MiniMax-H3 Video LoRA Training OOMs on 96GB VRAM: Optimization Tips Sought
isari_chan · reddit · 2026-08-05
A developer encountered severe memory and performance bottlenecks while attempting to train an audio-enabled LoRA for the MiniMax-H3 video model on a single 96GB VRAM RTX PRO 6000 workstation.
- Setup: Using an INT8 quantized DiT model with Batch size 1 and a 5-second video (107 frames), disabling gradient checkpointing immediately triggers CUDA OOM.
- Performance: With gradient checkpointing enabled, training runs but peaks at 48.3GB VRAM and takes 27.6 seconds per step. A 1,500-step run would take 11.5 hours. Testing musubi-tuner with CPU offloading was even slower (43.4s/step).
- Core Confusion: While ComfyUI can run inference on 24GB VRAM via system RAM offloading, training demands significantly more memory.
The author is seeking community advice on better memory management, torch.compile acceleration, and optimal hardware configurations.
More from Infra
- Running a Complex AI Task Costs Only $0.0042 — victormustar · 2026-08-06
- Opinion: AI and Robotics Will Make Labor Abundant, Compute and Energy Are the Next Oil — VraserX · 2026-08-06
- Dassault Systèmes Partners with NVIDIA to Accelerate Simulation via AI — NVIDIAAI · 2026-08-06
- Ezra Klein Podcast Discusses AI Compute Crisis and Data Center Moratoriums — kevinsxu · 2026-08-06
- Running MiniMax H3 on RTX 5090: Node Optimization Triples Speed — WARRIORPSIX · 2026-08-06
- Opinion: AI Data Centers Are the Future, Canada Must Overcome Backlash — LoganGrasby · 2026-08-05