result_final.json

{
  "kernel": "fused_moe_kernel_gptq_awq (int4 W4A16 MoE GEMM, vLLM Triton, ~66% device time, live on ROCm)",
  "gemm_tuning": {
    "method": "vllm benchmark_moe.py autotune (trimmed ROCm space, 600 cfgs, M=16..1024)",
    "micro_speedup_geomean": 1.2741,
    "config_file": "E=384,N=512,device_name=AMD_Instinct_MI300X,dtype=int4_w4a16.json",
    "e2e_result": "REGRESSED -7..8% (243 vs 261-266 tok/s)",
    "verdict": "REJECTED for deployment: overfit synthetic uniform-routing micro-bench; slower on real decode distribution"
  },
  "kernel_level": {
    "method": "PerfSkills: hoist per-iter int4 scale/zp pointer divides out of K-loop; contiguity hints; mask-collapse when block_k divisible",
    "micro_speedup_geomean": 1.1,
    "correctness": "PASS (rtol/atol=2e-2 vs installed-original under identical inputs)",
    "e2e_result": "NEUTRAL (261.7 vs 266.5 baseline; median TPOT 139.8 vs 140.9)",
    "verdict": "KEPT: numerically identical, isolated-faster, E2E-safe; spliced into installed fused_moe.py"
  },
  "gsm8k_exact_match": 0.965
}