Break reference cycle in load_state_dict #20397
Conversation
load_state_dict includes a recursive inner function `load` that captures Tensors through the close-over variable `state_dict`. Because it's recursive, it also captures itself leading to a reference cycle. This breaks the reference cycle so that any Tensors in state_dict can be collected immediately instead of waiting until the next GC cycle. Alternatively, we could have passed `state_dict` and `metadata` as arguments to load to prevent capture of Tensors. (That would still result in cyclic garbage, but not any cyclic garbage of Tensors). See: pytorch#20199 (comment)
|
Can you plz explain would there be cyclic garbage even when passing all the variables as arguments? |
@Stonesjtu because the function You can see this here: foo.py vs foo2.py. Anyways, cyclic-garbage isn't a problem in general unless it involves a Tensor. |
|
@colesbury , thanks! Should we add documentation to advise users against using Tensors as closure in recursive function? |
facebook-github-bot
left a comment
facebook-github-bot
left a comment
There was a problem hiding this comment.
@colesbury has imported this pull request. If you are a Facebook employee, you can view this diff on Phabricator.
|
@colesbury merged this pull request in c1fa449. |
6 participants
load_state_dict includes a recursive inner function
loadthat capturesTensors through the close-over variable
state_dict. Because it'srecursive, it also captures itself leading to a reference cycle.
This breaks the reference cycle so that any Tensors in state_dict can be
collected immediately instead of waiting until the next GC cycle.
Alternatively, we could have passed
state_dictandmetadataasarguments to load to prevent capture of Tensors. (That would still
result in cyclic garbage, but not any cyclic garbage of Tensors).
See:
#20199 (comment)