TRL v1.15 Released With Fused LM Head Cutting Peak VRAM by Up to 82%

Key Info

TRL v1.15 has been released as the project's biggest optimization yet, with a fused LM head enabled by default.

Highlights

• Up to 82% less peak VRAM usage
• Sequences up to 7x longer
• DPO capacity grows from 10k to 59k tokens
• GRPO capacity grows from 29k to 115k tokens