[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
@@ -51,8 +51,7 @@ def get_kv_connector_cache_layout(vllm_config: VllmConfig | None = None):
class KVOutputAggregator:
"""Utility class to aggregate the output of all workers into a single
output corresponding to Rank 0 for scheduler.
"""
output corresponding to Rank 0 for scheduler."""
def __init__(self, expected_finished_count: int):
# Complete transfer tracker. Used to track finished requests
@@ -278,8 +278,7 @@ class ExampleHiddenStatesConnector(KVConnectorBase_V1, SupportsHMA):
def _on_write_done(self, req_id: str, future: Future) -> None:
"""Surface any exception from the disk-write thread and drop the
completed future from the in-flight tracking dict.
"""
completed future from the in-flight tracking dict."""
self._req_futures.pop(req_id, None)
exc = future.exception()
if exc is not None:
@@ -231,8 +231,7 @@ class LMCacheMPRequestTracker:
####
def needs_retrieve(self) -> bool:
"""Check whether the current request needs retrieve, will be used
update_stage_after_alloc
"""
update_stage_after_alloc"""
return (
self.num_lmcache_hit_blocks > self.num_vllm_hit_blocks
and self.state != LMCacheMPRequestState.READY
@@ -240,8 +239,7 @@ class LMCacheMPRequestTracker:
def is_ready_for_retrieving(self) -> bool:
"""Check whether the current request is ready for retrieving,
will be used in process_loading_requests
"""
will be used in process_loading_requests"""
return (
self.state == LMCacheMPRequestState.WAITING_FOR_LOAD
and self.needs_retrieve()
@@ -718,8 +718,7 @@ class MooncakeConnectorScheduler:
def _get_remote_prefill_token_count(self, num_prompt_tokens: int) -> int:
"""D-side only. Returns N-1 for Mamba models since the decoder
always recomputes the last token and must start from h(N-1).
"""
always recomputes the last token and must start from h(N-1)."""
if self._has_mamba and num_prompt_tokens > 1:
return num_prompt_tokens - 1
return num_prompt_tokens
@@ -730,8 +729,7 @@ class MooncakeConnectorScheduler:
derive h(N) correctly.
Guarded by ``_p_side_truncated`` to avoid repeated truncation if the
request is preempted and rescheduled.
"""
request is preempted and rescheduled."""
params = request.kv_transfer_params
if (
params is not None
@@ -1677,8 +1675,7 @@ class MooncakeConnectorWorker:
def _bind_sender_thread_device(self) -> None:
"""ThreadPoolExecutor initializer — binds each pool thread to the
correct CUDA device. CUDA device selection is thread-local, so
without this, NVLink transfers fail for TP ranks > 0.
"""
without this, NVLink transfers fail for TP ranks > 0."""
current_platform.set_device(self.device_id)
def _send_blocks(
@@ -1904,8 +1901,7 @@ class MooncakeConnectorWorker:
def get_kv_connector_stats(self) -> KVConnectorStats | None:
"""Return transfer stats collected since the last call, or None
if nothing has been recorded in this interval.
"""
if nothing has been recorded in this interval."""
if self.xfer_stats.is_empty():
return None
return self.xfer_stats.clone_and_reset()
@@ -73,8 +73,7 @@ class ExternalCachedBlockPool:
class MooncakeStoreCoordinator:
"""Mirror of ``HybridKVCacheCoordinator.find_longest_cache_hit`` over an
``ExternalCachedBlockPool``.
"""
``ExternalCachedBlockPool``."""
def __init__(
self,
@@ -2629,8 +2629,7 @@ class LookupKeyClient:
non_block: bool = False,
) -> MooncakeLookupResult | None:
"""If non_block is True, will return None until the result is ready,
so the caller retries on a later step.
"""
so the caller retries on a later step."""
future = self.futures.get(req_id)
if future is None:
future = self.executor.submit(self._lookup, num_tokens, list(block_hashes))
@@ -548,8 +548,7 @@ class MoRIIOConnectorScheduler:
self, request_id: ReqId, transfer_id: TransferId | None = None
):
"""Unmap request_id/transfer_id. Uses transfer_id for lookup if
exact request_id match fails (handles input_processor mutation).
"""
exact request_id match fails (handles input_processor mutation)."""
if request_id in self.request_id_to_transfer_id:
tid = self.request_id_to_transfer_id[request_id]
del self.request_id_to_transfer_id[request_id]
@@ -400,8 +400,7 @@ class NixlBaseConnectorScheduler:
def _get_remote_prefill_token_count(self, num_prompt_tokens: int) -> int:
"""D-side only. The number of prompt tokens to load from the prefiller.
Stops short of the trailing ``_prefill_backoff()`` tokens that the decoder
will recompute locally.
"""
will recompute locally."""
backoff = self._prefill_backoff()
if backoff and num_prompt_tokens > backoff:
return num_prompt_tokens - backoff
@@ -414,8 +413,7 @@ class NixlBaseConnectorScheduler:
multi-module MTP it is the drafter's whole lookahead window.
Guarded by ``_p_side_truncated`` to avoid repeated truncation if the
request is preempted and rescheduled.
"""
request is preempted and rescheduled."""
backoff = self._prefill_backoff()
params = request.kv_transfer_params
if (
@@ -1873,8 +1873,7 @@ class NixlBaseConnectorWorker:
) -> np.ndarray:
"""Build remote desc regions (conv sub-projections + ssm) per layer.
For hetero-TP, each D rank reads only its sub-projection slice from
the P rank. Returns an Nx3 uint64 array.
