Musubi Tuner Fork Enables MiniMax H3 LoRA Training on 24GB VRAM
diogodiogogod · reddit · 2026-08-04
A developer shares their fork of Musubi Tuner that enables MiniMax H3 LoRA training on 24GB VRAM (e.g., RTX 4090). By using a pruned ConvRot INT8 model (21GB), BF16 LoRA training is possible with 19-20GB VRAM usage. Tested with 1024x1024 images, batch size 1, LoRA rank/alpha 16, 15 transformer blocks offloaded to CPU, completing two epochs. Advanced features are planned.
More from coding & agent
- Frontier closed models overengine small coding tasks, sparking an engineering crisis — robleclerc · 2026-08-04
- Open Source 'Meta-Skill' Automates High-Quality AI Skill Generation — vista8 · 2026-08-04
- OneStageROS: A Lightweight Web IDE for Complete ROS 2 Development — 4310sy · 2026-08-04
- Context Engine: An MCP Server That Lets Coding Agents Navigate Code as a Graph — artwelf · 2026-08-04
- Agent Burns Monthly Budget in 9 Hours, Dev Interrogates It — lvwerra · 2026-08-04
- 3 Weeks on MCP Registry: Zero Tool Calls, Only Crawlers — suzuridev · 2026-08-04