Add pydocstyle to the ruff rules (#52136)
Signed-off-by: Harry Mellor <19981378+hmellor@users.noreply.github.com> Signed-off-by: Harry Mellor <hej.mellor@gmail.com> Co-authored-by: Claude Opus 5 (1M context) <noreply@anthropic.com>
This commit is contained in:
co-authored by
Claude Opus 5
parent
804e5377aa
commit
8c1557a79c
@@ -493,7 +493,6 @@ async def test_dp_pause_abort(expert_parallel: bool):
|
||||
@pytest.mark.parametrize("expert_parallel", [False, True])
|
||||
async def test_dp_pause_keep_then_resume(expert_parallel: bool):
|
||||
"""Start generation, pause after a few tokens (keep mode), resume; verify gap."""
|
||||
|
||||
pause_duration = 2.0
|
||||
min_tokens_before_pause = 3
|
||||
|
||||
@@ -548,7 +547,8 @@ async def test_dp_pause_keep_then_resume(expert_parallel: bool):
|
||||
async def test_dp_pause_keep_race_staggered_engines():
|
||||
"""Race: send pause(keep) to engine 0, then add two requests,
|
||||
then pause(keep) to engine 1. Ensures no deadlock when pause
|
||||
requests are staggered and requests arrive in between."""
|
||||
requests are staggered and requests arrive in between.
|
||||
"""
|
||||
if DP_SIZE != 2:
|
||||
pytest.skip("test_dp_pause_keep_race_staggered_engines requires DP_SIZE=2")
|
||||
|
||||
@@ -607,8 +607,7 @@ async def test_dp_pause_keep_race_staggered_engines():
|
||||
|
||||
@pytest.mark.asyncio
|
||||
async def test_dp_pause_barrier_request_deadlock():
|
||||
"""
|
||||
Test that start_dp_wave is ignored while paused.
|
||||
"""Test that start_dp_wave is ignored while paused.
|
||||
|
||||
Sequence:
|
||||
1. Pause all engines (PAUSED_ALL).
|
||||
@@ -719,7 +718,8 @@ async def test_dp_pause_barrier_request_deadlock():
|
||||
async def test_dp_pause_wait_mode_drains_in_flight():
|
||||
"""mode="wait" through the DP consensus path: the pause lets the
|
||||
in-flight request run to completion while the consensus is pending,
|
||||
resolves only once drained, and the engine generates again on resume."""
|
||||
resolves only once drained, and the engine generates again on resume.
|
||||
"""
|
||||
with ExitStack() as after:
|
||||
engine = AsyncLLM.from_engine_args(_get_dp_pause_engine_args(True))
|
||||
after.callback(engine.shutdown)
|
||||
@@ -750,7 +750,8 @@ async def test_dp_pause_wait_mode_drains_in_flight():
|
||||
async def test_dp_pause_while_asleep():
|
||||
"""Pausing a sleeping DP engine: sleeping ranks skip dummy batches yet
|
||||
must still reach the pause consensus, and the completion barrier must
|
||||
be harmless on workers whose memory is unmapped."""
|
||||
be harmless on workers whose memory is unmapped.
|
||||
"""
|
||||
with ExitStack() as after:
|
||||
engine = AsyncLLM.from_engine_args(_get_dp_pause_engine_args(True))
|
||||
after.callback(engine.shutdown)
|
||||
@@ -786,7 +787,8 @@ async def test_dp_pause_completion_implies_device_idle(
|
||||
"""A resolved pause future promises an idle device. Inflate the dummy
|
||||
batches the pause consensus manufactures so any work the pause fails to
|
||||
wait on stays visible, then require a quiet worker stream the moment
|
||||
pause_generation() returns."""
|
||||
pause_generation() returns.
|
||||
"""
|
||||
with ExitStack() as after:
|
||||
# The probes below ship functions through collective_rpc.
|
||||
monkeypatch.setenv("VLLM_ALLOW_INSECURE_SERIALIZATION", "1")
|
||||
|
||||
@@ -1,7 +1,6 @@
|
||||
# SPDX-License-Identifier: Apache-2.0
|
||||
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
|
||||
"""
|
||||
Test Dual Batch Overlap (DBO) with Data Parallelism + Expert Parallelism.
|
||||
"""Test Dual Batch Overlap (DBO) with Data Parallelism + Expert Parallelism.
|
||||
|
||||
DBO is specifically designed for DP+EP scenarios to hide communication latency
|
||||
by overlapping computation of two batches. This test validates that DBO works
|
||||
@@ -54,9 +53,7 @@ DEEPEP_BACKENDS = [
|
||||
),
|
||||
)
|
||||
def test_dbo_dp_ep_gsm8k(all2all_backend: str, num_gpus_available):
|
||||
"""
|
||||
Test DBO with DP+EP using GSM8K evaluation.
|
||||
"""
|
||||
"""Test DBO with DP+EP using GSM8K evaluation."""
|
||||
lm_eval = pytest.importorskip("lm_eval")
|
||||
required_gpus = DP_SIZE
|
||||
|
||||
|
||||
@@ -24,7 +24,8 @@ TP_SIZE = int(os.getenv("TP_SIZE", "1"))
|
||||
|
||||
class ExternalLBServerManager:
|
||||
"""Manages data parallel vLLM server instances for external
|
||||
load balancer testing."""
