We are thrilled to release v0.1.0, the first release of Meshy, a ray-less asynchronous RL Engine for LLMs.
Highlights
- Ray-less, role-driven architecture
Meshy models inference, training, and rollout as independent services coordinated through a shared data plane. - Native asynchronous RL support
The same service architecture supports on-policy, bounded off-policy, and fully asynchronous training workflows. - Declarative and flexible topology
Define your training topology in a Python recipe. Each process derives its own placement locally under an SPMD paradigm, without requiring centralized service discovery. - Unified data and control plane
TransferQueue transports tensors, synchronization signals, and GPU ownership tokens, eliminating the need for direct RPC coordination between services. - Flexible GPU colocation
Services can be colocated or disaggregated across GPUs, with arbitrary service groups sharing resources through TransferQueue-managed ownership. - Long-context RL training
Meshy supports context parallelism, dynamic batching, custom advantage shaping, and up to 128K-token contexts. - Reproducible training recipes
This release includes recipes for GRPO on GSM8K and MATH, JustRL reproduction, and fully asynchronous RL training. - Lightweight and debuggable runtime
Per-service logs, complete tracebacks, and queue-level visibility make distributed training runs easier to inspect and diagnose
Recipes
- GRPO + GSM8K for minimal baseline and smoke test.
- JustRL, including async and fully async variants, for reproducing JustRL and demonstrating Meshy's native asynchronous capabilities.
- GRPO + MATH for long context CP-enabled training on MiniCPM5-2B model.
Docker Image
docker pull ztonyzhao/meshy:0.1.0
Contributors
Contributors
zTonyZhao and MayDomine