[CI] Ignore ruff D209, rejoin the docstrings it split, and silence incompatible-rule warnings (#57212)

Signed-off-by: Harry Mellor <19981378+hmellor@users.noreply.github.com>
Co-authored-by: Claude Opus 5 (1M context) <noreply@anthropic.com>
This commit is contained in:
Harry Mellor
2026-09-16 10:57:24 -07:00
committed by GitHub
co-authored by Claude Opus 5
parent 1085b64425
commit 651a88c09a
693 changed files with 2088 additions and 4171 deletions
@@ -123,8 +123,7 @@ def _check_writes_only_its_own_shard(
device: torch.device, tp_size: int, rank: int
) -> None:
"""Two ranks that swapped both weight rows and write offsets still sum to
the right total, so inspect each slice before the final collective.
"""
the right total, so inspect each slice before the final collective."""
transform = _build_transform(device)
runner = _tail_runner(transform, tp_size)
group = get_tp_group().device_group
@@ -130,8 +130,7 @@ def test_disabled_reproduces_the_plain_residual_sum(
def test_taps_the_consumer_layer_when_one_follows(recorder, monkeypatch):
"""The value the next layer reads is the mixture against *its* weights,
so the tap has to reach forward rather than use the current layer's.
"""
so the tap has to reach forward rather than use the current layer's."""
_set_last_rank(monkeypatch, True)
_call(_stub_model(enabled=True), 1, torch.zeros(2), None, torch.zeros(2))
@@ -182,8 +181,7 @@ def test_pending_mlp_output_is_folded_in_rather_than_passed_as_delta(
):
"""The kernel writes an applied delta back into the prefix in place, which
would double-add it into the live residual stream, so the pending output
has to arrive already summed into the prefix with `delta` left None.
"""
has to arrive already summed into the prefix with `delta` left None."""
_set_last_rank(monkeypatch, True)
prefix_sum = torch.tensor([1.0, 2.0])
pending = torch.tensor([0.5, 0.25])
+1 -2
View File
@@ -727,8 +727,7 @@ def test_cudagraph_capture_metadata_avoids_device_to_host_copy(
"""`build_for_cudagraph_capture` is not capture-only: a DP rank with nothing
scheduled re-stages its FULL-graph metadata through it on every dummy step,
so it must not synchronize the device. The host-side draft counts have to
come from `query_start_loc_cpu` and match the device-derived values.
"""
come from `query_start_loc_cpu` and match the device-derived values."""
num_speculative_tokens = 2
batch = BatchSpec(seq_lens=[50, 30], query_lens=[3, 3])
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")