feat: DDTree draft + hybrid tree-attention verify for DFlash
Port Lucebox DDTree (top-K extract, best-first tree build, follow) into common/, wire draft-dflash for chain and tree drafts, and add GDN/SSM TREE_MODE CUDA kernels plus ancestor KQ mask. Server path packs a DFS batch with parent_ids/visibility, walks target logits, then checkpoint- restores and chain-commits the accepted path. Enable with LLAMA_DFLASH_DDTREE=1 and LLAMA_DFLASH_TREE_VERIFY=1. Smoke shows higher mean accept length than chain verify, with higher verify cost until intermediate-state commit is available.
A
andi committed
220335d8fff5c64175a006804d319ed53bff1db3
Parent: abe60e4