Mistral-7B @ 8K: FP16=1073MB, 8bit=537MB, Naive=537MB, Triton=467MB(2.3x)
Llama-3-8B @ 8K: FP16=1073MB, 8bit=537MB, Naive=537MB, Triton=526MB(2.04x)
Key finding: Naive uint8 = same as uniform 8-bit on actual GPU
Triton true packing = 15% better than 8-bit on Mistral
Zero perplexity degradation on both models
