Skip to content

Commit

Permalink
Logging KT's key order warning only once (#2548)
Browse files Browse the repository at this point in the history
Summary:

this warning is very noisy, changing to print it only once instead of every time

Reviewed By: TroyGarden

Differential Revision: D65700079
  • Loading branch information
gag1jain authored and facebook-github-bot committed Nov 12, 2024
1 parent a97cf28 commit 16f3541
Showing 1 changed file with 11 additions and 4 deletions.
15 changes: 11 additions & 4 deletions torchrec/sparse/jagged_tensor.py
Original file line number Diff line number Diff line change
Expand Up @@ -3380,14 +3380,21 @@ def _kt_unflatten(
return KeyedTensor(context[0], context[1], values[0])


print_flatten_spec_warn = True


def _kt_flatten_spec(kt: KeyedTensor, spec: TreeSpec) -> List[torch.Tensor]:
_keys, _length_per_key = spec.context
# please read https://fburl.com/workplace/8bei5iju for more context,
# you can also consider use short_circuit_pytree_ebc_regroup with KTRegroupAsDict
logger.warning(
"KT's key order might change from spec from the torch.export, this could have perf impact. "
f"{kt.keys()} vs {_keys}"
)
global print_flatten_spec_warn
if print_flatten_spec_warn:
logger.warning(
"KT's key order might change from spec from the torch.export, this could have perf impact. "
f"{kt.keys()} vs {_keys}"
)
print_flatten_spec_warn = False

res = permute_multi_embedding([kt], [_keys])
return [res[0]]

Expand Down

0 comments on commit 16f3541

Please sign in to comment.