Llama.cpp DSpark PC Tree Fork (up to 3%-29.5% faster!)

Hello gang, I made an implementation of DSpark PC Tree (Parent conditioned drafting tree). This is an implementation of this research paper: arxiv.org/abs/2608.02123 Unaffiliated, just found it and implemented it. And I have to preface: This is just a first shot, I have no feedback from anyone yet! These are some stats im getting with Qwen 3.0: GPU: SM120 (RTX5090) llama-bench combined Configuration tok/s vs plain vs DSpark n3 Acceptance ━━━━━━━━━━━━━━━ ━━━━━━━━ ━━━━━━━━━━ ━━━━━━━━━━━━━━ ━━━━━━━━━━━━ Plain 94.27 1.00x -39.43% n/a ─────────────── ──────── ────────── ────────────── ──────────── DSpark n2 138.83 1.47x -10.80% 55.66% ─────────────── ──────── ────────── ────────────── ──────────── DSpark n3 155.64 1.65x baseline 47.19% ─────────────── ──────── ────────── ────────────── ──────────── PCTree k2/n8 152.29 1.62x -2.15% 58.83% ─────────────── ──────── ────────── ────────────── ──────────── PCTree k3/n16 159.00 1.69x +2.16% 67.87% ─────────────── ──────── ────────── ────────────── ──────────── PCTree k4/n22 157.99 1.68x +1.51% 72.16% k3/n16 beat linear DSpark in 9 of 11 categories. Its largest gain was summarization at +6.56%. Only RAG (-0.57%) and multilingual (-1.35%) slightly favored linear. k4/n22 accepted more but drafts but could not repay the larger verifier batch. Benchmark conditions: SPEED-Bench qualitative, all 11 categories, four samples per category, 44/44 completed with zero failures, 512 output tokens, temperature 0, 8K context, one slot, BF16 target and draft. CPU-only llama-bench, 12 threads, three repetitions: Model pp512 tg128 ━━━━━━━ ━━━━━━━━━━━━━━━━━━━━━━ ━━━━━━━━━━━━━━━━━━━━━ BF16 76.59 +/- 0.33 tok/s 3.82 +/- 0.02 tok/s ─────── ────────────────────── ───────────────────── Q8_0 76.07 +/- 0.79 tok/s 7.12 +/- 0.01 tok/s The PCTree launch parameters are: --spec-type draft-dspark --spec-draft-n-max 3 --spec-dspark-pctree --spec-dspark-pctree-k 3 --spec-dspark-pctree-n 16 Example complete command: llama-server.exe -m Qwen3-8B-BF16.gguf -md dspark-Qwen3-8B-BF16.gguf --spec-type draft-dspark --spec-draft-n-max 3 --spec-dspark-pctree --spec-dspark-pctree-k 3 --spec-dspark-pctree-n 16 -ngl all -ngld all -c 8192 --parallel 1 --temp 0 --jinja --host 127.0.0.1 ` --port 8080 It would be cool if you could test it out with some models and tell me if it works or not. kinda hoped someone on llamacpp would look at it, but seemingly no one wanted to :P I tested Qwen 3.8 27B Q4, but that was sadly worse with k2-4. Im interested to see in how DS Flash would perform for example. You can read a bit more below in the issue I created. Repo: github.com/getraid/llama.cpp-dspark-pctree Related research issue: github.com/ggml-org/llama.cpp/issues/27499