diff --git a/scripts/autotune_helion_kernels.py b/scripts/autotune_helion_kernels.py index c02d2a0206..864aabfc09 100644 --- a/scripts/autotune_helion_kernels.py +++ b/scripts/autotune_helion_kernels.py @@ -37,6 +37,7 @@ try: get_kernel_by_name, get_registered_kernels, ) + from vllm.kernels.helion.ops import import_all_kernels from vllm.kernels.helion.utils import get_canonical_gpu_name from vllm.logger import init_logger from vllm.utils.import_utils import has_helion @@ -382,6 +383,9 @@ def main(): args = parser.parse_args() + # import all helion kernels + import_all_kernels() + import logging if args.verbose: diff --git a/vllm/kernels/helion/__init__.py b/vllm/kernels/helion/__init__.py index 8c05c428bb..8a20c9eb9e 100644 --- a/vllm/kernels/helion/__init__.py +++ b/vllm/kernels/helion/__init__.py @@ -2,7 +2,6 @@ # SPDX-FileCopyrightText: Copyright contributors to the vLLM project """Helion integration for vLLM.""" -import vllm.kernels.helion.ops # noqa: F401 Auto-register all Helion ops from vllm.kernels.helion.case_key import CaseKey from vllm.kernels.helion.config_manager import ( ConfigManager, diff --git a/vllm/kernels/helion/ops/__init__.py b/vllm/kernels/helion/ops/__init__.py index eacd483bbb..588c070a12 100644 --- a/vllm/kernels/helion/ops/__init__.py +++ b/vllm/kernels/helion/ops/__init__.py @@ -1,11 +1,33 @@ # SPDX-License-Identifier: Apache-2.0 # SPDX-FileCopyrightText: Copyright contributors to the vLLM project -"""Auto-import all Helion op modules to trigger kernel registration.""" +"""Helion kernel implementation. + +Importing this package does NOT register any kernels. Runtime code imports the +specific op module it needs, e.g.:: + + from vllm.kernels.helion.ops import scaled_mm # noqa: F401 + +which triggers that op's ``@register_kernel`` as an import side effect. + +Tools that need the full registry (e.g. scripts/autotune_helion_kernels.py) +call ``import_all_ops()`` to force every op module to register. +""" import importlib import pkgutil -# Automatically import all submodules so that @register_kernel -# decorators execute and register ops with torch.ops.vllm_helion. -for _module_info in pkgutil.iter_modules(__path__): - importlib.import_module(f"{__name__}.{_module_info.name}") + +def import_all_kernels() -> list[str]: + """Import every kernel submodule so all ``@register_kernel`` decorators run. + + Returns: + The fully-qualified module names that were imported. + """ + imported: list[str] = [] + for module_info in pkgutil.iter_modules(__path__): + if module_info.ispkg: + continue + module_name = f"{__name__}.{module_info.name}" + importlib.import_module(module_name) + imported.append(module_name) + return imported