[megatron] 11/n towards Kimi K2.6: file-backed offload of frozen masters so Tinker API LoRA training fits host RAM - #2062

Open

casper-hansen wants to merge 4 commits into

NovaSky-AI:mainfrom

casper-hansen:casper/kimi-11-frozen-mmap-offload

Commits on Aug 19, 2026

[megatron] Offload frozen masters to file-backed mmap; keep LoRA buffers resident

casper-hansen
and
cursoragent
committed
c3a70d3

Commits on Sep 4, 2026

[megatron] Decide offload buffer residency from the is_lora flag

avigyabb
and
claude
committed
aa604a0
unlinking file so no other process can use other model weights
avigyabb
committed
c321e58
got rid of .tobytes() which created 2x copies of the weights on cpu
avigyabb
committed
41c8571