pullrequests Search Results · bucket language:JavaScript language:JavaScript language:JavaScript language:Java language:Python
Filter by
1M results
The decode context bucket is chosen from the computed-token counts of every slot in the padded decode batch. Slots of
finished requests keep their last counts, so after a long request finishes, later ...
Description
The scripts behind the bucket_sort pi05 experiments (Studio pi05 29%, physicalai fp32 90%). Draft for sharing, not meant
to merge as is.
| Trained policy (overview and wrist camera ...
Library
Summary
This PR is extracted from the original larger PR #9370 and contains the end-to-end implementation required to make
partition-level bucket counts safe and usable in Flink.
This PR delivers ...
Paired: GitHub #1313 ↔ GitCode !1853
Plan compatible FSDP collectives within explicit model module scopes and split buckets at dependency, effect, storage
and communication boundaries. Fuse reduce ...