Skip to content

Commit 52740a9

Browse files
AnbeeldAnbeeld
committed
Optimize KVarN materialize hot paths
Add the fast-vs-generic materialize oracle and extend KVarN backend perf coverage to bits 3 and 5. Recover CUDA materialize throughput with counter-backed K/V fast-path changes while preserving the generic fallback.
1 parent 13d15a3 commit 52740a9

3 files changed

Lines changed: 540 additions & 131 deletions

File tree

0 commit comments

Comments
 (0)