Documentation
Bundle OKF 0.2 · 17 conceitos · salient-data/b200-qwen38-27b-fp8
Open source Repository Open in the app JSON
About
# Documentation
Serving **Qwen3.8-27B** on an on-demand RunPod **B200**, joined to your tailnet
as `b200`, with the engine port never published. This bundle is the source of
truth for what is measured, what is assumed, and what breaks;
[`README.md`](../README.md) is the landing page that links here.
**Shipped configuration:** `Qwen/Qwen3.8-27B-FP8` with in-checkpoint MTP
speculative decoding at k=3, on `lmsysorg/sglang:latest`, `trtllm_mha`,
`mem-fraction-static 0.90`.
## Start here
- [Serving Qwen3.8-27B on a rented B200](reference/serving-qwen3-8-27b-on-a-rented-b200.md)
— what this serves, the 48/16 layer split that explains the host's behaviour,
and how to bring it up.
## Decisions
- [Serve Qwen FP8 with MTP k=3](adr/serve-qwen-fp8-with-mtp-k-3.md) — the
default, and which of its four premises is a measurement and which is a prior.
- [Reach the pod only over the tailnet](adr/reach-the-pod-only-over-the-tailnet.md)
— no published port, no auth proxy, and the tailnet-name failu
Details
- Kind
- OKF bundles
- Topic
- No topic detected
- Publisher
- salient-data
- Origin
- okf_github
- Category
- dados
- Version
- 0.2
- Added
- 2026-09-09 12:02:17
- Updated
- 2026-09-09 12:02:17
- Origin id
salient-data/b200-qwen38-27b-fp8:docs/index.md