- The core challenge is balancing throughput and stability when using low-precision quantization like NVFP4 in reinforcement learning (RL) for large language models, as sampling and training instabilities compound.
- A baseline recipe uses NVFP4 for MoE layers (97% of parameters) with per-token activation scaling to avoid issues like token-dependent quantization and future-to-past leakage.
- Gradient stability is improved via dequantized backward passes, aligning backward operands with forward quantization decisions to reduce mismatches and gradient norm spikes.
- The four-over-six (4/6) technique reduces quantization error by adaptively choosing between maximal values of ±4 or ±6 for both weights and activations, crucial for RL due to disproportionate impacts on pre-trained weights.