|
||||
load balancer testing.
|
||||
"""
|
||||
|
||||
def __init__(
|
||||
self,
|
||||
|
||||
@@ -30,7 +30,8 @@ DP_SIZE_LOCAL = DP_SIZE // NUM_NODES # 2 ranks per node
|
||||
class HybridLBServerManager:
|
||||
"""Manages hybrid data parallel vLLM server instances where each node
|
||||
runs a single logical API server that balances requests only to the
|
||||
DP engines running on that same node."""
|
||||
DP engines running on that same node.
|
||||
"""
|
||||
|
||||
def __init__(
|
||||
self,
|
||||
|
||||
@@ -107,7 +107,8 @@ async def _run_request_bursts(
|
||||
|
||||
class MultinodeInternalLBServerManager:
|
||||
"""Manages multi-node data parallel vLLM server instances for internal
|
||||
load balancer testing using --headless mode."""
|
||||
load balancer testing using --headless mode.
|
||||
"""
|
||||
|
||||
def __init__(
|
||||
self,
|
||||
@@ -241,7 +242,8 @@ class MultinodeInternalLBServerManager:
|
||||
|
||||
class APIOnlyServerManager:
|
||||
"""Manages API-only server (Node 0) and headless engines server (Node 1)
|
||||
for testing separated API server and engine configuration."""
|
||||
for testing separated API server and engine configuration.
|
||||
"""
|
||||
|
||||
def __init__(
|
||||
self,
|
||||
@@ -261,7 +263,6 @@ class APIOnlyServerManager:
|
||||
|
||||
def __enter__(self) -> list[tuple[RemoteOpenAIServer, list[str]]]:
|
||||
"""Start API-only server and headless engines server."""
|
||||
|
||||
# Start API-only server (Node 0) - no engines, only API server
|
||||
api_server_args = self.base_server_args.copy()
|
||||
api_server_args.extend(
|
||||
@@ -607,7 +608,6 @@ async def test_api_only_multinode_dp_completion(
|
||||
model_name: str,
|
||||
) -> None:
|
||||
"""Test API-only server with all engines on separate headless server."""
|
||||
|
||||
# Test single request
|
||||
result = await _make_completion_request(api_only_client, model_name)
|
||||
assert result is not None
|
||||
@@ -645,7 +645,8 @@ async def test_api_only_multinode_dp_completion_streaming(
|
||||
model_name: str,
|
||||
) -> None:
|
||||
"""Test API-only server streaming with all engines on separate
|
||||
headless server."""
|
||||
headless server.
|
||||
"""
|
||||
prompt = "What is an LLM?"
|
||||
|
||||
async def make_streaming_request():
|
||||
|
||||
@@ -82,7 +82,8 @@ async def _generate(engine: AsyncLLM, prompt: str, max_tokens: int) -> int:
|
||||
@pytest.mark.parametrize("async_scheduling", [True, False])
|
||||
async def test_pp_dp_v2_low_concurrency(async_scheduling: bool):
|
||||
"""A single in-flight request at a time, repeated, to exercise the
|
||||
PP slot ring under empty batches between decodes."""
|
||||
PP slot ring under empty batches between decodes.
|
||||
"""
|
||||
with ExitStack() as after:
|
||||
engine = AsyncLLM.from_engine_args(_engine_args(async_scheduling))
|
||||
after.callback(engine.shutdown)
|
||||
@@ -96,7 +97,8 @@ async def test_pp_dp_v2_low_concurrency(async_scheduling: bool):
|
||||
@pytest.mark.parametrize("async_scheduling", [True, False])
|
||||
async def test_pp_dp_v2_mid_concurrency(async_scheduling: bool):
|
||||
"""64 concurrent requests, staggered, to exercise the steady-state
|
||||
DP all-reduce + PP slot-ring path."""
|
||||
DP all-reduce + PP slot-ring path.
|
||||
"""
|
||||
with ExitStack() as after:
|
||||
engine = AsyncLLM.from_engine_args(_engine_args(async_scheduling))
|
||||
after.callback(engine.shutdown)
|
||||
@@ -112,8 +114,8 @@ async def test_pp_dp_v2_mid_concurrency(async_scheduling: bool):
|
||||
@pytest.mark.asyncio
|
||||
async def test_pp_dp_v2_abort_mid_decode():
|
||||
"""Cancel half the in-flight requests mid-stream and confirm the
|
||||
engine survives the abort storm."""
|
||||
|
||||
engine survives the abort storm.
|
||||
"""
|
||||
with ExitStack() as after:
|
||||
engine = AsyncLLM.from_engine_args(_engine_args(async_scheduling=True))
|
||||
after.callback(engine.shutdown)
|
||||
@@ -153,8 +155,8 @@ async def test_pp_dp_v2_abort_mid_decode():
|
||||
@pytest.mark.asyncio
|
||||
async def test_pp_dp_v2_pause_resume():
|
||||
"""Pause an engine with a request in flight, then resume and confirm
|
||||
new requests still work."""
|
||||
|
||||
new requests still work.
|
||||
"""
|
||||
with ExitStack() as after:
|
||||
engine = AsyncLLM.from_engine_args(_engine_args(async_scheduling=True))
|
||||
after.callback(engine.shutdown)
|
||||
|
||||
Reference in New Issue
Block a user