[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:
co-authored by
Claude Opus 5
parent
1085b64425
commit
651a88c09a
@@ -15,8 +15,7 @@ from vllm.v1.kv_offload.cpu.policies.lru import LRUCachePolicy
|
||||
class _DummyCachePolicy(CachePolicy):
|
||||
"""Minimal CachePolicy for CachePolicyFactory registration tests. Loaded
|
||||
by module path, so it must be importable at module scope (mirrors
|
||||
tests/v1/kv_offload/test_factory.py's SingleArgExternalOffloadingSpec).
|
||||
"""
|
||||
tests/v1/kv_offload/test_factory.py's SingleArgExternalOffloadingSpec)."""
|
||||
|
||||
def __init__(self, cache_capacity: int) -> None:
|
||||
self.cache_capacity = cache_capacity
|
||||
@@ -55,8 +54,7 @@ class TestCachePolicyFactory:
|
||||
|
||||
def test_pre_registered_policies_can_be_imported(self):
|
||||
"""If someone moves a policy module but forgets to update
|
||||
factory.py, CI fails.
|
||||
"""
|
||||
factory.py, CI fails."""
|
||||
for name in CachePolicyFactory._registry:
|
||||
cls = CachePolicyFactory._registry[name]()
|
||||
assert issubclass(cls, CachePolicy)
|
||||
@@ -90,8 +88,7 @@ class TestCachePolicyFactory:
|
||||
register_cache_policy() call -- this is how external projects
|
||||
integrate a custom CachePolicy without forking/patching vLLM.
|
||||
Mirrors tests/v1/kv_offload/test_factory.py's
|
||||
test_dynamic_load_via_spec_module_path.
|
||||
"""
|
||||
test_dynamic_load_via_spec_module_path."""
|
||||
policy_cls = CachePolicyFactory.get_cache_policy_cls(
|
||||
"_DummyCachePolicy", "tests.v1.kv_offload.cpu.policies.test_factory"
|
||||
)
|
||||
@@ -99,8 +96,7 @@ class TestCachePolicyFactory:
|
||||
|
||||
def test_manager_resolves_policy_via_module_path(self):
|
||||
"""End-to-end: CPUOffloadingManager resolves an unregistered policy
|
||||
purely from cache_policy_module_path.
|
||||
"""
|
||||
purely from cache_policy_module_path."""
|
||||
manager = CPUOffloadingManager(
|
||||
num_chunks=4,
|
||||
cache_policy="_DummyCachePolicy",
|
||||
@@ -110,7 +106,6 @@ class TestCachePolicyFactory:
|
||||
|
||||
def test_unregistered_policy_without_module_path_raises(self):
|
||||
"""eviction_policy not in registry + no cache_policy_module_path ->
|
||||
ValueError, same shape as the OffloadingSpecFactory error path.
|
||||
"""
|
||||
ValueError, same shape as the OffloadingSpecFactory error path."""
|
||||
with pytest.raises(ValueError, match="Unknown cache policy"):
|
||||
CachePolicyFactory.get_cache_policy_cls("nonexistent", None)
|
||||
|
||||
@@ -126,8 +126,7 @@ def _canonical_handler(gpu_tensor, cpu_tensor, mapping, gpu_to_cpu):
|
||||
def test_gpu_roundtrip_assembles_canonical_page_across_ranks():
|
||||
"""Two TP2-style rank handlers must scatter into one shared canonical
|
||||
CPU page such that each rank reloads bit-exact and a whole-page (TP1)
|
||||
reader sees both shards — the cross-topology contract.
|
||||
"""
|
||||
reader sees both shards — the cross-topology contract."""
|
||||
torch.manual_seed(0)
|
||||
num_blocks = 4
|
||||
gpu_rank = [
|
||||
@@ -204,8 +203,7 @@ def _nhd_shard_mapping(tp: int, rank: int) -> CanonicalPageMapping:
|
||||
|
||||
def _head_shard(full_kv: torch.Tensor, tp: int, rank: int) -> torch.Tensor:
|
||||
"""This rank's local page rows out of the (blocks, 2, tokens, heads,
|
||||
head_bytes) ground truth.
|
||||
"""
|
||||
head_bytes) ground truth."""
|
||||
local_heads = _TOTAL_HEADS // tp
|
||||
shard = full_kv[:, :, :, rank * local_heads : (rank + 1) * local_heads, :]
|
||||
return shard.reshape(full_kv.shape[0], -1).contiguous()
|
||||
@@ -216,8 +214,7 @@ def _head_shard(full_kv: torch.Tensor, tp: int, rank: int) -> torch.Tensor:
|
||||
def test_cross_topology_roundtrip(writer_tp: int, reader_tp: int):
|
||||
"""KV written at one tp must be readable at another: writer ranks scatter
|
||||
head shards into a shared canonical region, reader ranks gather their own
|
||||
shards, and every reader must see the writers' ground-truth bytes.
