Fixing MiniMax H3 Black Frames on Legacy GPUs: FP16 Mix Cuts Inference 11x
Bubbly_Lawfulness_43 · reddit · 2026-08-08
MiniMax H3 officially only supports bf16/fp32, forcing older GPUs like the V100 to run at a crawl in fp32, while forcing fp16 outputs pure black frames.
A developer investigated and found that three specific places in the model's DiT architecture overflow fp16's max value (text-conditioning projection, attention-sink rows, etc.). By creating a custom node that keeps only these three spots in fp32 while running the rest in fp16, they achieved flawless inference on the V100 without quality loss. Speed dropped from 350s/step to 30s/step, an 11x speedup.
The fix is open-source, plugs directly into ComfyUI, and safely self-disables on modern bf16-supported hardware.
More from Infra
- Zero's next-gen IVM engine runs in a single DO, live-querying a 100GiB R2 database — threepointone · 2026-10-16
- Meme: driving to the chip factory in 2028 captures AI's compute anxiety — beffjezos · 2026-10-16
- One volunteer maintains the time zone file behind 4 billion phones — danbri · 2026-10-16
- DeepSeek V4.1 Flash on 4x DGX Spark: weights alone are 511GB, the 494 tok/s claim doesn't fit — Aroochacha · 2026-10-16
- Open-sourced Intel GPU Vulkan inference upgrades: 2x ingestion, 60%+ faster inference — Fun-Aardvark-1143 · 2026-10-16
- ELSFP pain points: >3dB insertion loss, costly PMF, harsh packaging requirements — jwt0625 · 2026-10-16