@adelie/gpu-model-serving 0.1.0 → 0.6.3

package/package.json

@@ -1,9 +1,16 @@
 {
   "name": "@adelie/gpu-model-serving",
-  "version": "0.1.0",
+  "version": "0.6.3",
   "description": "Stand up an OpenAI-compatible LLM endpoint on a rented or owned GPU box: pick the instance, fetch and quantize weights, serve with SGLang — speculative decoding, prefix caching and tool calling for agent workloads — then verify it and expose the port.",
   "license": "Apache-2.0",
+  "repository": {
+    "type": "git",
+    "url": "git+https://github.com/lmliheng/Adelie.git",
+    "directory": "plugins/gpu-model-serving"
+  },
   "keywords": [
+    "adelie-plugin",
+    "penguin-harness",
     "gpu",
     "llm",
     "sglang",
@@ -13,20 +20,14 @@
     "quantization",
     "speculative-decoding"
   ],
-  "repository": {
-    "type": "git",
-    "url": "git+https://github.com/lmliheng/Adelie.git",
-    "directory": "plugins/gpu-model-serving"
-  },
   "files": [
     "plugin.json",
     "icon.svg",
+    "README.md",
     "skills",
-    "hooks",
-    "LICENSE",
-    "README.md"
+    "LICENSE"
   ],
   "publishConfig": {
     "access": "public"
   }
-}
+}

package/plugin.json

@@ -4,5 +4,6 @@
   "short_description": "Self-host an LLM endpoint on a GPU box with SGLang.",
   "short_description_zh": "在 GPU 机器上用 SGLang 自部署模型接口。",
   "version": "2026.10.30.1",
-  "category": "software-development"
+  "category": "software-development",
+  "preinstall": false
 }

package/skills/gpu-model-serving/SKILL.md

@@ -9,6 +9,14 @@ You are standing up a **self-hosted, OpenAI-compatible LLM endpoint** on a GPU m
 
 Two things make this harder than it looks, and both are scheduling, not modelling: the **build toolchain** (CUDA headers, nvcc, JIT kernel compilation) and the **memory budget** of a single card. Read [`reference/pitfalls.md`](reference/pitfalls.md) before you start typing; every entry there is a real failure that costs 10–60 minutes when it first appears.
 
+## Before you start
+
+If the message only names this skill with no job behind it, ask what the endpoint is for before touching a machine: the agent framework that will call it, and roughly how much traffic and how much context it needs. Those two answers decide the model class, the quantization and the card.
+
+**Renting costs money by the hour, and prepaid balances stop the instance without warning.** Confirm which machine to use — a box the user owns, or a rented instance and which one — and say what an hour costs before starting a download that takes thirty minutes. Do not rent a larger card than the job needs; a 24–48GB single card covers most agent workloads, and the arithmetic in [`reference/sglang-serve.md`](reference/sglang-serve.md) is what says whether it is enough.
+
+Nothing here is destructive to existing data, but a `sglang serve` occupies the whole card and its port. Before taking both, read back which service, if any, is already running there and get agreement to replace it.
+
 ## The shape of the job
 
 1. **Pick the machine and where the bytes live.** A single 24–48GB card runs a 27B-class model at 4-bit with room for a draft model and KV cache. Put weights, virtualenv and scripts on a **persistent data disk**, never on the small system disk — a stopped/restarted instance keeps the data disk and loses nothing else, and re-downloading 20–30GB is the slowest thing in the whole job.