A high-throughput and memory-efficient inference and serving engine for LLMs
Mirrored from https://github.com/tduncan-lum/Nvidia-Optimization-Stack---VLLM.git
Synced Nov 14, 05:30 UTC
Revert "[Bug] Dynamo Unsupported due to `BasevLLMParameter.torch_function` calling disabled super()" (#25681)
Signed-off-by: yewentao256 <zhyanwentao@126.com> Co-authored-by: Wentao Ye <44945378+yewentao256@users.noreply.github.com>
Michael Goin committed
Sep 25, 2025 at 12:45 UTC
916bd9204d4f895d7a6d93199d87a3fbe3e1546d
1 file changed
-10
vllm/model_executor/parameter.py
-10
@@ -12,7 +12,6 @@ from torch.nn import Parameter
12
from vllm.distributed import (get_tensor_model_parallel_rank,
13
get_tensor_model_parallel_world_size)
14
from vllm.logger import init_logger
15
-from vllm.utils import is_torch_equal_or_newer
15
16
__all__ = [
17
"BasevLLMParameter", "PackedvLLMParameter", "PerTensorScaleParameter",
@@ -115,15 +114,6 @@ class BasevLLMParameter(Parameter):
114
115
@classmethod
116
def __torch_function__(cls, func, types, args=(), kwargs=None):
118
- if not is_torch_equal_or_newer("2.8.0"):
119
- logger.warning_once(
120
- "Torch %s detected (<2.8.0): returning NotImplemented in "
121
- "BasevLLMParameter.__torch_function__ to avoid potential "
122
- "TorchDynamo issues.",
123
- torch.__version__,
124
- )
125
- return NotImplemented
126
-
117
if kwargs is None:
118
kwargs = {}
119
return super().__torch_function__(func, types, args, kwargs)