zechengtang/Spark-H3

VRAM-efficient MiniMax H3 for ComfyUI: QKV/RoPE/MLP activation chunking, automatic OOM chunk reduction, and CK/SLA/Sol/Hybrid attention for SM75 and SM80+ GPUs.

Veda Sparse Attention for ComfyUI

MiniMax-H3 33B audio+video DiT accelerated 1.54x on a single NVIDIA DGX Spark (GB10, sm_121): NVIDIA Sol-Engine FirstBlockCache + batched VAE decode ported to single-GPU ComfyUI, Sol-Attn patched for compute 12.1, kijai Triton INT8/TMA kernels. Benchmarks, charts, one-shot installer, agent recipe.

Video Delta Net hybrid attention for MiniMax-H3 as a native ComfyUI node.

MiniMax-H3 multi-GPU inference acceleration: ~12x on 8x RTX 6000D (step distillation + SageAttention2 + Cache-DiT + torch.compile, TP2+Ulysses4 on sglang)

MiniMax H3 Ref2VA acceleration for ComfyUI on an RTX 4070 12GB: 4-bit FC1 weights + Streaming VSA sparse attention. 1024x1792 video with audio in about 3 min 17 s, about 3.4x faster than dense attention.

NVIDIA Sol-Attn for ComfyUI / Triton kernel on SM89 - SM121, with zero-copy MiniMax H3 nodes: memory-efficient attention, scheduled tau with graph preview, and feed-forward chunking. Measured 1.14–1.44× vs SageAttention and −37% MLP peak VRAM on H3

24GB-optimized VDN-H3 hybrid attention implementation for MiniMax H3 in ComfyUI