FP8 Quantization

Cosmos-RL supports FP8 quantization for policy and rollout. Currently, only rowwise quantization is supported.

Enable FP8 Quantization

To enable FP8 quantization, you need to set the following configuration:

[train.fp8]
enable_fp8 = true
fp8_recipe = "dynamic_scaling"
quant_recipe = "rowwise"

[rollout]
quantization = "fp8"

For policy:

• fp8_recipe could only be set to dynamic_scaling now.

• quant_recipe could be set to rowwise or tensorwise. rowwise is recommended for better accuracy.

For rollout:

• quantization should be set to fp8. Then the rollout will dynamically quantize weights from policy to fp8 during weight synchronization in rowwise manner.