Speed up 27B LLM coding with draft-p-min configuration
Old-Sherbert-4495 · reddit · 2026-08-28
Shared a specific configuration to speed up code generation for a 27B model. By setting --spec-draft-n-max (6-16) and --spec-draft-p-min 0.8, the TPS increased from an average of 25-35 to 35-40+.
Parameter Notes:
- Higher n-max consumes more VRAM but offers greater speedup potential.
- Higher p-min makes the filtering stricter.
The author is still experimenting with the best p-min values for other scenarios.
More from coding & agent
- Paseo: Open-source agent orchestrator with full mobile support — jasonkneen · 2026-08-28
- LangChain Tutorial: Building Self-Correcting Memory in OpenWiki — adnan_hashmi · 2026-08-28
- Uber Cuts AI Costs by 34% While Scaling Agent Usage 9.4x — dosco · 2026-08-28
- LLMRouter: Unified Infrastructure for Developing and Deploying LLM Routers — adnan_hashmi · 2026-08-28
- Warp introduces self-improvement loops for Agents — vikvang1 · 2026-08-28
- Warp opens early access to Factories: agents-as-code platform running 200k agent runs/day — vikvang1 · 2026-08-28