this is pure witchcraft, thank you a lot!
Dude, I managed to get those models, even the HQ ones, to run flawlessly on my modest setup (RTx 5060 TI 16gb vram + 32gb RAM) at a very decent speed, no oom, i'm in owe!
I had to tweak the MiniMax H3 Mem Eff Sage Attention Patch to deal with the "exotic" architecture, i'm right now using MiniMax-H3_Ref2VA-DT-sQKV-INT8-ConvRot-HQ.safetensors, everything is computed on the gpu, the cpu and my SSD are idle,
thank you, thank you, thank you, thank you, thank you, thank you,
Thank you for the feedback — I’m really glad my build runs this well on a 16 GB GPU. In my opinion, DT-sQKV HQ is indeed one of the most successful variants. It restores three physically separate Q, K, and V projections and retains the original FP32 time MLP.
The official ComfyOrg quant replaces that time path with a 1025×8 table, while my stock-compatible quants use a denser 4097×16 table. DT-sQKV has no time table at all: it evaluates the time embedding dynamically at every denoising step. Since H3 is designed to model complex, non-uniform temporal dynamics, this may better preserve combinations of slow and fast motion within the same video. It is not a direct per-frame speed controller, but the model’s original continuous time path remains intact.
Separate Q/K/V does not reduce the stored weight size by itself, but my memory-efficient Sage Attention patch can process the projections separately and may reduce peak temporary memory. A controlled VRAM A/B test is still needed to confirm the exact saving, but your result is already very interesting.
I had to also tweak the turbo lora (the attention layers were dismissed, it went unnoticed since the quality is so good), now it's flawless, I tested many flavors of minimax h3, yours is definitly my model of choice, keep on going, and again: thank you!