"""
the P rank. Returns an Nx3 uint64 array."""
assert nixl_agent_meta.kv_caches_base_addr, (
"Remote KV cache base addresses must not be empty."
)
@@ -2657,8 +2656,7 @@ class NixlBaseConnectorWorker:
@cached_property
def _attention_kv_caches(self) -> list[torch.Tensor]:
"""Device KV caches of attention layers (mamba states excluded),
as consumed by the receive post-process.
"""
as consumed by the receive post-process."""
assert self.device_kv_caches, (
"_attention_kv_caches accessed before register_kv_caches"
)
@@ -315,8 +315,7 @@ class NixlPushConnectorWorker(NixlBaseConnectorWorker):
executes on the writer; the handshake runs on the background executor
and the request is re-queued onto ``_reg_send_inbox`` once it
completes (at which point ``_ensure_handshake`` returns ``None`` and we
send directly).
"""
send directly)."""
remote_pp_size = reg_data.get("remote_pp_size", 1)
fut = self._ensure_handshake(
reg_data["remote_engine_id"],
@@ -508,8 +507,7 @@ class NixlPushConnectorWorker(NixlBaseConnectorWorker):
``BlockIds`` is canonically a tuple of per-group lists, but some
registration payloads collapse a single-group case to a flat
list. Re-wrap that case so downstream group-aware helpers see a
consistent shape.
"""
consistent shape."""
if block_ids and not isinstance(block_ids[0], (list, tuple)):
return (list(block_ids),)
return block_ids
@@ -37,8 +37,7 @@ def canonical_format_id(kv_cache_layout: str) -> str:
Canonical pages keep the worker's KV layout family, so the id couples the
format version with that family; consumers must match it exactly. The family
keeps its historical NHD/HND spelling so ids stay stable for KV persisted
before the layout enum existed.
"""
before the layout enum existed."""
layout = KVCacheLayout[kv_cache_layout]
legacy = {KVCacheLayout.LBNHC: "nhd", KVCacheLayout.LBHNC: "hnd"}
family = legacy.get(layout, layout.name.lower())
@@ -122,8 +121,7 @@ def _interleave_cp_tokens(
ctx: _RankContext,
) -> tuple[CopyRun, ...]:
"""Place each region's num_tokens rows at their canonical token positions,
one run per chunk of interleaved tokens.
"""
one run per chunk of interleaved tokens."""
runs: list[CopyRun] = []
for region in regions:
if ctx.cp_size == 1:
@@ -258,8 +256,7 @@ def _attention_byte_regions(
cp_size: int,
) -> list[ByteRegion] | None:
"""Byte regions of an attention page, given this rank's head shard.
None when the physical layout is not recognized (fail closed).
"""
None when the physical layout is not recognized (fail closed)."""
bs, heads, head_size = spec.block_size, spec.num_kv_heads, spec.head_size
if tuple(kv_cache.shape) == (num_blocks, heads, bs, 2 * head_size):
return _packed_kv_regions(kv_cache, spec, head_shard, num_head_shards, cp_size)
@@ -373,8 +370,7 @@ def _is_exact_partition(intervals: list[tuple[int, int]], size: int) -> bool:
def _verify_tiling(layer_name: str, per_rank: list[CanonicalPageMapping]) -> None:
"""Whichever ranks a block elects as writers must tile the canonical page
exactly once, and each rank's runs must cover exactly its local page.
"""
exactly once, and each rank's runs must cover exactly its local page."""
size = per_rank[0].canonical_page_size_bytes
num_writers = per_rank[0].num_writers
for mapping in per_rank:
@@ -79,8 +79,7 @@ def get_offloading_event_group_spec(
class _OffloadEventMetadata:
"""BlockStored payload snapshot for one OffloadKey, captured while the
Request is available and kept until the final matching removal event.
``medium`` and ``ownership`` are forwarded from the OffloadingEvent.
"""
``medium`` and ``ownership`` are forwarded from the OffloadingEvent."""
# The chunk's constituent block hashes; the last one is the OffloadKey.
block_hashes: tuple[BlockHash, ...]
@@ -256,8 +255,7 @@ class OffloadingEventsTracker:
def reset(self) -> None:
"""Drop all tracked state; pending payloads are stale after a
manager cache reset.
"""
manager cache reset."""
self._pending_event_metadata.clear()
def _build_event_metadata(
@@ -268,8 +266,7 @@ class OffloadingEventsTracker:
) -> _OffloadEventMetadata:
"""Build the payload snapshot for one offloaded chunk: its
constituent per-block hashes, the whole chunk's tokens, and the
per-block ``block_size``.
"""
per-block ``block_size``."""
hashes_per_chunk = group_config.hashes_per_chunk
assert hashes_per_chunk > 0
assert chunk_idx >= 0
@@ -676,8 +676,7 @@ class OffloadingConnectorScheduler:
start_chunk_idx: int,
) -> int | None:
"""Return the number of consecutive offloaded chunks from the start,
or None if the backend deferred a lookup.
"""
or None if the backend deferred a lookup."""
hit_count = 0
defer_lookup = False
for local_idx, key in enumerate(keys):
@@ -712,8 +711,7 @@ class OffloadingConnectorScheduler:
"""Return the end index (in `keys`) of the last run of
`sliding_window_size` consecutive hits, scanning from the end.
The first run may need a larger window for a partial rightmost chunk.
Returns 0 on miss, None if the backend deferred a lookup.
"""
Returns 0 on miss, None if the backend deferred a lookup."""
defer_lookup = False
pending_in_window = False
consecutive_hits = 0