What optimizer was used to pretrain Chronos-2?

[SulemanQB]

Hi, thanks for the Chronos-2 release and technical report , really nice work.

I wanted to confirm one detail that I couldn't find explicitly stated: which optimizer was Chronos-2 pretrained with?

The Chronos-1 paper explicitly states AdamW (Section 5.2), and the legacy training script in this repo (scripts/training/train.py) defaults its optim argument to adamw_torch_fused. However, the Chronos-2 technical report
(arXiv:2510.15821) describes the training procedure (two-stage context length, quantile regression objective, robust scaling) in detail but doesn't state the optimizer, and I couldn't find a public Chronos-2-specific pretraining script to check directly.

Was Chronos-2 also pretrained with AdamW (or a variant like adamw_torch_fused), or was a different optimizer used? Just trying to make sure I attribute this correctly before building on it.

Thanks!