inference/research
draftlow · 0.50updated 2026-10-15

VibeSys bespoke serving engine (agent-built)

rocm, hip
This record is a draft. Not yet verified. Provisional. Treat every number on this page as provisional.

Fields

id
engines/vibesys-bespoke-engine
Repo
vibesys-playground/qwen3.5-397b-mi300a-bespoke-vibesys-20260920
Languages
PythonHIP
License
unknown
First release
2026
Design goal
throughput-firstmodel-hardware-workload-specialized
Scheduling
continuous-batching
Backends
rocmhip
Formats
mxfp4
Notable features
agent-generated bespoke engineHIP graph capturefused MXFP4 MoE kernelDeltaNet snapshot prefix cacheMTP speculative decoding
Hardware caveats
written specifically for 4x MI300A gfx942 + Qwen3.5-397B; not a general engine
Best for
single-deployment peak-goodput serving where agents write the engine
Avoid for
general multi-model serving
notes
~19.4k LOC incl 520 HIP written by Claude-Code agents over 105h; reached 2242 tok/s goodput on Qwen3.5-397B MXFP4 vs 961 for tuned SGLang on 4x MI300A.

Notes

~19.4k LOC incl 520 HIP written by Claude-Code agents over 105h; reached 2242 tok/s goodput on Qwen3.5-397B MXFP4 vs 961 for tuned SGLang on 4x MI300A.

Cross-references

This record references no other records — it is a leaf. It is cited by 1.