- 引入 MemoryRouterXL 与 v5/v6 流式多线程训练/编码管线 - 修复 prepare_memory_router_dataset 候选池重建缺陷(mega 家族 3568x 加速,输出逐字节相同) - 修复 v5 被破坏的拒答与多跳标签(train 未知样本 319 -> 16319,multi_hop 平均正例 1.00 -> 2.00) - 同存储预算下 V2-128 v6 逐轴 22/22 通过:Top-1 41.12% -> 94.62%,未知拒答 0.00% -> 100.00% - 记录三条被实测推翻的显然优化(logits_to_keep=1 反而慢 55%、XL 容量未带来收益) - 记忆手术跨架构可移植性 14/14,读写关闭时与原生模型逐位相同
26 lines
899 B
JSON
26 lines
899 B
JSON
{
|
|
"V2-128 v6 final": {
|
|
"parameters": 2037774,
|
|
"rounds": 7,
|
|
"single_ms_min": 1.149349998740945,
|
|
"single_ms_median": 1.2042000016663224,
|
|
"single_ms_max": 1.3496499996108469,
|
|
"single_ms_spread_pct": 17.427241579094304,
|
|
"batch64_qps_median": 53455.38981773372,
|
|
"batch256_qps_median": 190102.7742132792,
|
|
"batch64_qps_spread_pct": 20.531002864672395,
|
|
"batch256_qps_spread_pct": 13.093009305269373
|
|
},
|
|
"REPLAY-128 final": {
|
|
"parameters": 2037774,
|
|
"rounds": 7,
|
|
"single_ms_min": 1.167499998700805,
|
|
"single_ms_median": 1.2087499999324791,
|
|
"single_ms_max": 1.2399500010360498,
|
|
"single_ms_spread_pct": 6.205567658746657,
|
|
"batch64_qps_median": 48526.02206733375,
|
|
"batch256_qps_median": 192599.94881524373,
|
|
"batch64_qps_spread_pct": 78.07160732122072,
|
|
"batch256_qps_spread_pct": 33.117117099551436
|
|
}
|
|
} |