Commit Graph

2 Commits

Author SHA1 Message Date
ZacharyZcR 6d3a6168e4 cuda: keep ragged KV resident across decode steps 2026-07-18 18:34:46 +08:00
ZacharyZcR 1223f9ca2e cuda: batch ragged attention across independent streams 2026-07-18 02:01:12 +08:00