llama: properly handle KV on training (#28520)
* llama: properly handle KV on training * improve
X
Xuan-Son Nguyen committed
81ff93ea1d48c0482508d30e75b814df99c25bf0
Parent: 60e9cf7
Committed by GitHub <noreply@github.com>
on 9/30/2026, 4:09:08 PM