Inference Optimisation Skill
Curated by Skill Harbor: Hill-climbs your local LLM server speed, decode, prefill and startup, with noise-controlled A/B tests and a ledger of every run.
- What
- Curated by Skill Harbor: Hill-climbs your local LLM server speed, decode, prefill and startup, with noise-controlled A/B tests and a ledger of every run.
- Cost
- Free
- Needs
- a running local LLM server (vLLM, SGLang, llama.cpp, ExLlama/TabbyAPI, MLX or TensorRT-LLM); Python for the harness scripts; access to the machine's GPU tooling (for example nvidia-smi or docker) for the memory watchdog.
- Install
- Copy the installer prompt below into your Muse — your agent does the rest.
Curated by Skill Harbor: Hill-climbs the serving speed of a local LLM server on any hardware and any quantization, from decode and prefill to KV cache, startup time and time to first token. It computes the theoretical ceiling first, then runs noise-controlled A/B experiments, one change at a time, and keeps a ledger of every run. It works with vLLM, SGLang, llama.cpp, ExLlama and TabbyAPI, MLX and TensorRT-LLM, and climbs through the cheapest levers first: persistent caches, decode graphs, speculative decoding, concurrency and KV sizing, then prefill tuning and kernel sizes. Its ground rules are strict: one change per experiment, at least six decode samples per run, controls on both sides, and a win counts only if it beats both controls and regresses nothing by more than about three percent. It also enforces one GPU job at a time, a fixed power profile and a memory cap watchdog, and it forbids printing or copying any credential. Run-to-run noise is typically five to ten percent, so small gains may be noise, and results depend on your hardware, model and quantization. Some trades are labelled lossy, such as int8 prefill, and a server must never be force-killed mid-batch. By @sybil-solutions, listed here with credit to its creator. Discovered via GitHub. Skill Harbor never reviews the code, review it yourself before use.
Version:
Quick answer
Curated by Skill Harbor: Hill-climbs the serving speed of a local LLM server on any hardware and any quantization, from decode and prefill to KV cache, startup time and time to first token. It computes the theoretical ceiling first, then runs noise-controlled A/B experiments, one change at a time, and keeps a ledger of every run.
Install
Copy the install package below, then paste it into Muse📌 Pinned version: commit fda968ePrerequisites: a running local LLM server (vLLM, SGLang, llama.cpp, ExLlama/TabbyAPI, MLX or TensorRT-LLM); Python for the harness scripts; access to the machine's GPU tooling (for example nvidia-smi or docker) for the memory watchdog. Install "Inference Optimisation Skill" for me. It hill-climbs the serving speed of a local LLM server on any hardware and any quantization, from decode and prefill to KV cache, startup time and time to first token, with noise-controlled A/B experiments and a ledger of every run. Repository: https://github.com/sybil-solutions/inference-optimisation-skill Pinned version: install the code at commit fda968e7fc1996aa23c67139c755fe541c9d4be0 exactly (Skill Harbor tested this version). Do not substitute a newer version. If this commit is unreachable, stop and tell me. 1. Fetch the SKILL.md file (and any helper files) from the repository path skills/inference-optimisation/ into a temporary folder. 2. Safety check: review the SKILL.md and scripts for anything suspicious (unexpected network calls, shell commands, credential harvesting). This repo should contain zero secrets in code, credentials only via the secure vault, allowed hosts declared in the SKILL.md. Verify that holds here; STOP on any red flag and tell me. 3. Install it as a skill: copy SKILL.md and its helper files into the agent's skills directory, in a folder named "inference-optimisation". 4. Verify with no network calls: frontmatter valid, files in place. 5. Report what was installed, where, and what I still need to do myself (e.g. point it at my running local server, set my memory cap and target metrics). GitHub is optional: if I have a GitHub account or the gh CLI, you may use it; otherwise public access is fine. Never require it unless it's in the prerequisites above. Rules: don't touch anything outside the temp folder and the install target. If anything looks off, stop and ask me.
Saved to your recent installs. Find it anytime on /connect.
🔒 Security, provenance and what this touches
The install prompt below already includes the vetting steps: your agent follows the community checklist before installing anything with executable code. Want more?
Verification certificate
Dated facts about what was checked for this listing, and how. This is not a safety certification.
- Pinned version
- fda968e
- Sandbox install test
- No sandbox install test on record for this listing.
- Creator identity
- Not verified
- Fingerprint watch
- Last checked · Nov 19, 2026 : no change detected in the files linked by this install prompt
Skill Harbor never reviews the code of a listed skill. These are dated facts about the install prompt and our install test, not a safety certification. Review a skill yourself before you install it. How we verify
Questions
How do I install a skill?
Every product page includes a copy-paste install prompt. Paste it into your Muse and it sets the skill up for you — no manual configuration.
Where does my money go?
Straight to the seller. Skill Harbor never processes payments: checkout happens on the seller’s own page, usually Stripe.
What does the ✓ next to a creator’s name mean?
It means we confirmed the identity of the person behind the listing. It says nothing about the code itself — always check a skill before installing it.