Toward RL on mixed-precision, mixed-accelerator samplers
We trained a model on agentic coding tasks using fully asynchronous RL and mixed-precision, mixed-accelerator samplers. With off-policy correction, training stayed stable through convergence, on par with homogeneous samplers, at a lower cost.
Read the full story at Modal Blog ↗
Timeline · 1 report
- 2026-10-13 12:00 · Modal Blog
Toward RL on mixed-precision, mixed-accelerator samplers