MoonMath challenge
Higgs quantization of AR-DiT
Quantize NVIDIA LongLive autoregressive DiT inference to 4-bit with Higgs and QJL, then evaluate generation quality.
Quantize both weights and activations, including the KV cache, to 4-bit. Benchmark whether the quantized model still produces acceptable generations.
Focus the evaluation on output quality. Inference speed, throughput, memory savings, and GPU selection are outside the scope of this challenge.