Release v0.5.4

What's Changed

• Fixed typos: advatanges -> advantages by @songxxzp in #570
• Only decode the queries once for multiple remote rm by @zhuzilin in #572
• overlap vllm init and actor/reward model loading by @zhuzilin in #575
• correct the order of multiplication in grad acc by @zhuzilin in #577
• Support ring-attention during sft phase by @UbeCc in #576
• Add better error message for empty datasets by @frrad in #581
• Fix nan for sft-ring when labels are all IGNORE_INDEX by @UbeCc in #583
• explicitly ignore attention_mask for packing_samples. by @xiaoxigua999 in #588
• Set default grad_accum_dtype to None @xiaoxigua999
• update global batch size in eval model compatible to ring-attn-size by @ShomyLiu in #590

New Contributors

• @songxxzp made their first contribution in #570
• @UbeCc made their first contribution in #576
• @frrad made their first contribution in #581
• @ShomyLiu made their first contribution in #590

Full Changelog: v0.5.3...v0.5.4