[CI] Disable JIT warmup by default in VllmRunner (#58452)
Signed-off-by: mgoin <mgoin64@gmail.com>
Misha Goin committed
Sep 23, 2026 at 19:53 UTC
89d9fd83e2888405bea16ac36bea0045e352a074
1 file changed
+4
tests/conftest.py
+4
@@ -982,6 +982,8 @@ class VllmRunner:
982
- `enable_chunked_prefill`: Set to `False` instead of `None` for
983
test reproducibility.
984
- `enforce_eager`: Set to `False` to test CUDA graph.
985
+ - `kernel_config.enable_jit_warmup`: Set to `False` to reduce test startup
986
+ time.
987
"""
988
989
def __init__(
@@ -1010,6 +1012,8 @@ class VllmRunner:
1012
else set_default_torch_num_threads(default_torch_num_threads)
1013
)
1014
1015
+ kwargs.setdefault("kernel_config", {"enable_jit_warmup": False})
1016
+
1017
if not kwargs.get("compilation_config", None):
1018
# Note(@tdoublep): This is set to 4 because some tests (e.g., hybrid
1019
# model tests) may set max_num_seqs=4. If min cudagraph_capture_size is