v0.1.0

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

@zTonyZhao, @MayDomine