inference-optimisation-skillsybil-solutions/inference-optimisation-skill
[Core Features]: A skill for hill-climbing local LLM server speed on any hardware and quantization,…
InstallOne-line install
UnconfirmedWhether this repository follows the Agent Skills spec (that is, whether it contains a SKILL.md) has not been verified yet. Until that is settled, clone it and use what you need:
git clone https://github.com/sybil-solutions/inference-optimisation-skill.git
OverviewOverview
[Core Features]: A skill for hill-climbing local LLM server speed on any hardware and quantization, covering single-user and concurrent decode, speculative decoding, prefill, KV cache size, startup time and time to first token under a memory cap with full-context stability checks. The workflow computes a theoretical ceiling, changes one thing at a time, measures against interleaved controls, logs every result including failures and pins a clean build before publishing, with measured speedups on Intel Arc iGPU among the examples. [Applicable Scenarios]: Inference tuning, KV cache sizing, speculative decoding and benchmark measurement. [Target Roles]: Inference engineer, performance engineer and AI infrastructure team.
TopicsTopics
RatingUser ratings
Nobody has rated this skill yet. Yours would be the first on the site — your vote becomes the initial average.
RepositoryRepository
| Skill ID | 5994 |
|---|---|
| Data source | GitHub |
| Repository | sybil-solutions/inference-optimisation-skill |
| Category | Data & Infrastructure Skills / Local LLM Inference Tuning Skill |
| Primary language | Not specified |
| License | Not specified |
| Default branch | |
| GitHub repository ID | |
| Rank | #760 (position within the source ranking) |
| Site-wide popularity rank | #871 (computed live from Stars, across all 5897 entries) |
GrowthStar / Fork growth
TimelineTimeline
CommitLatest commit
New Feature: inference-optimisation skill: hardware- and quant-agnostic LLM serving hill-climb SKILL.md procedure (ceilings first, interleaved controls, full-context gates, publishing), references (ceilings, methodology, levers with measured evidence, gotchas, publishing), engine-agnostic harness (bench.py, cei