pyMotor¶
pyMotor (MindIE-Motor) is an Ascend-native distributed inference serving
framework that runs vLLM-style prefill/decode (PD) disaggregation on Atlas
hardware. A pyMotor cluster is described as Kubernetes-style configs — a
Coordinator plus Engine Pods driven by user_config.json and deploy.py —
which makes deployment, scaling, and failover declared and reproducible.
pyMotor serves vLLM-compatible engines on Ascend and ships with features such as manual scaling, primary/standby failover, request tracing, PD-role rescheduling, and container snapshot.
UCM plugs into pyMotor as a persistent KV cache store backend
(UCMConnector / UcmPipelineStore): prefill writes the KVCache once, and
later requests that share the same prefix reuse it instead of recomputing.
Deployment¶
Installation and deployment are covered by the official MindIE-Motor documentation:
| Task | Guide |
|---|---|
| Documentation home | MindIE-Motor docs |
| Environment preparation | Official guide |
| Quick start | Official guide |
| Kubernetes deployment — PD disaggregation | Official guide |
| Kubernetes deployment — PD aggregation | Official guide |
| Standalone Coordinator deployment | Official guide |
UCM as the KV cache store backend¶
To run pyMotor with UCM as its persistent KV cache store, follow the dedicated integration guide:
It walks through installing the UCM wheel in the prefill role, mounting the
shared cache storage under motor_deploy_config.storage, wiring the
MultiConnector (Mooncake in front, UCMConnector in position 2) for prefill,
and verifying KVCache hits on repeat requests.