devils-agent / reports /runtime-optimization.json
devildasdf's picture
Optimize CPU feature encoding and add bounded batch inference with measured parity
c7d6933 verified
Raw History Blame Contribute Delete
812 Bytes
{
"scope": "Windows CPU, two Torch threads, synthetic data; unchanged model weights",
"splits": {
"validation": {
"rows": 480,
"encoding_bit_exact": true,
"raw_step_match": 1.0
},
"test": {
"rows": 480,
"encoding_bit_exact": true,
"raw_step_match": 1.0
},
"novel_wording": {
"rows": 480,
"encoding_bit_exact": true,
"raw_step_match": 0.5583333333333333
}
},
"decision_parity_rows": 160,
"old_predict": {
"median_ms": 2.1032000076957047,
"p95_ms": 3.303700010292232
},
"new_predict": {
"median_ms": 0.8837499772198498,
"p95_ms": 1.3269999762997031
},
"batch16": {
"median_ms": 6.252999999560416,
"p95_ms": 9.634000016376376
},
"speedup": 2.37985862733719
}