[Bugfix] Fix layerwise reload dropping params after a composed weight loader (#44814)
Signed-off-by: hallerite <git@hallerite.com> Co-authored-by: Claude Opus 4.8 (1M context) <noreply@anthropic.com> Co-authored-by: Kyle Sayers <kylesayrs@gmail.com>
This commit is contained in:
co-authored by
Claude Opus 4.8
Kyle Sayers
parent
6850839c6f
commit
c9e5bf8135
@@ -185,4 +185,18 @@ def get_numel_loaded(
|
||||
"""
|
||||
with CopyCounter() as counter:
|
||||
return_value = weight_loader(*args.args, **args.kwargs)
|
||||
return counter.copied_numel, return_value
|
||||
|
||||
# A weight loader fills a single destination parameter, so the number of
|
||||
# loaded elements is at most that parameter's size. Some loaders copy into
|
||||
# the parameter more than once -- e.g. ``composed_weight_loader`` runs an
|
||||
# in-place post-load transform (``param.copy_(fn(param))``) on top of the
|
||||
# initial copy -- which would make CopyCounter report twice the parameter
|
||||
# size. Over-counting inflates the layer's loaded-element total and can
|
||||
# finalize the layer before every parameter is loaded, silently dropping
|
||||
# the trailing parameter(s) (e.g. Mamba ``mixer.D``). Cap the count at the
|
||||
# destination size to keep the per-layer accounting correct.
|
||||
numel = counter.copied_numel
|
||||
param = args.arguments.get("param", None)
|
||||
if isinstance(param, torch.Tensor):
|
||||
numel = min(numel, param.numel())
|
||||
return numel, return_value
|
||||
|
||||
Reference in New Issue
Block a user