Skip to content

Make tune_chunk_size a top-level config field#210

Open
GMNGeoffrey wants to merge 1 commit into
aqlaboratory:mainfrom
GMNGeoffrey:chunk-tuner-top-level-config
Open

Make tune_chunk_size a top-level config field#210
GMNGeoffrey wants to merge 1 commit into
aqlaboratory:mainfrom
GMNGeoffrey:chunk-tuner-top-level-config

Conversation

@GMNGeoffrey
Copy link
Copy Markdown
Contributor

Summary
Currently tune_chunk_size is a field reference used in various subconfigs. Setting it in any of htem changes the value in all of them. I think given that, it makes sense to have one obviously canonical top-level place to set it. Otherwise you have to pick one arbitrary module to set it in, and I think having done that it's surprising that it gets set everywhere. This matches the behavior of similar fields per_sample_token_cutoff, per_sample_atom_cutoff, and low_mem_validation.

An alternative would be to unlink the module-level settings so they could be set independently (and optionally having a top-level setting that controls the default).

Changes
Add tune_chunk_size to settings.memory.eval.

Related Issues
None

Testing
Ran examples manually using the setting here or in any one of the modules and confirmed chunk tuning is turned off everywhere.

Setting this in any location sets it for all locations because it's a
field reference, so I think it makes sense to have this settable at the
top-level along with chunk size. Otherwise you have to pick one
arbitrary module to set it in, and I think having done that it's
surprising that it gets set everywhere. This matches the behavior of
similar fields `per_sample_token_cutoff`, `per_sample_atom_cutoff`, and
`low_mem_validation`.

An alternative would be to unlink the module-level settings so they
could be set independently (and optionally having a top-level setting
that controls the default).
@GMNGeoffrey
Copy link
Copy Markdown
Contributor Author

@christinaflo

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant