From 365cc2ade5c915aae3ce0bc2179125d8aab017d0 Mon Sep 17 00:00:00 2001 From: Alison Shao <54658187+alisonshao@users.noreply.github.com> Date: Tue, 2 Jun 2026 14:33:52 -0700 Subject: [PATCH] jit_kernel tests: bump multiprocess_test timeout 90s -> 240s (cold JIT cache) (#26994) --- python/sglang/jit_kernel/tests/utils.py | 12 ++++++++++-- 1 file changed, 10 insertions(+), 2 deletions(-) diff --git a/python/sglang/jit_kernel/tests/utils.py b/python/sglang/jit_kernel/tests/utils.py index 56f3511ab..6560308fd 100644 --- a/python/sglang/jit_kernel/tests/utils.py +++ b/python/sglang/jit_kernel/tests/utils.py @@ -6,8 +6,16 @@ from typing import Callable import pytest -def multiprocess_test(file: str, nproc: int, timeout: int = 90) -> None: - """Launch this script as a torchrun worker and assert success.""" +def multiprocess_test(file: str, nproc: int, timeout: int = 240) -> None: + """Launch this script as a torchrun worker and assert success. + + The default budget covers the cold-cache first invocation, where the + worker pays the full triton + cutlass JIT compile cost (60-180s observed + on H200). The previous 90s default tripped intermittently on the first + parametrisation of `test_tp_qknorm` (seen on `main` runs too, not only + on fresh-venv PRs); subsequent parametrisations finished in ~60s once + the JIT cache was warm. + """ cmd = [ "torchrun", f"--nproc_per_node={nproc}",