[Refactor] Clean up parallel runtime comments (#40632)

This commit is contained in:
Cheng Wan
2026-09-21 14:32:22 -07:00
committed by GitHub
parent f532ad1f9a
commit acac4dd9d9
63 changed files with 199 additions and 1122 deletions
@@ -93,8 +93,6 @@ def _make_prefill_aware_swa_runner(
page_size=1,
attn_cp_size=1,
tp_size=1,
# The backend still reads the runner's frozen record for these two;
# same single-rank placement, stated where it looks for it.
ps=SimpleNamespace(attn_cp_size=1, tp_size=1),
is_draft_worker=False,
server_args=server_args,
@@ -33,12 +33,7 @@ GROUP_SIZE = 32 # MXFP4 block size
@pytest.fixture
def stated_tp_group():
"""A TP group for a test that runs in a process without one.
The production call passes the group *into* `use_symmetric_memory`, so
stubbing that context manager does not stop the read -- the argument is
evaluated first. Stating it on the context answers every spelling.
"""
"""Provide a TP-group placeholder for kernels with mocked symmetric memory."""
from sglang.srt.runtime_context import get_parallel
with get_parallel().override(tp_group=None):
@@ -19,12 +19,7 @@ register_cuda_ci(est_time=14, stage="base-b", runner_config="1-gpu-small")
@pytest.fixture
def stated_tp_group():
"""A TP group for a test that runs in a process without one.
The production call passes the group *into* `use_symmetric_memory`, so
stubbing that context manager does not stop the read -- the argument is
evaluated first. Stating it on the context answers every spelling.
"""
"""Provide a TP-group placeholder for kernels with mocked symmetric memory."""
from sglang.srt.runtime_context import get_parallel
with get_parallel().override(tp_group=None):
@@ -65,12 +65,7 @@ GROUP_SIZE = 32 # MXFP4 block size
@pytest.fixture
def stated_tp_group():
"""A TP group for a test that runs in a process without one.
The production call passes the group *into* `use_symmetric_memory`, so
stubbing that context manager does not stop the read -- the argument is
evaluated first. Stating it on the context answers every spelling.
"""
"""Provide a TP-group placeholder for kernels with mocked symmetric memory."""
from sglang.srt.runtime_context import get_parallel
with get_parallel().override(tp_group=None):