Flash-dLLM attacks diffusion-model inference I/O

What's this about?

Flash-dLLM fuses KV-cache operations to reduce GPU memory movement, then lets the diffusion model draft and verify parallel tokens itself.

Why is memory I/O the bottleneck?

Reusing cached attention states saves computation but repeatedly moving them can dominate runtime. A fused kernel avoids redundant transfers.

Why does this matter?

1. The method is training-free and needs no auxiliary draft model.
2. Reported speedups are 5.1× on GSM8K and 11× on HumanEval.
3. Broader models, hardware and quality audits are still needed.