[Feat] runtime: Include FlashInfer’s default tactic in startup tuning. (#43927)

Co-authored-by: Po-Han Huang <pohanh@nvidia.com>
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
This commit is contained in:
Lee Nau
2026-10-15 15:30:45 +08:00
committed by GitHub
co-authored by Po-Han Huang Claude Opus 5.5
parent a776851c97
commit 33618d4f00
2 changed files with 4 additions and 0 deletions
@@ -263,6 +263,8 @@ def _drop_diverged_autotune_cache(
@contextlib.contextmanager
def flashinfer_autotune_context(model_runner: ModelRunner, *, run_lm_head: bool):
from flashinfer import autotune_v2
# The gate below decides on the same inputs load_configs does.
from flashinfer.autotuner import AutoTuner, _collect_metadata, autotune
@@ -331,6 +333,7 @@ def flashinfer_autotune_context(model_runner: ModelRunner, *, run_lm_head: bool)
cache=None if reuse_cache else str(autotune_cache),
skip_ops=skip_ops,
),
autotune_v2(mode="tune", persistent_cache=False),
mega_context,
autotune_dummy_run_mode(run_lm_head=run_lm_head),
):
@@ -80,6 +80,7 @@ class TestMegaMoEAutotuneStartup(CustomTestCase):
)
general = Mock(side_effect=lambda *_args, **_kwargs: nullcontext())
modules = {
"flashinfer": SimpleNamespace(autotune_v2=lambda **_: nullcontext()),
"flashinfer.autotuner": SimpleNamespace(
AutoTuner=SimpleNamespace(
get=lambda: SimpleNamespace(