[Pytorch Mobile] Remove caching (in code) of interned strings #50390
Add this suggestion to a batch that can be applied as a single commit.
This suggestion is invalid because no changes were made to the code.
Suggestions cannot be applied while the pull request is closed.
Suggestions cannot be applied while viewing a subset of changes.
Only one suggestion per line can be applied in a batch.
Add this suggestion to a batch that can be applied as a single commit.
Applying suggestions on deleted lines is not supported.
You must change the existing code in this line in order to create a valid suggestion.
Outdated suggestions cannot be applied.
This suggestion has been applied or marked resolved.
Suggestions cannot be applied from pending reviews.
Suggestions cannot be applied on multi-line comments.
Suggestions cannot be applied while the pull request is queued to merge.
Suggestion cannot be applied right now. Please check back later.
Stack from ghstack:
Currently, there is a massive switch/case statement that is generated in the
InternedStrings::string()
method to speed up Symbol -> string conversion without taking a mock (mutex). The relative call rate of this on mobile is insignificant, so unlikely to have any material impact on runtime even if the lookups happen under a lock. Plus, parallelism is almost absent on mobile, which is where locks/mutexes cause the most problem (taking a mutex without contention is usually very fast and just adds a memory barrier iirc).The only impact that caching interned strings has is avoiding taking a lock when interned strings are looked up. They are not looked up very often during training, and based on basic testing, they don't seem to be looked up much during inference either.
During training, the following strings were looked up at test startup:
Command used to trigger training:
buck test fbsource//xplat/papaya/client/executor/torch/store/transform/feature/test:test
During inference, the only symbol that was looked up was
tolist
.Differential Revision: D25861786