diff --git a/tests/conftest.py b/tests/conftest.py index cd1d11917b..45c629b875 100644 --- a/tests/conftest.py +++ b/tests/conftest.py @@ -982,6 +982,8 @@ class VllmRunner: - `enable_chunked_prefill`: Set to `False` instead of `None` for test reproducibility. - `enforce_eager`: Set to `False` to test CUDA graph. + - `kernel_config.enable_jit_warmup`: Set to `False` to reduce test startup + time. """ def __init__( @@ -1010,6 +1012,8 @@ class VllmRunner: else set_default_torch_num_threads(default_torch_num_threads) ) + kwargs.setdefault("kernel_config", {"enable_jit_warmup": False}) + if not kwargs.get("compilation_config", None): # Note(@tdoublep): This is set to 4 because some tests (e.g., hybrid # model tests) may set max_num_seqs=4. If min cudagraph_capture_size is