LTX-2.5 22B LoRA Training Working on Dual RTX 3060 12GB
Ok-Beautiful-3479 · reddit · 2026-08-24
The author published a LoRA trainer for LTX-2.5 designed for low-VRAM setups. The core technique is multi-GPU model sharding, distributing the 48 transformer blocks across GPUs instead of requiring one GPU to hold the entire 22B model.
- Setup: 2× RTX 3060 12GB, 4-bit BNB NF4.
- Results: 7-9 GB VRAM per GPU. Successfully trained a face + voice LoRA (138 images + 37 voice/video segments, 2,000 steps) and loaded it back into LTX-2.5.
- Future: Testing 1x2 to 6x6 spatial tiling to trade VRAM for compute.
More from Infra
- Ramp's local model adoption prompts industry rethink on API costs — annetgriffin · 2026-08-24
- Out of GitHub Actions minutes? Time to move to a self-hosted VPS — MicahBerkley · 2026-08-24
- Ubuntu aims to become the reference OS for RISC-V CPUs — bookwormengr · 2026-08-24
- Neocloud offtake contracts exceed $250B through 2026 — AccBalanced · 2026-08-24
- US AI Data Centers Spark Death Threats; GOP Sees Electoral Risk — rohanpaul_ai · 2026-08-24
- OpenAI VP: Elon's edge comes from co-designing hardware down to the materials — hsu_byron · 2026-08-24