MG Python API#1307
Conversation
viclafargue
added
improvement
| index->dtype.code = kDLFloat; | ||
| index->dtype.bits = 32; | ||
| index->addr = reinterpret_cast<uintptr_t>(_mg_distribute<float>(res, filename)); |
There was a problem hiding this comment.
Is this change intentional? (do we only support float32 here, and you meant to remove support for the other dtypes?)
There was a problem hiding this comment.
This functions deserializes a single-GPU index and distributes it across multiple GPUs. The IVF-PQ index does not have a T template parameter, only IdxT (which is always int64_t). Because of this, a file containing a serialized IVF-PQ index is not prepended with an index type (float, half, uint8_t, int8_t). This makes it impossible to find the index type necessary to instantiate a multi-GPU index. This is fundamentally an issue in the multi-GPU API that should be remediated. In the meantime, we should assume that for this specific case (the replication of a single-GPU IVF-PQ index), the user is planning to search with floats.
|
lgtm! |
|
/merge |
rapids-bot
Bot
merged commit d02e0fc
into
rapidsai:branch-25.10
github-project-automation
Bot
moved this from In Progress
to Done
in Unstructured Data Processing
3 participants
No description provided.