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
2026-09-19 11:11:31 +08:00
commit 643e22ecb9
484 changed files with 306821 additions and 0 deletions
@@ -0,0 +1,78 @@
+param(
+    [int]$Steps = 100000,
+    [int]$EvalInterval = 500,
+    [int]$CheckpointInterval = 5000,
+    [int]$GpuMemoryGb = 9
+)
+
+$ErrorActionPreference = 'Stop'
+
+$Python = 'C:\Users\Administrator\miniconda3\envs\LLM\python.exe'
+$WorkDir = 'H:\Memory'
+$OutputDir = 'H:\Memory\V2_dpskw\checkpoints\natural_memory_v2_router_512'
+$FeatureCacheDir = 'H:\Memory\V2_dpskw\checkpoints\natural_memory_v2_router_v3\feature_cache'
+$TrainFile = 'V2_dpskw/data/router_training_v3/train.jsonl'
+$EvalFile = 'V2_dpskw/data/router_training_v3/eval.jsonl'
+
+if (-not (Test-Path -LiteralPath $Python)) {
+    throw "Python executable not found: $Python"
+}
+if (-not (Test-Path -LiteralPath (Join-Path $WorkDir $TrainFile))) {
+    throw "Training file not found: $TrainFile"
+}
+if (-not (Test-Path -LiteralPath (Join-Path $WorkDir $EvalFile))) {
+    throw "Evaluation file not found: $EvalFile"
+}
+if (-not (Test-Path -LiteralPath $FeatureCacheDir)) {
+    throw "Feature cache not found: $FeatureCacheDir"
+}
+if ($Steps -lt 1) { throw 'Steps must be positive.' }
+if ($EvalInterval -ne 500) { throw 'This production protocol requires EvalInterval=500.' }
+if ($CheckpointInterval -lt 1) { throw 'CheckpointInterval must be positive.' }
+
+New-Item -ItemType Directory -Force -Path $OutputDir | Out-Null
+
+$existing = Get-CimInstance Win32_Process | Where-Object {
+    $_.Name -eq 'python.exe' -and
+    $_.CommandLine -match 'train_memory_router_large' -and
+    $_.CommandLine -match 'natural_memory_v2_router_512'
+}
+if ($existing) {
+    throw "A 512d router training process is already running: $($existing.ProcessId -join ', ')"
+}
+
+$stdout = Join-Path $OutputDir 'training_stdout.log'
+$stderr = Join-Path $OutputDir 'training_stderr.log'
+$argumentLine = @(
+    '-m V2_dpskw.train_memory_router_large',
+    '--train-file V2_dpskw/data/router_training_v3/train.jsonl',
+    '--eval-file V2_dpskw/data/router_training_v3/eval.jsonl',
+    '--model-path V2_dpskw/qwen3_5_4b_natural_memory_v2',
+    '--output-dir V2_dpskw/checkpoints/natural_memory_v2_router_512',
+    '--feature-cache-dir V2_dpskw/checkpoints/natural_memory_v2_router_v3/feature_cache',
+    '--router-dim 512',
+    '--num-heads 8',
+    '--max-hops 3',
+    '--gpu-memory-gb ' + $GpuMemoryGb,
+    '--encode-batch-size 1',
+    '--steps ' + $Steps,
+    '--batch-size 64',
+    '--eval-batch-size 128',
+    '--eval-interval ' + $EvalInterval,
+    '--checkpoint-interval ' + $CheckpointInterval
+) -join ' '
+
+$process = Start-Process -FilePath $Python -WorkingDirectory $WorkDir `
+    -ArgumentList $argumentLine -RedirectStandardOutput $stdout `
+    -RedirectStandardError $stderr -WindowStyle Hidden -PassThru
+
+[pscustomobject]@{
+    pid = $process.Id
+    router_dim = 512
+    router_parameters = 4741902
+    steps = $Steps
+    eval_interval = $EvalInterval
+    checkpoint_interval = $CheckpointInterval
+    output_dir = $OutputDir
+    feature_cache = $FeatureCacheDir
+} | ConvertTo-Json -Compress