Llama.cpp DSpark PC Tree fork boosts inference by up to 29.5%
RapidRaid · reddit · 2026-08-22
A developer implemented the DSpark PC Tree algorithm in llama.cpp, providing benchmark results on RTX 5090 and CPU.
Performance (Qwen 3.0, RTX 5090):
- Plain: 94.27 tok/s
- DSpark n3: 155.64 tok/s (+65%)
- PCTree k3/n16: 159.00 tok/s (+69% vs plain)
The PCTree k3/n16 configuration beat linear DSpark in 9 of 11 categories, with a peak gain of +6.56% in summarization. CPU tests showed consistent stability.
Launch Args:
--spec-type draft-dspark --spec-draft-n-max 3 --spec-dspark-pctree --spec-dspark-pctree-k 3 --spec-dspark-pctree-n 16
More from Infra
- Muse Spark Open runs locally on a Mac M3 Ultra via llama.cpp: frontier model on desktop — pcuenq · 2026-10-16
- Muse-Spark open-weights drop: 816B-A42B MoE model gets 359GB MLX 3-bit build — ns123abc · 2026-10-16
- vLLM ships Day-0 LoRA RL for Muse Spark Open, lifting AIME from 0.56 to 0.85 — vllm_project · 2026-10-16
- DSpark draft model for Muse Spark Open proposes 7 tokens per pass with 23-42% acceptance — vllm_project · 2026-10-16
- Cloudflare Stream's WebRTC ultra-low-latency live streaming is now Generally Available — ritakozlov · 2026-10-16
- YC-backed Overshoot launches inference optimization, claims $5M saved for AI teams — ycombinator · 2026-10-16