MoonMath challenge
Higgs quantization of AR-DiT
Quantize NVIDIA LongLive autoregressive diffusion inference to four bits and evaluate the output quality.
Use Higgs and QJL to quantize NVIDIA’s LongLive inference to 4-bit. Quantize both weights and activations, including the KV cache, and benchmark output quality.
Inference performance, throughput, and GPU choice are not evaluation criteria for this challenge.