← Back to live feed1 story
Hugging Face released TRL v1.15, which reduces peak VRAM usage by up to 82% via a fused LM head enabled by default. The update supports sequence lengths up to 7 times greater than previous versions, increasing token capacity for DPO from 10k to 59k and GRPO from 29k to 115k.
Sign in to suggest edits
Key sources
- SOURCE@qgallouedec“DPO 10k → 59k tokens, GRPO 29k → 115k”x.com
- SOURCEmarketbrief.now