33 lines
1.2 KiB
JSON
33 lines
1.2 KiB
JSON
{
|
|
"contract": {
|
|
"moe_mapping": "standalone moe_grouped_gemm and serving MoE both include expert prepare/finalize plus expert GEMMs",
|
|
"real_anchor_ms": 410.0,
|
|
"serving_sample": "longest execute window on critical TP rank",
|
|
"simulator_sample": "mean of first nine q8192/ctx0 single-request chunks"
|
|
},
|
|
"counterfactual": {
|
|
"moe_only_counterfactual_bias": -0.02523382099105694,
|
|
"moe_only_counterfactual_total_ms": 399.65413339366665,
|
|
"moe_residual_fraction": 0.8075231403622317,
|
|
"moe_residual_ms": 43.40535640199994,
|
|
"moe_scale": 1.2536568227334743,
|
|
"total_residual_ms": 53.751223008333284
|
|
},
|
|
"decision": "inject_tp2_prefill_moe_and_replay",
|
|
"gates": {
|
|
"moe_explains_at_least_70pct": true,
|
|
"moe_only_counterfactual_within_5pct": true,
|
|
"serving_reproduces_anchor_within_5pct": true
|
|
},
|
|
"measurements_ms": {
|
|
"real_anchor": 410.0,
|
|
"serving_execute_wall": 408.19065,
|
|
"serving_moe": 214.52378299999995,
|
|
"serving_non_moe": 193.66686700000005,
|
|
"simulator_moe_grouped_gemm": 171.118426598,
|
|
"simulator_non_moe": 185.1303503936667,
|
|
"simulator_total": 356.2487769916667
|
|
},
|
|
"schema": "frontier-tp2-prefill-serving-smoke-verdict.v1"
|
|
}
|