feat(runtime): update to v1 API for leanvec ood #262
Conversation
| virtual ~LeanVecTrainingData(); | ||
|
|
||
| /* Build LeanVec training data (compression matrices) from the provided | ||
| * data. |
There was a problem hiding this comment.
Should we call "Leanvec transformation matrices" instead of training data?
There was a problem hiding this comment.
LeanVecTrainingData was purposefully chosen to be generic. In hindsight, I think we could have been more specific. But changing this would require an API update and therefore conflicts with your suggestion in the other comment.
| size_t n, | ||
| const float* x, | ||
| size_t n_train, | ||
| const float* q, |
There was a problem hiding this comment.
Can we keep these two new arguments in the end and default initialize to 0 and nullptr? In that case, the older v0 calls to build will still work without any modifications. Also "n_train" is kind of confusing as both data/queries are used for training? How about we explicitly say "n_queries"?
There was a problem hiding this comment.
That's exactly the discussion I wanted to have. Your suggestion would indeed allow us to stick with v0 (only that we'd need a copy of the function and we can't use default values, due to ABI compability, but that's just a detail).
In my opinion, the order (n_data, const float* data, n_query, const float* queries, size_t leanvec_dims) just makes more sense than (n_data, const float* data, size_t leanvec_dims, n_query, const float* queries). But does it justify bumping to v0?
Your preference would be to stick with v0 for as long as possible?
Increments runtime API version to
v1, becauseLeanVecTrainingData::build()changed significantly for OOD suport.Unchanged entities from
v0are aliased inv1.Successful runtime lib build pending changes from svs prebuilt binaries.