|
||||
"""
|
||||
shards, and every reader must see the writers' ground-truth bytes."""
|
||||
torch.manual_seed(0)
|
||||
num_blocks = 3
|
||||
full_kv = torch.randint(
|
||||
|
||||
@@ -167,8 +167,7 @@ def _mp_race_construct_and_write(
|
||||
) -> None:
|
||||
"""Race to construct a SharedOffloadRegion, write fill_value, then wait
|
||||
for the parent's cleanup signal before tearing down. The wait gives the
|
||||
parent a window to read the raw mmap before the creator removes the file.
|
||||
"""
|
||||
parent a window to read the raw mmap before the creator removes the file."""
|
||||
try:
|
||||
region = SharedOffloadRegion(
|
||||
engine_id=engine_id,
|
||||
@@ -197,8 +196,7 @@ def _mp_barrier_construct_and_hold(
|
||||
replicated: bool = False,
|
||||
) -> None:
|
||||
"""Construct with a real cross-process barrier, write, then hold the
|
||||
mapping (no cleanup) until the parent SIGKILLs this process.
|
||||
"""
|
||||
mapping (no cleanup) until the parent SIGKILLs this process."""
|
||||
try:
|
||||
populate_calls = 0
|
||||
get_populate_write_fn = region_module._get_populate_write_fn
|
||||
@@ -252,8 +250,7 @@ def iid():
|
||||
|
||||
def test_create_next_worker_view_shape_and_stride(iid):
|
||||
"""Returned tensor must have shape (num_chunks, tensor_page_size) and
|
||||
stride (row_stride, 1) where row_stride = cpu_page_size * num_workers.
|
||||
"""
|
||||
stride (row_stride, 1) where row_stride = cpu_page_size * num_workers."""
|
||||
with _region(iid, num_chunks=4, cpu_page_size=2 * PAGE_SIZE) as r:
|
||||
t = r.create_next_worker_view(PAGE_SIZE)
|
||||
assert t.shape == (4, PAGE_SIZE)
|
||||
@@ -293,8 +290,7 @@ def test_create_next_worker_view_row_stride_with_multiple_workers(iid):
|
||||
|
||||
def test_create_next_worker_view_cursor_advances(iid):
|
||||
"""Each create_next_worker_view call must advance _worker_offset by
|
||||
tensor_page_size.
|
||||
"""
|
||||
tensor_page_size."""
|
||||
with _region(iid, cpu_page_size=3 * PAGE_SIZE) as r:
|
||||
assert r._worker_offset == 0
|
||||
r.create_next_worker_view(PAGE_SIZE)
|
||||
@@ -351,8 +347,7 @@ def test_create_next_worker_view_overflow_does_not_mutate_cursor(iid):
|
||||
|
||||
def test_create_next_worker_view_write_visible_in_raw_mmap(iid):
|
||||
"""Writes into a create_next_worker_view view must appear at the correct
|
||||
raw mmap offset
|
||||
"""
|
||||
raw mmap offset"""
|
||||
with _region(iid, num_chunks=4) as r:
|
||||
t = r.create_next_worker_view(PAGE_SIZE)
|
||||
t[2, :] = 42 # write to chunk row 2
|
||||
@@ -385,8 +380,7 @@ def test_create_next_worker_view_multi_tensor_layout(iid):
|
||||
|
||||
def test_create_next_worker_view_multiprocess_slots(iid):
|
||||
"""Each worker process calls create_next_worker_view and writes distinct data;
|
||||
the parent verifies each slot lands at the correct interleaved offset.
|
||||
"""
|
||||
the parent verifies each slot lands at the correct interleaved offset."""
|
||||
num_workers = 2
|
||||
num_chunks = 4
|
||||
|
||||
@@ -671,8 +665,7 @@ def test_multi_worker_race_shared_memory_visible(iid):
|
||||
|
||||
def test_multiprocess_race_construct_and_write(iid):
|
||||
"""N processes race to construct the same SharedOffloadRegion, each writes
|
||||
fill_value = rank+1 into their slot; parent verifies interleaved layout.
|
||||
"""
|
||||
fill_value = rank+1 into their slot; parent verifies interleaved layout."""
