deepseek v3 q4_k_m vs q8_0 perplexity gap on code is bigger than claimed

tested both quants yesterday on same hardware (2x4090, batch=1, temp=0.2). claimed perplexity difference is ~2.1%, measured difference on our code generation workload (python, rust, typescript) is 4.7%. q4_k_m: avg perplexity 3.84 q8_0: avg perplexity 3.66 tested on 340 samples, context length 12k-18k. the gap gets worse past 15k context - q4_k_m degrades way harder is this pattern or our setup is wrong)