{
 "conv_layer_ms": 18.4,
 "attn_layer_fixed_ms": 15.0,
 "lm_head_ms": 51.9,
 "rows": [
  {
   "context": 256,
   "attn_core_ms": 0.12,
   "full_layer_ms": 15.1,
   "sliding_layer_ms": 15.1,
   "lfm3_step_ms": 456,
   "all_full_24_step_ms": 414,
   "ratio": 0.91
  },
  {
   "context": 1024,
   "attn_core_ms": 0.35,
   "full_layer_ms": 15.3,
   "sliding_layer_ms": 15.1,
   "lfm3_step_ms": 458,
   "all_full_24_step_ms": 419,
   "ratio": 0.92
  },
  {
   "context": 4096,
   "attn_core_ms": 3.41,
   "full_layer_ms": 18.4,
   "sliding_layer_ms": 15.1,
   "lfm3_step_ms": 476,
   "all_full_24_step_ms": 493,
   "ratio": 1.03
  },
  {
   "context": 16384,
   "attn_core_ms": 55.88,
   "full_layer_ms": 70.8,
   "sliding_layer_ms": 15.1,
   "lfm3_step_ms": 791,
   "all_full_24_step_ms": 1752,
   "ratio": 2.21
  },
  {
   "context": 65536,
   "attn_core_ms": 205.41,
   "full_layer_ms": 220.4,
   "sliding_layer_ms": 15.1,
   "lfm3_step_ms": 1688,
   "all_full_24_step_ms": 5341,
   "ratio": 3.16
  },
  {
   "context": 131072,
   "attn_core_ms": 377.97,
   "full_layer_ms": 392.9,
   "sliding_layer_ms": 15.1,
   "lfm3_step_ms": 2724,
   "all_full_24_step_ms": 9482,
   "ratio": 3.48
  }
 ],
 "method": "best of 3 runs x best of 15 reps per primitive; random weights, real shapes; int8 dynamic linears; fp32 attention core; 2 threads; shared box at load ~10 (so absolute ms are inflated; ratios are the point)"
}