Dedicate a GPU to Activation Weights for Higher Res MiniMax H3 Video
yamosin · reddit · 2026-08-21
A user shared a multi-GPU optimization technique for MiniMax H3. By storing UNet weights on cuda:0 and dedicating cuda:1 purely for activations during inference, users gain a full GPU's worth of VRAM for activations with practically zero performance hit.
Test Setup: 3x RTX 3090. Workflow loads CLIP+VAE on cuda:0 and ejects CLIP after conditioning. Half the UNet acts as a storage pool on cuda:0, while the other half serves as the compute device on cuda:1.
Results: Shifting UNet weight distribution (e.g., 3GB vs 16.5GB) resulted in negligible sampling time differences (83s-87s). However, this setup allows significantly higher resolutions and longer video generation within VRAM limits.
Conclusion: Two 16GB cards could potentially handle 0.7MP + 15s videos entirely in VRAM (activations 12GB). This offers massive performance gains compared to system RAM swapping.
More from Infra
- Muse Spark Open runs locally on a Mac M3 Ultra via llama.cpp: frontier model on desktop — pcuenq · 2026-10-16
- Muse-Spark open-weights drop: 816B-A42B MoE model gets 359GB MLX 3-bit build — ns123abc · 2026-10-16
- vLLM ships Day-0 LoRA RL for Muse Spark Open, lifting AIME from 0.56 to 0.85 — vllm_project · 2026-10-16
- DSpark draft model for Muse Spark Open proposes 7 tokens per pass with 23-42% acceptance — vllm_project · 2026-10-16
- Cloudflare Stream's WebRTC ultra-low-latency live streaming is now Generally Available — ritakozlov · 2026-10-16
- YC-backed Overshoot launches inference optimization, claims $5M saved for AI teams — ycombinator · 2026-10-16