DFlash for MiniMax-M3 (WIP): synthesis thinking-mode mix - #1749

Draft

yeyu-nvidia wants to merge 19 commits into

mainfrom

yeyu/dflash-minimax-m3

Commits on Jun 16, 2026

specdec: per-conversation thinking-mode mix in data synthesis (for MiniMax-M3)

yeyu-nvidia
and
claude
committed
specdec(server_generate): accept OAI 'messages' input + OAI-native output

yeyu-nvidia
and
claude
committed

Commits on Jun 22, 2026

specdec(dump): enable vLLM hidden-state dump for MiniMax-M3 (OMNIML-4747)

yeyu-nvidia
and
claude
committed
specdec(recipe): add MiniMax-M3 DFlash offline recipe + train template (OMNIML-4747)

yeyu-nvidia
and
claude
committed
specdec(recipe): default MiniMax-M3 DFlash export YaRN to full 1M (factor 128)

yeyu-nvidia
and
claude
committed

Commits on Jun 25, 2026

specdec(recipe): point MiniMax-M3 DFlash dump at the cleaned synth dataset

yeyu-nvidia
and
claude
committed

Commits on Jun 29, 2026

Commits on Jun 30, 2026

Commits on Jul 2, 2026

Commits on Jul 7, 2026

specdec(recipe): M3 offline training on 1 node (2-node C10d rendezvous unstable)

yeyu-nvidia
and
claude
committed