed=None):
Expand Down
Loading