⚡ Fanout Diffusion: 1-Step Generative Query Expansion

1-Step EDM Consistency Model with 1-Bit MMA Tensor Core Transformer

Generate 10 semantically aligned, diverse subquery expansion slots in < 1 millisecond (0.33 ms native CUDA MMA, ~1.2 ms CPU ONNX).

40 120

Inference Performance & Semantic Metrics

Decoded 10-Slot Subquery Expansions

Examples