|
||||
num_workers = 4
|
||||
num_chunks = 3
|
||||
|
||||
@@ -798,8 +791,7 @@ def test_cleanup_unregisters_every_pinned_chunk(iid, monkeypatch):
|
||||
def test_cleanup_after_create_next_worker_view_releases_mmap(iid):
|
||||
"""cleanup() must close the mmap even after create_next_worker_view was called.
|
||||
create_next_worker_view returns a view that shares storage with _base; both must be
|
||||
released before mmap.close() can succeed.
|
||||
"""
|
||||
released before mmap.close() can succeed."""
|
||||
r = _make_region(iid)
|
||||
mmap_obj = r.mmap_obj
|
||||
|
||||
@@ -978,8 +970,7 @@ def test_ftruncate_failure_cleans_up_creator(monkeypatch):
|
||||
|
||||
def test_backing_file_unlinked_after_barrier(iid):
|
||||
"""The file must exist until the barrier releases (late joiners need the
|
||||
name) and be gone right after, so no exit path can leak it.
|
||||
"""
|
||||
name) and be gone right after, so no exit path can leak it."""
|
||||
path = f"/dev/shm/vllm_offload_{iid}.mmap"
|
||||
seen_at_barrier = []
|
||||
|
||||
@@ -1001,8 +992,7 @@ def test_backing_file_unlinked_after_barrier(iid):
|
||||
|
||||
def test_barrier_failure_unlinks_creator_and_raises(iid):
|
||||
"""A failed rendezvous must remove the creator's file and re-raise, not
|
||||
leave a stub that wedges the next start in _wait_for_file_size.
|
||||
"""
|
||||
leave a stub that wedges the next start in _wait_for_file_size."""
|
||||
path = f"/dev/shm/vllm_offload_{iid}.mmap"
|
||||
with pytest.raises(RuntimeError, match="peer died"):
|
||||
_make_region(iid, barrier=MagicMock(side_effect=RuntimeError("peer died")))
|
||||
@@ -1012,8 +1002,7 @@ def test_barrier_failure_unlinks_creator_and_raises(iid):
|
||||
@pytest.mark.parametrize("replicated", [False, True])
|
||||
def test_mp_barrier_unlinks_file_and_survives_sigkill(iid, replicated):
|
||||
"""With a real cross-process barrier every worker maps one shared inode,
|
||||
the name is gone while they run, and SIGKILL leaks nothing.
|
||||
"""
|
||||
the name is gone while they run, and SIGKILL leaks nothing."""
|
||||
num_workers = 2
|
||||
path = f"/dev/shm/vllm_offload_{iid}.mmap"
|
||||
ctx = get_mp_context()
|
||||
@@ -1052,8 +1041,7 @@ def test_mp_barrier_unlinks_file_and_survives_sigkill(iid, replicated):
|
||||
@pytest.mark.parametrize("failure_phase", ["shm", "memory", "population"])
|
||||
def test_setup_failure_before_barrier_releases_peers(iid, monkeypatch, failure_phase):
|
||||
"""A worker that dies before the rendezvous must still arrive at the
|
||||
barrier, or its peers block in the collective until it times out.
|
||||
"""
|
||||
barrier, or its peers block in the collective until it times out."""
|
||||
failure = MagicMock(side_effect=ValueError("Insufficient space"))
|
||||
memory_check = failure if failure_phase == "memory" else None
|
||||
if failure_phase == "shm":
|
||||
@@ -1084,8 +1072,7 @@ def test_setup_failure_before_barrier_releases_peers(iid, monkeypatch, failure_p
|
||||
|
||||
def test_mmap_failure_unlinks_creator_before_releasing_peers(iid, monkeypatch):
|
||||
"""A creator that fails after sizing the file must drop it before arriving
|
||||
at the barrier, so the next start does not land on a stale file.
|
||||
"""
|
||||
at the barrier, so the next start does not land on a stale file."""
|
||||
path = f"/dev/shm/vllm_offload_{iid}.mmap"
|
||||
monkeypatch.setattr("mmap.mmap", MagicMock(side_effect=OSError("mmap")))
|
||||
seen_at_barrier = []
|
||||
@@ -1098,8 +1085,7 @@ def test_mmap_failure_unlinks_creator_before_releasing_peers(iid, monkeypatch):
|
||||
|
||||
def test_barrier_release_failure_keeps_original_error(iid, monkeypatch):
|
||||
"""When releasing the peers fails too, the setup error that explains the
|
||||
failure must be the one that propagates.
|
||||
"""
|
||||
failure must be the one that propagates."""
|
||||
import vllm.v1.kv_offload.cpu.shared_offload_region as region
|
||||
|
||||
monkeypatch.setattr(
|
||||
|
||||
Reference in New Issue
Block a user