QGallouedec· @QGallouedec · X·· 2 天前AI 评分
Hugging Face TRL v1.15 发布,默认启用融合 LM head 优化
TRL v1.15 发布,官方称这是迄今最大的一次优化,融合 LM head 默认开启。峰值显存最多降低 82%,序列长度提升 7 倍,DPO 从 10k 提升到 59k tokens,GRPO 从 29k 提升到 115k,更多细节见 github.com/huggingface/trl/r…。
TRL v1.15 is out. Our biggest optimization ever.
Fused LM head, on by default:
→ up to 82% less peak VRAM
→ 7x longer sequences
→ DPO 10k → 59k tokens, GRPO 29k → 115k
And more... github.com/huggingface/trl/r…
来源:QGallouedec · x.com