Redline

Redline is a directed-workflow agent skill that finds the safe context ceiling for a vLLM-served model on a fixed GPU. It climbs --max-model-len and --max-num-batched-tokens one variable at a time, clean-rebooting between rungs and recording the real KV-cache and concurrency signal from vLLM boot logs.

The workflow is the program. The agent is the runtime.

Quick Start

Clone this repo alongside the model-tuning workspace or make it available to your agent:

git clone https://github.com/TGPSKI/redline.git

Invoke the workflow by pointing your agent at the skill:

@redline/SKILL.md

Which model and GPU profile are we climbing?
The ladder file should be ./ladder.yml unless you find an existing one.

Redline expects a known-good low-context boot command. It does not discover the starting floor; it climbs from one.

How It Works

The workflow persists state in a ladder file: ladder.yml following assets/rung-result.schema.yml.

Each rung records:

• the model identity and exact vllm serve parameters
• the GPU and allocator environment
• whether boot succeeded, OOMed, or failed for another reason
• vLLM's reported KV-cache capacity and maximum concurrency
• smoke-test status for near-ceiling prefill behavior

The router in SKILL.md reads the last rung and dispatches to the next phase:

Phase	Purpose
references/01-run-rung.md	Compute and run the next single-axis tuning attempt
references/02-record-ceiling.md	Classify a failed rung and write the ceiling summary
references/03-smoke-test.md	Verify a successful boot under near-ceiling request load

Repository Layout

redline/
├── SKILL.md                         # Top-level skill router
├── references/                      # Top-level phase files
│   ├── 01-run-rung.md
│   ├── 02-record-ceiling.md
│   └── 03-smoke-test.md
├── assets/                          # Top-level schema and example assets
│   ├── rung-result.schema.yml
│   └── example-ladder.yml
├── AGENTS.md                        # Contributor and agent guidance
├── README.md                        # You are here
└── LICENSE

Design Rules

• Do not trust model-card context claims as final. Treat them as targets to test toward.
• Raise max_model_len and max_num_batched_tokens independently.
• Classify OOM site before calling a failure a real ceiling.
• Keep ladder.yml append-only for boot outcomes.
• Treat Maximum concurrency as the useful serving signal, not just raw context length.

License

GNU General Public License v3.0

Author

Tyler Pate (@TGPSKI), 2026