MiniMax H3 Slowdown on 12GB Cards? A ComfyUI VRAM Troubleshooting Guide
BrooklynBrawl · reddit · 2026-08-06
This post details the troubleshooting process for a severe generation slowdown (degrading from 18s/it to 400+ s/it) when running the MiniMax H3 model via ComfyUI on a 12GB RTX 4070 Ti.
Root Cause
The author identifies ComfyUI's newer comfy-aimdo DynamicVRAM feature as the culprit. Under VRAM pressure and an existing memory leak, its aggressive allocation tipped memory over into the drastically slower system shared memory overflow.
The Fix
Adding --disable-dynamic-vram and --disable-pinned-memory to the launch flags resolved the slowdown and memory climbing issues. The article also shares a timeline of fixing initial int8 crashes, OOM errors, and configuring SageAttention.
More from Infra
- Maple 20B Hits 9,885 tokens/s on Single GH200 with 64 Concurrent Requests — MaziyarPanahi · 2026-08-06
- Open-Source LLM Inference Engine TokenSpeed Joins PyTorch Ecosystem — zhyncs42 · 2026-08-06
- Stranded Gas Wells to Power AI: Converting Methane into Methanol for Data Centers — Ghost_Pilot_MD · 2026-08-06
- FPTalks 2026 Focuses on Low-Precision LLM Pretraining — nmwsharp · 2026-08-06
- Gemma 4 Runs Offline on iPhone with Just 500MB of RAM, Outsmarting Siri — cyb3rops · 2026-08-06
- Targon Launches Organizations: Team Collaboration with Confidential Compute — JosephJacks_ · 2026-08-06