fix: fp16 config and grad scaling

[yxyOo]

Summary

Ensure FP16 training is consistent and safe by setting required optimizer params, aligning fp16/bf16 flags, and avoiding duplicate grad scaling.

Changes

• Set required FP16 optimizer parameters in code to avoid missing user‑configurable CLI options.
• Set store_param_remainders=false to reduce memory usage, since this option isn’t exposed by Megatron’s public config.
• Skip duplicate grad scaling checks during FP16 training.
• Refs: [megatron] feat: fp16 training (dense and MoE supported) verl-project/verl#4086

Results

fp16 shows faster convergence speed and lower mismatch between training and rollout.

[zhuzilin]

I think that these params can be set with the

    for f in dataclasses.fields(OptimizerConfig):
        if hasattr(args, f.name):
            kwargs[f.name] = getattr(args, f.name)

For example, you can set the initial_loss_scale with --initial-loss-scale 32768.

[yxyOo]

Yes, configurable, but these params must be enabled with FP16—otherwise performance degrades and users may miss them.

[yxyOo]

And store_param_remainders cannot be set this way (not exposed in Megatron’s public config), while other params are permissible.