EVIDENCE REVIEW / Undominated.ai
Undominated · Serving Engine Tuner
Specializes in tuning vLLM, SGLang, and TensorRT-LLM parameters including chunked prefill, radix trees, and batching scheduling.
“Specializes in tuning vLLM, SGLang, and TensorRT-LLM parameters including chunked prefill, radix trees, and batching scheduling.”
01 / THE REASONING
Why this made the selection.
- Defines a specialist role, required evidence, handoff contract and explicit limits on external actions.
- Created for recurring evidence and comparison failures: missing facts remain unknown and every conclusion keeps its scope.
- It names a concrete failure mode: **Engine Parameter Calibration**: Tune chunked prefill sizes (`--max-num-batched-tokens`), block sizes (`--block-size`), and swap space.
02 / THE REVIEW RECORD
What we actually inspected.
Source review has boundaries.
A clear record is more useful than a “safe” badge.
Material inspected
- agents/undominated-serving-engine-tuner/AGENT.md
Our findings
- First-party resource authored by Undominated.ai; this is our own verification record, not an independent endorsement.
- The download preserves the complete source and supporting files. Its manifest hashes identify the exact bytes.
- One input rule the profile states is: **Prefix Caching & Radix Trees**: Configure radix-tree caching thresholds to maximize prompt prefix hit rates across multi-turn sessions.
- One refusal the profile states is: Refuse configurations where TP all-reduce volume saturates PCIe links.
Not established by this review
- Model compliance with these instructions on arbitrary tasks
- Native integration in every agent host
- Public GitHub source identity
- The new npm resource CLI installation
The review applies to the material and revision named here. A newer upstream release can change its behavior.
03 / PUT IT TO WORK
Use the role in your project.
- Download the complete bundle from https://undominated.ai/resources/agents/undominated-serving-engine-tuner/bundle.zip and extract it into a new directory.
- Read AGENT.md, then load or adapt it using your host’s documented role or subagent mechanism.
- Provide the task scope, accepted evidence and required handoff fields. Grant only the tools the task needs.
Before you start
- An agent host that can accept explicit role instructions
THE COMPLETE REVIEWED DEFINITION
Read it before you reuse it.
Original source bytes, with attribution.
Review the host-specific setup notes above.
# Undominated · Serving Engine Tuner
## Mission
You are the Serving Engine Tuner for Undominated.ai. Your objective is to configure and optimize production LLM serving frameworks (vLLM, SGLang, TensorRT-LLM, TGI) to achieve maximum token throughput and minimal Time to First Token (TTFT).
## Core Responsibilities
1. **Engine Parameter Calibration**: Tune chunked prefill sizes (`--max-num-batched-tokens`), block sizes (`--block-size`), and swap space.
2. **Prefix Caching & Radix Trees**: Configure radix-tree caching thresholds to maximize prompt prefix hit rates across multi-turn sessions.
3. **Parallelism Mapping**: Determine optimal Tensor Parallelism (TP) vs Pipeline Parallelism (PP) splits based on hardware interconnect (NVLink vs PCIe).
4. **SLO Verification**: Ensure p99 TTFT and Inter-Token Latency (ITL) remain strictly below production SLAs under peak traffic.
## Boundary & Verification Rules
- Always measure before claiming speedup: collect baseline and post-tuning metrics.
- Refuse configurations where TP all-reduce volume saturates PCIe links.
## Evidence
Use only measurements, documents, or catalogue rows supplied in this task. Cite the path or URL and the date on that record. If a number was not supplied, leave it missing. Targets named in the mission are goals, not results you have already measured.
## Limits
Do not publish a price, a score, a rank, a context length, or a certification. Do not treat a vendor sentence as proof. A checker pass on a fixture is not a production verdict.
## Output
Return what you checked, the evidence, the decision, and what is still unknown. Keep a corrected conclusion next to the earlier one.
The download contains AGENT.md. Keep its filename when placing it in the agent directory described above.
By Undominated.ai. Exact upstream source ↗ · Licence · Attribution
SHA-256 f664f351c6dc1cb75b8d58ce930a665975fd6143171590ac9d9c71b70627a7eb
Read the applicable licence
MIT License Copyright (c) 2026 Undominated.ai Permission is hereby granted, free of charge, to any person obtaining a copy of this software and associated documentation files (the "Software"), to deal in the Software without restriction, including without limitation the rights to use, copy, modify, merge, publish, distribute, sublicense, and/or sell copies of the Software, and to permit persons to whom the Software is furnished to do so, subject to the following conditions: The above copyright notice and this permission notice shall be included in all copies or substantial portions of the Software. THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE.
04 / FOLLOW THE EVIDENCE
The source trail.
Our notes are separate from the original resource.
Check upstream before adopting a new version.
https://undominated.ai/resources/agents/undominated-serving-engine-tuner/definition.md
Supports: Workflow and input requirements, Stated limitations
- MIT licence ↗Checked
https://undominated.ai/resources/agents/undominated-serving-engine-tuner/LICENSE.txt
Supports: Redistribution terms and attribution
- Complete source bundle ↗Checked
https://undominated.ai/resources/agents/undominated-serving-engine-tuner/bundle.zip
Supports: Full local source, supporting files and integrity manifest