FastVideo/FastVideo-FastH3-4-step-Preview-v1-VSA-DataFree

Full fine-tune vs MiniMaxAI/MiniMax-H3. Fine-tune: 99% of the weights changed, across 0 of 0 layers.

Tensors changed: Attention 212/212, other 207/208, Embeddings 6/6, Norms 17/212

2 derivatives checked