V6.5-attn-v2 cleanup: remove legacy v6_5_7ds_ewc_w8a8_benchmark.json
Browse files
v6_5_7ds_ewc_w8a8_benchmark.json
DELETED
|
@@ -1,41 +0,0 @@
|
|
| 1 |
-
{
|
| 2 |
-
"benchmark": "EWC+W8A8 eval with active dequantization (V6.5-7ds)",
|
| 3 |
-
"config": {
|
| 4 |
-
"model": "TestModel(64-128-32)",
|
| 5 |
-
"n_linears": 2,
|
| 6 |
-
"n_bits": 8,
|
| 7 |
-
"alpha_smoothquant": 0.5,
|
| 8 |
-
"calibration_samples": 32,
|
| 9 |
-
"n_iterations": 100
|
| 10 |
-
},
|
| 11 |
-
"results": {
|
| 12 |
-
"baseline_float": {
|
| 13 |
-
"penalty": 7.579145386815071,
|
| 14 |
-
"time_ms_per_call": 0.04622459411621094
|
| 15 |
-
},
|
| 16 |
-
"w8a8_with_dequant": {
|
| 17 |
-
"penalty": 7.581265166401863,
|
| 18 |
-
"time_ms_per_call": 0.1355147361755371,
|
| 19 |
-
"relative_error": 0.00027968583245277626
|
| 20 |
-
},
|
| 21 |
-
"w8a8_no_dequant_broken": {
|
| 22 |
-
"penalty": 1102808.2418839484,
|
| 23 |
-
"time_ms_per_call": 0.09852170944213867,
|
| 24 |
-
"relative_error": 145504.6191166922
|
| 25 |
-
}
|
| 26 |
-
},
|
| 27 |
-
"analysis": {
|
| 28 |
-
"dequant_preserves_accuracy": true,
|
| 29 |
-
"dequant_relative_error": 0.00027968583245277626,
|
| 30 |
-
"int8_relative_error": 145504.6191166922,
|
| 31 |
-
"dequant_overhead_ms": 0.08929014205932617,
|
| 32 |
-
"dequant_overhead_pct": 193.16587579946358,
|
| 33 |
-
"conclusion": "V6.5-7ds EWC+W8A8 eval com dequant: erro dequant=0.000280 (< 0.1 = OK), erro int8 direto=145504.619117."
|
| 34 |
-
},
|
| 35 |
-
"ewc_config": {
|
| 36 |
-
"eval_mode_penalty": true,
|
| 37 |
-
"skip_som_filled_neurons": true,
|
| 38 |
-
"lambda_ewc": 100.0,
|
| 39 |
-
"fisher_n_samples": 32
|
| 40 |
-
}
|
| 41 |
-
}
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|