inference/research
verifiedmedium · 0.80updated 2026-10-15

SyFI/VibeSys: 4x MI300A, bespoke agent-built engine, Qwen3.5-397B-A17B MXFP4, 96 sessions, goodput

2242.0 tok/s aggregate goodput (96 concurrent sessions, SLO-filtered)

Fields

id
benchmarks/syfi-mi300a-vibesys-bespoke-qwen35-397b-mxfp4-96sess-goodput
Metric
tps_aggregate
Value
2,242
Unit
tok/s aggregate goodput (96 concurrent sessions, SLO-filtered)
Model
amd/Qwen3.5-397B-A17B-MXFP4
Methodology
THIRD-PARTY MEASURED (SyFI Lab, UW). Bespoke serving engine written by VibeSys coding agents (~19.4k LOC incl 520 HIP) on 4x AMD MI300A. Qwen3.5-397B-A17B MXFP4, MTP speculation (84% draft accept). Multi-turn chat SLO p95 TPOT<=250ms & TTFT<=10s. Peak goodput 2242 tok/s @96 sessions = 2.33x tuned SGLang (961 @48). AGGREGATE across sessions.
Measured by
third_party
Reproducible
yes
notes
Engine = bespoke agent-built engine; engine record vibesys-bespoke-engine created for it. NOT vLLM/SGLang.

Notes

Engine = bespoke agent-built engine; engine record vibesys-bespoke-engine created for it. NOT vLLM/SGLang.

Cross-references