Mixed int8+fp16 weights fix MOSS image model NaN issue, 28% faster on RTX 2060 Super
mkrGG · reddit · 2026-11-18
After digging through ComfyUI GitHub PRs, the author found MOSS image base model's fp16 black-screen/NaN issue stems from peaks in certain weights; KBlueLeaf published a fixed fp16 repo on Hugging Face, but full fp16 exceeds 8GB VRAM.
The fix: mix ComfyUI's int8 weights with KBlueLeaf's fp16 weights, plus an AI-written ComfyUI custom node for fp16 loading with a clamp — no ComfyUI code edits needed. A "heal" weight version is also released for LoRA training.
RTX 2060 Super at 2MP: int8convrot 6.64s/it → mixed int8+fp16 4.78s/it (28% faster); base int8 with the node hits 4.80s/it, though high-MP LoRA use may show clamp artifacts. Weights and node are open-sourced on HF and GitHub.
More from Infra
- Inside Microsoft's massive Outlook outage: what broke and why fixes took so long — tomwarren · 2026-11-18
- vLLM ships DSpark draft model for Kimi K3.1 Pro with up to 30% higher acceptance length — vllm_project · 2026-11-18
- Wonderful runs in-house voice models on Together AI at 274ms first sentence — togethercompute · 2026-11-18
- Anthropic's S-1 Shows Only 25% of Compute Goes to Inference — Sparking a Bear Case — toptickcrypto · 2026-11-18
- LanceDB Raises $57.5M Series B for AI-Native Multimodal Data Stack — pzakin · 2026-11-18
- Multi-Branch Transformer scales LLMs via parallel branches, saving up to 1.26x training compute — KyeGomezB · 2026-11-18