Research & academia
Toward RL on mixed-precision, mixed-accelerator samplers
Can RL training stay stable when its samplers use different accelerators, precisions and kernels from the trainer? The authors report that off-policy correction can make this workable.
Why it matters: The results suggest sampler and trainer hardware need not be matched for stability: correction can let each use its own accelerators and precision. The reported cost saving is per update for sampling, not total training cost.