Create int bfloat16 vector formats#136627
Conversation
elasticsearchmachine
added
v9.3.0
Team:Search Relevance
|
Pinging @elastic/es-search-relevance (Team:Search Relevance) |
server/src/main/java/org/elasticsearch/index/codec/vectors/es93/ES93Int8FlatVectorFormat.java
Outdated
Show resolved
Hide resolved
thecoop
force-pushed
the
int-hnsw-bfloat16
branch
from
66172e3 to
1287a0b
Compare
...r/src/main/java/org/elasticsearch/index/codec/vectors/ES814ScalarQuantizedVectorsFormat.java
Show resolved
Hide resolved
benwtrent
left a comment
benwtrent
left a comment
There was a problem hiding this comment.
I need to remember why I started that Flat vs Format pattern for bit vectors that you have continued here, I am not sure its actually warranted.
...in/java/org/elasticsearch/index/codec/vectors/es93/ES93ScalarQuantizedFlatVectorsFormat.java
Outdated
Show resolved
Hide resolved
This reverts commit 064392f.
|
I've removed the intermediate format, and pushed everything into the top-level formats |
john-wagster
left a comment
john-wagster
left a comment
There was a problem hiding this comment.
this lgtm. It did pop into my head that I don't remember where we've validated that Lucene104ScalarQuantizedVectorScorer is overall better and we don't need some option to be able to roll back but it's also not clear that that has anything to do with this PR just an errant thought I figured I'd write down.
|
We haven't yet. We'll be able to change the format after this has merged, so we can validate the performance of the new formats before we GA it |
4 participants
Create classes for quantized int vector formats that support bfloat16 and direct IO. This uses the new Lucene104 quantizer.