Natural Memory NM2.1: 记忆路由器分叉、数据集缺陷修复与全轴评测证据

- 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
This commit is contained in:
WpyQwq
2026-09-19 11:11:31 +08:00
commit 643e22ecb9
484 changed files with 306821 additions and 0 deletions
+86
View File
@@ -0,0 +1,86 @@
{
"V2-128 deployed(v3)": {
"parameters": 2037774,
"rounds": 5,
"single_ms_min": 1.1413500014896272,
"single_ms_median": 1.155849999122438,
"single_ms_max": 1.1732999992091209,
"single_ms_spread_pct": 2.799316395303309,
"batch64_qps_median": 54038.53621622001,
"batch256_qps_median": 199554.12120495233,
"batch64_qps_spread_pct": 22.594620069656955,
"batch256_qps_spread_pct": 7.049645899749284
},
"V2-128 v6 best": {
"parameters": 2037774,
"rounds": 5,
"single_ms_min": 1.1358499996276805,
"single_ms_median": 1.167899999927613,
"single_ms_max": 1.196150000396301,
"single_ms_spread_pct": 5.308799646818349,
"batch64_qps_median": 54149.18098951274,
"batch256_qps_median": 196829.20449441153,
"batch64_qps_spread_pct": 20.236040711481348,
"batch256_qps_spread_pct": 59.54630444438826
},
"V2-128 v6 final": {
"parameters": 2037774,
"rounds": 5,
"single_ms_min": 1.1389999999664724,
"single_ms_median": 1.165349998700549,
"single_ms_max": 1.1786999984906288,
"single_ms_spread_pct": 3.4855134789574196,
"batch64_qps_median": 53471.46796235134,
"batch256_qps_median": 198025.9290094514,
"batch64_qps_spread_pct": 16.333515374225275,
"batch256_qps_spread_pct": 15.161085444193706
},
"XL-128 v6 best": {
"parameters": 6714639,
"rounds": 5,
"single_ms_min": 1.5637500000593718,
"single_ms_median": 1.578349998453632,
"single_ms_max": 1.6234499998972751,
"single_ms_spread_pct": 3.817745792846468,
"batch64_qps_median": 38630.547104467114,
"batch256_qps_median": 141500.569342867,
"batch64_qps_spread_pct": 9.66360135427591,
"batch256_qps_spread_pct": 6.638103530033454
},
"XL-128 v6 final": {
"parameters": 6714639,
"rounds": 5,
"single_ms_min": 1.5639999983250163,
"single_ms_median": 1.5958999993017642,
"single_ms_max": 1.6187000001082197,
"single_ms_spread_pct": 3.497442573004151,
"batch64_qps_median": 36843.43841344889,
"batch256_qps_median": 134976.95897715187,
"batch64_qps_spread_pct": 16.69087035116882,
"batch256_qps_spread_pct": 13.146859485706125
},
"XL-512 v6 best": {
"parameters": 7898127,
"rounds": 5,
"single_ms_min": 1.545650002299226,
"single_ms_median": 1.5855000019655563,
"single_ms_max": 1.6239000015048077,
"single_ms_spread_pct": 5.062594965818999,
"batch64_qps_median": 39909.70428801726,
"batch256_qps_median": 94647.25411335309,
"batch64_qps_spread_pct": 15.919053379775894,
"batch256_qps_spread_pct": 5.0702261917617895
},
"V2-512 v6 best": {
"parameters": 4741902,
"rounds": 5,
"single_ms_min": 1.1235000001761364,
"single_ms_median": 1.162250000561471,
"single_ms_max": 1.1778999996749917,
"single_ms_spread_pct": 4.84201152561876,
"batch64_qps_median": 53757.11860745415,
"batch256_qps_median": 154085.05975293368,
"batch64_qps_spread_pct": 14.384722867748353,
"batch256_qps_spread_pct": 12.811712690590992
}
}