DoRA + rsLoRA 2026 Production LoRA Starting Config: r=16, target_modules='all-linear', rsLoRA Enabled for High-Rank Stability

Meta Intelligence·medium signal

The 2026 production consensus for LoRA fine-tuning is DoRA (Directional Low-Rank Adaptation) with r=16 and target_modules='all-linear' as the default starting configuration, superseding the older r=8 + selected modules approach. Enable rsLoRA (rank-stabilized LoRA) when experimenting with ranks above 32 to prevent training instability. DoRA decomposes weight updates into magnitude and direction components, improving learning dynamics particularly for instruction-following fine-tunes. Use Unsloth on consumer hardware for 2x throughput.