According to @_avichawla, NVIDIA's closed-form KV cache transfer skips prefill and runs 2.7–25x faster, retaining up to 98% accuracy across model scales.