[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
@@ -185,8 +185,7 @@ def test_trtllm_bf16_monolithic_routing_replay_records_valid_experts(
) -> None:
"""The capture callback should receive the int16 routed-expert IDs the
kernel actually used, the values should be valid expert indices, and
each token should pick ``top_k`` distinct experts.
"""
each token should pick ``top_k`` distinct experts."""
if top_k > _DSV3_N_GROUP * _DSV3_TOPK_GROUP:
pytest.skip(
f"DSV3 requires top_k <= n_group * topk_group "
@@ -265,8 +264,7 @@ def test_trtllm_bf16_monolithic_routing_replay_non_dsv3(
routing_method: RoutingMethodType,
) -> None:
"""Routing replay works for non-DeepSeekV3 routing methods too.
FlashInfer's ``routing_replay_out`` is routing-method-agnostic.
"""
FlashInfer's ``routing_replay_out`` is routing-method-agnostic."""
torch.manual_seed(0)
device = torch.device("cuda:0")
@@ -319,8 +317,7 @@ def test_trtllm_bf16_monolithic_routing_replay_non_dsv3(
def test_trtllm_bf16_monolithic_capture_disabled_skips_buffer_alloc() -> None:
"""With no callback installed the kernel should not see a
``routing_replay_out`` tensor — verify the helper short-circuits.
"""
``routing_replay_out`` tensor — verify the helper short-circuits."""
torch.manual_seed(0)
device = torch.device("cuda:0")
experts, _, _ = _make_bf16_monolithic_experts(
@@ -342,8 +339,7 @@ def test_trtllm_bf16_monolithic_capture_disabled_skips_buffer_alloc() -> None:
def test_trtllm_bf16_monolithic_supports_capture_for_all_routing() -> None:
"""FlashInfer's ``routing_replay_out`` is supported by all routing
methods, so ``supports_routing_replay_capture`` should be True
regardless of routing method.
"""
regardless of routing method."""
device = torch.device("cuda:0")
for routing_method in (
RoutingMethodType.DeepSeekV3,
@@ -365,8 +361,7 @@ def test_trtllm_bf16_monolithic_supports_capture_for_all_routing() -> None:
def test_trtllm_bf16_monolithic_capture_buffer_shape_and_dtype() -> None:
"""When capture is installed, the allocated buffer is int16 and shaped
``(num_tokens, experts_per_token)``.
"""
``(num_tokens, experts_per_token)``."""
device = torch.device("cuda:0")
experts, _, _ = _make_bf16_monolithic_experts(
num_experts=_DSV3_NUM_EXPERTS,
@@ -597,8 +592,7 @@ def test_trtllm_fp8_block_scale_monolithic_routing_replay_records_valid_experts(
top_k: int,
) -> None:
"""End-to-end: ``TrtLlmFp8ExpertsMonolithic`` (DeepSeekFp8 block-scale
path, DSV3 routing) captures valid expert IDs.
"""
path, DSV3 routing) captures valid expert IDs."""
if top_k > _DSV3_N_GROUP * _DSV3_TOPK_GROUP:
pytest.skip(
f"DSV3 requires top_k <= n_group * topk_group "
@@ -784,8 +778,7 @@ def _run_nvfp4_monolithic(
routing_bias: torch.Tensor,
) -> torch.Tensor:
"""The monolithic NVFP4 apply expects packed fp4 hidden states + the
matching fp8 per-block scale stored in the ``a1q_scale`` slot.
"""
matching fp8 per-block scale stored in the ``a1q_scale`` slot."""
# Stash the weight tensors on the experts in the locations the apply()
# implementation reads from (it pulls them from quant_config / scales
# already; w1/w2 come in as args).
@@ -813,8 +806,7 @@ def test_trtllm_nvfp4_monolithic_routing_replay_records_valid_experts(
top_k: int,
) -> None:
"""End-to-end: ``TrtLlmNvFp4ExpertsMonolithic`` captures valid expert IDs
on the DSV3 routing path.
"""
on the DSV3 routing path."""
if top_k > _DSV3_N_GROUP * _DSV3_TOPK_GROUP:
pytest.skip(
f"DSV3 requires top_k <= n_group * topk_group "