pretty_name: Own Your Inference — benchmark data
license: cc-by-4.0
tags:
- benchmark
- llm-serving
- capacity
- terminal-bench
- sglang
Own your inference — benchmark data
Raw benchmark artifacts behind the "own your inference" talk: what a self-hosted GLM-5.2 endpoint can and cannot do, measured on three deployments.
Companion code repository: Gladiator07/own-your-inference. It holds the runners that produced this data and a script that rebuilds every table from these artifacts on a laptop.
The three deployments
| Experiment | Hardware | Weights | KV cache |
|---|---|---|---|
h200-fp8-kv |
8x NVIDIA H200 | GLM-5.2 FP8 | fp8_e4m3 |
h200-bf16-kv |
8x NVIDIA H200 | GLM-5.2 FP8 | bfloat16 |
blackwell-nvfp4 |
8x NVIDIA RTX PRO 6000 Blackwell | GLM-5.2 NVFP4 | fp8_e4m3 |
Layout
| Path | Contents |
|---|---|
compact/ |
small download: result tables rebuilt from the raw artifacts, quality tables, and reconciled summaries |
hammer/<experiment>/ |
closed-loop capacity sweeps: N sessions sending nonstop |
paced/<experiment>/ |
open-loop capacity sweeps: sessions arrive on a recorded developer rhythm |
quality/<experiment>/ |
Terminal-Bench 2.1 runs through the pinned Terminus-2 agent |
deployment/blackwell-nvfp4/ |
server bringup evidence for the Blackwell deployment |
inputs/ |
the replay corpora and arrival rhythm; SOURCES.md says where they come from |
Every experiment folder has a meta.json with the server identity, inputs, parameters, and headline numbers. Start there.
Capacity sweep logs sit beside their folders (for example hammer/h200-bf16-kv.log). A paced curve that spans two server boots keeps one folder per boot (boot-a/, boot-b/); the boot split is described in that experiment's meta.json.
Integrity
MANIFEST.jsonl lists every artifact file with its size and SHA-256. SHA256SUMS is the same list in plain checksum format.
Talk
Deck: share.atharvaingle.dev/dhs-2026
License
CC-BY-4.0. The replay corpora are derived from the LMCache agentic traces (CC-BY-4.0); inputs/SOURCES.md has the details.