🤖 AI Tools
SGLang Serving Config from a Model Inventory (No Invented Endpoints or Keys)
Write an SGLang serving config from a model inventory. No invented endpoints, API keys, or extra engines.
0Reviews
Prompt
Act as an SGLang serving-config editor who only uses a pasted model inventory. You write launch flags the inventory already names. You do not invent API keys, listen endpoints, or extra engines. This is not a vLLM serve dump, not a TGI config, and not an Ollama Modelfile. You work only from Inputs. Do not invent stats, citations, quotes, URLs, names, IDs, or records that are not in Inputs. Inputs: - Model inventory I lock (names, sizes, flags I allow): [Models] - SGLang version I lock (or UNKNOWN): [Version] - Listen address I lock (or UNKNOWN): [Listen] - Words I must not use: [Banned] - What I must never invent (API keys, endpoints, extra engines, GPU counts): [Never] - Output format: [Format] - Language for comments: [Lang] - Secrets style I lock (env names only): [Secrets] Generate: 1. Honesty ledger: Models, Version, Listen, Secrets, Lang. Forbidden: invented keys, extra engines, extra endpoints. 2. Flag lock table: each flag in Models quoted. Flags not listed stay NOT IN INPUTS. 3. Config: python -m sglang.launch_server from Models. tp-size and context-length only if Models named them. Listen from Listen or LISTEN UNKNOWN. 4. Secrets: env names from Secrets only. Never paste sk- values. 5. Refuse list: vLLM serve, TGI config.yaml, TensorRT-LLM engine, Ollama FROM, raw API keys, invented http://0.0.0.0:30000 if Listen is UNKNOWN. 6. Version: quote Version or write VERSION UNKNOWN. 7. Never: do not add --enable-lora if Models omitted it. Do not invent a Hugging Face token. 8. Compliance pass: quote Banned and Never hits. Cut them. Format as Format. Constraints: - SGLang serving config from Models only. Not vLLM and not Ollama. - Never invent API keys, extra engines, or endpoints. - Secrets stay env names. - No emojis.
Instructions
Replace every [bracket] with your details before running. Works on ChatGPT, Claude, and Gemini.
Examples
Example Input
Models: model path as pasted llama-family-local; dtype bfloat16 as pasted. tp-size NOT IN INPUTS. context-length NOT IN INPUTS. LoRA NOT IN INPUTS. Version: UNKNOWN (do not invent 0.4.1) Listen: UNKNOWN (do not invent 0.0.0.0:30000) Banned: sk-, hf_, always, guaranteed Never: invent API keys, invent endpoints, invent extra engines, invent GPU counts Format: ledger + flag-lock + config + refuse + gaps Lang: English comments Secrets: SGLANG_API_KEY env name only as pasted. No raw key.
Example Output
1. Ledger. Models: llama-family-local; dtype bfloat16. tp-size NOT IN INPUTS. context-length NOT IN INPUTS. LoRA NOT IN INPUTS. Version UNKNOWN. Listen UNKNOWN. Secrets SGLANG_API_KEY env. Lang English. Forbidden: sk-, hf_, always, guaranteed, extra engines, extra endpoints, GPU counts. 2. Flag lock. --model-path llama-family-local KEEP. --dtype bfloat16 KEEP. --tp-size NOT IN INPUTS. --context-length NOT IN INPUTS. 3. Config sketch. # VERSION UNKNOWN. LISTEN UNKNOWN. Do not invent 0.0.0.0:30000. python -m sglang.launch_server --model-path llama-family-local --dtype bfloat16 # --tp-size omitted. NOT IN INPUTS. # --context-length omitted. NOT IN INPUTS. # --enable-lora omitted. NOT IN INPUTS. # --host/--port omitted. LISTEN UNKNOWN. # SGLANG_API_KEY env name only. 4. Secrets. SGLANG_API_KEY env. No sk- value. No hf_ token. 5. Refuse. vLLM serve: refused. TGI config.yaml: refused. TensorRT-LLM engine: refused. Ollama FROM: refused. 0.0.0.0:30000: refused. sk- key: refused. GPU count 8: refused. 6. Version. VERSION UNKNOWN. 7. Never. No LoRA flag. No minted listen. No Hugging Face token. 8. Compliance. Banned hits none. Format ledger+flag-lock+config+refuse+gaps. Gaps: SGLang version, listen address, tp-size, context-length, LoRA adapters. Missing-data policy: if a field was blank, write NOT IN INPUTS rather than guessing. Lock any tool version named in Inputs; if unnamed, write unknown. No invented testimonials, star ratings, or press logos. If legal, clinical, insurance, HR, education-plan, or veterinary content appears, add a one-line not-advice and de-identify banner. Quote banned-word hits and cut them. End with a gaps list of five bullets the user still owes you. Character and byte caps in the job are hard; print counts when relevant. Refuse to backfill DOIs, exam dumps, PHI, PII, or compensation promises not in Inputs.