[checkpointio] fix zero optimizer async save memory (#6151)

* [checkpointio] fix zero optimizer async save memory

* [checkpointio] fit new tensornvme api

* [checkpointio] fit new tensornvme api
This commit is contained in:
Hongxin Liu
2024-11-25 14:46:31 +08:00
committed by GitHub
parent 8ecff0cb7f
commit ab856fd308
7 changed files with 57 additions and 42 deletions

View File

@@ -56,7 +56,7 @@ class GeneralCheckpointIO(CheckpointIO):
if use_async:
from tensornvme.async_file_io import AsyncFileWriter
writer = AsyncFileWriter(open(checkpoint, "wb", buffering=0), self.N_WRITE_ENTRIES, backend="pthread")
writer = AsyncFileWriter(checkpoint, self.N_WRITE_ENTRIES, backend="pthread")
if id(model) not in self.pinned_state_dicts:
self.pinned_state_dicts[id(model)] = create_pinned_state_dict(state_dict)
self.async_writers.append(writer)