The development team described the update as the largest optimization in the history of the library. TRL is a tool used by AI researchers and developers to fine-tune and align large language models through reinforcement learning techniques.
Wednesday, October 21, 2026
Story
Hugging Face TRL v1.15 Cuts Peak VRAM by 82%
AI12 days ago
The Transformer Reinforcement Learning library now supports sequence lengths up to 7 times greater than previous versions. This capability is introduced in the release of TRL v1.15, which features a fused LM head enabled by default that reduces peak VRAM usage by up to 82%. These changes allow token capacity for Direct Preference Optimization (DPO) to increase from 10k to 59k tokens and Group Relative Policy Optimization (GRPO) to rise from 29k to 115k tokens.
Looking for more context or analysis?
Tilt chat provides reliable, tailored insights, alerts, and automations.