[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
+9 -18
View File
@@ -43,8 +43,7 @@ def _make_batch(
slot_pool: np.ndarray | None = None,
) -> InputBatch:
"""Consistent batch with random unique request slots, 1 + k_i logits
rows per request at the tail of each query segment.
"""
rows per request at the tail of each query segment."""
if slot_pool is None:
slot_pool = np.arange(max_num_reqs)
slots = rng.choice(slot_pool, size=num_reqs, replace=False)
@@ -142,8 +141,7 @@ def _owned_rows(cu: np.ndarray, owned: np.ndarray) -> np.ndarray:
def test_local_batch_partition(tp_size: int, seed: int):
"""Local sub-batches partition the batch by owner, preserve batch order
within an owner, and carry exactly the owned requests' entries of every
per-request and per-row array.
"""
per-request and per-row array."""
rng = np.random.default_rng(seed)
max_num_reqs = 64
batch = _make_batch(rng, num_reqs=23, max_num_reqs=max_num_reqs, max_spec=3)
@@ -225,8 +223,7 @@ def test_local_batch_partition(tp_size: int, seed: int):
def test_all_to_all_reassembly_simulated(tp_size: int):
"""Vocab shards computed from owner-sorted hidden states and routed
through the simulated all-to-all reassemble into exactly the full-vocab
logits of each rank's owned rows.
"""
logits of each rank's owned rows."""
rng = np.random.default_rng(7)
max_num_reqs = 32
batch = _make_batch(rng, num_reqs=11, max_num_reqs=max_num_reqs, max_spec=4)
@@ -264,8 +261,7 @@ def test_all_to_all_reassembly_simulated(tp_size: int):
@requires_cuda
def test_shard_no_spec():
"""Batches without draft tokens (one logits row per request) degenerate
cleanly: rows mirror requests.
"""
cleanly: rows mirror requests."""
rng = np.random.default_rng(11)
tp_size = 4
max_num_reqs = 32
@@ -284,8 +280,7 @@ def test_shard_no_spec():
@pytest.mark.parametrize("tp_size", [2, 4])
def test_shard_grammar_output(tp_size: int):
"""The grammar filter keeps exactly the owned requests' bitmask rows, in
grammar order, and the per-rank pieces partition the global bitmask.
"""
grammar order, and the per-rank pieces partition the global bitmask."""
rng = np.random.default_rng(5)
max_num_reqs = 32
num_reqs = 12
@@ -340,8 +335,7 @@ def test_shard_grammar_output(tp_size: int):
@pytest.mark.parametrize("tp_size", [2, 4])
def test_gather_src_indices_round_trip(tp_size: int):
"""Padded per-rank result blocks indexed by gathered_src_indices restore
batch order, including when some rank owns zero requests.
"""
batch order, including when some rank owns zero requests."""
rng = np.random.default_rng(3)
max_num_reqs = 40
all_slots = np.arange(max_num_reqs)
@@ -376,8 +370,7 @@ def test_gather_src_indices_round_trip(tp_size: int):
def test_gather_sampler_output_kernels(tp_size: int):
"""The fused pack/unpack kernels round-trip per-rank sampler results
into batch order: token ids zero-filled past the local width, counts
cast back to int32, identical on every rank.
"""
cast back to int32, identical on every rank."""
from dataclasses import replace as dc_replace
from vllm.v1.worker.gpu.sample.output import SamplerOutput
@@ -472,8 +465,7 @@ def test_gather_sampler_output_kernels(tp_size: int):
def test_gather_sampler_output_logprobs_and_nans():
"""num_nans is reduced per request and gathered; LogprobsTensors are
packed into fixed per-request blocks, width-aligned across ranks (pad
and truncate paths), and restored to global row order.
"""
and truncate paths), and restored to global row order."""
from dataclasses import replace as dc_replace
from vllm.v1.outputs import LogprobsTensors
@@ -606,8 +598,7 @@ def test_gather_sampler_output_logprobs_and_nans():
def test_finish_requests_frees_slots_in_sorted_order():
"""Request slots must be freed in the same order on every TP rank:
ownership derives from slot indices, and `finished_req_ids` is a set
whose iteration order is per-process hash-randomized.
"""
whose iteration order is per-process hash-randomized."""
from vllm.v1.worker.gpu.model_runner import GPUModelRunner
removed: list[str] = []