Fixing MiniMax H3 Local OOMs: 3090 Runs 30s Video with Tweaked Args
knoll_gallagher · reddit · 2026-08-05
While running the MiniMax H3 model locally on an RTX 3090 (64GB), a developer discovered a surprising cause for OOM (Out of Memory) errors: GPU demands being too low, which triggered dynamic memory allocation issues.
By tweaking ComfyUI launch arguments (like --vram-headroom 3), the author successfully generated a 30-second video at 0.2 megapixels. Using int8 and kj sage optimizations, a single video takes about 16 minutes. The author recommends keeping the H3 environment separate from the main ComfyUI instance due to its unique scheduling behavior.
Related event: Running MiniMax H3 on a Single RTX 3090(2 posts)→
More from Infra
- MLCommons launches MLPerf Endpoints, a benchmark that measures inference endpoints the way buyers actually purchase them — TheKanter · 2026-10-16
- Neocloud M&A decoded: CoreWeave, Nebius, Crusoe, Nscale all bet margins vanish beyond GPU scarcity — AccBalanced · 2026-10-16
- AI-Driven Timing-Aware Power Optimization Uses Learned Ranking With Baseline-Preserving Repair — Amit Sharma · 2026-10-16
- Nscale AI datacenter plagued by coolant leaks, transformer fire, 10 safety violations — saranormous · 2026-10-16
- WSL 3.0 ships with native containers and Linux 6.18 kernel, syscall overhead down ~16% — solyarisoftware · 2026-10-16
- Are 22 GB modded RTX 2080 Ti cards a viable budget local LLM option? — Dave86ch · 2026-10-16