llama.cpp fuses the DFlash speculative encoder into KV-cache injection, cutting a device-to-host round trip per round

GitHub·medium signalXBlueskyLinkedInCopy link

Build b10715, released 2026-08-31, folds the DFlash encoder into the decoder's embd branch so a single `llama_decode` handles it (#27310). The encoder is only a fully-connected layer plus a norm, but running it as a separate `llama_encode` forced its output on a device-to-host round trip before the injection decode could re-upload it, and built a second graph per round. The commit is co-authored with Claude Fable, which is itself a signal about how these kernel-level patches are now written.