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 17, 01:45 UTC
[Misc] remove old comments (#4866)
youkaichao committed
May 16, 2024 at 11:07 UTC
10fa9eea21ae757d17c1369afa6172598db3be92
1 file changed
-10
vllm/worker/model_runner.py
-10
@@ -887,16 +887,6 @@ class ModelRunner:
887
# This usually takes < 10 seconds.
888
logger.info("Graph capturing finished in %.0f secs.", elapsed_time)
889
890
- def __del__(self) -> None:
891
- # Delete the CUDA graphs before deleting the pynccl communicator.
892
- # NOTE(woosuk): This is necessary because otherwise deadlocks can
893
- # happen.
894
- # FIXME(woosuk): This is a bit hacky. Find a more robust solution.
895
- # TODO(youkaichao): when we get enough user feedback that pynccl is
896
- # more stable than cupy, we can remove this, e.g. in v0.4.1.
897
- self.graph_runners.clear()
898
- self.pynccl_backend = None
899
-
890
@property
891
def vocab_size(self) -> int:
892
return self.model_config.get_vocab_size()