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 +1,107 @@
+param(
+    [double]$RefreshSeconds = 0.5
+)
+
+$TrainingRoot = 'H:\Memory\V2_dpskw\checkpoints\natural_memory_v2_router_v3_100k'
+$MetricsPath = Join-Path $TrainingRoot 'metrics.jsonl'
+$SummaryPath = Join-Path $TrainingRoot 'memory_router_large_training.json'
+$TargetSteps = 100000
+
+$Host.UI.RawUI.WindowTitle = 'Natural Memory v2 Router V3 - Live Monitor'
+
+function Get-TrainingProcess {
+    Get-CimInstance Win32_Process -ErrorAction SilentlyContinue |
+        Where-Object {
+            $_.CommandLine -and
+            $_.CommandLine -like '*train_memory_router_large*' -and
+            $_.CommandLine -like '*router_v3*'
+        } |
+        Select-Object -First 1
+}
+
+function Read-Events {
+    if (-not (Test-Path -LiteralPath $MetricsPath)) {
+        return @()
+    }
+    $events = @()
+    foreach ($line in (Get-Content -LiteralPath $MetricsPath -Tail 300 -ErrorAction SilentlyContinue)) {
+        try {
+            $events += ($line | ConvertFrom-Json -ErrorAction Stop)
+        } catch {
+            # Ignore the one line that may be being appended right now.
+        }
+    }
+    return $events
+}
+
+function Get-GpuStatus {
+    $raw = & nvidia-smi --query-gpu=name,memory.used,memory.total,utilization.gpu --format=csv,noheader,nounits 2>$null
+    if (-not $raw) {
+        return $null
+    }
+    $parts = $raw -split ',' | ForEach-Object { $_.Trim() }
+    if ($parts.Count -lt 4) {
+        return $null
+    }
+    return [pscustomobject]@{
+        Name = $parts[0]
+        UsedMiB = $parts[1]
+        TotalMiB = $parts[2]
+        Utilization = $parts[3]
+    }
+}
+
+while ($true) {
+    $procInfo = Get-TrainingProcess
+    $events = Read-Events
+    $lastTrain = @($events | Where-Object { $_.event -eq 'train' } | Select-Object -Last 1)[0]
+    $lastEval = @($events | Where-Object { $_.event -in @('eval', 'eval_resume') } | Select-Object -Last 1)[0]
+    $gpu = Get-GpuStatus
+
+    Clear-Host
+    Write-Host 'Natural Memory v2 - Router V3 Live Monitor' -ForegroundColor Cyan
+    Write-Host ('Updated: {0}' -f (Get-Date -Format 'yyyy-MM-dd HH:mm:ss'))
+    Write-Host ('Output: {0}' -f $TrainingRoot)
+    Write-Host ''
+
+    if ($procInfo) {
+        $proc = Get-Process -Id $procInfo.ProcessId -ErrorAction SilentlyContinue
+        $ram = if ($proc) { '{0:N2} GB' -f ($proc.WorkingSet64 / 1GB) } else { 'N/A' }
+        $cpu = if ($proc) { '{0:N1}s' -f $proc.CPU } else { 'N/A' }
+        Write-Host ('Status: RUNNING   PID={0}   RAM={1}   CPU={2}' -f $procInfo.ProcessId, $ram, $cpu) -ForegroundColor Green
+    } else {
+        Write-Host 'Status: training process not found' -ForegroundColor Yellow
+    }
+
+    if ($lastTrain) {
+        $percent = [math]::Min(100, 100 * [double]$lastTrain.step / $TargetSteps)
+        Write-Host ('Progress: {0}/{1} ({2:N1}%)' -f $lastTrain.step, $TargetSteps, $percent)
+        Write-Host ('Loss={0:N5}  Route={1:N5}  Need={2:N6}  Hop={3:N6}  LR={4:E3}' -f [double]$lastTrain.loss, [double]$lastTrain.route_loss, [double]$lastTrain.need_loss, [double]$lastTrain.hop_loss, [double]$lastTrain.lr)
+    } else {
+        Write-Host 'Progress: waiting for training log...' -ForegroundColor DarkYellow
+    }
+
+    if ($lastEval) {
+        Write-Host ''
+        Write-Host ('Last evaluation: step {0}' -f $lastEval.step) -ForegroundColor Magenta
+        Write-Host ('Top1={0:P2}  RecallAt3={1:P2}  MRR={2:P2}' -f [double]$lastEval.route_top1, [double]$lastEval.route_recall_at3, [double]$lastEval.route_mrr)
+        Write-Host ('NeedF1={0:P2}  Abstention={1:P2}  Hop={2:P2}' -f [double]$lastEval.need_f1, [double]$lastEval.abstention_accuracy, [double]$lastEval.hop_accuracy)
+    }
+
+    if ($gpu) {
+        Write-Host ''
+        Write-Host ('GPU: {0}   VRAM: {1}/{2} MiB   Utilization: {3}%' -f $gpu.Name, $gpu.UsedMiB, $gpu.TotalMiB, $gpu.Utilization) -ForegroundColor DarkCyan
+    }
+
+    Write-Host ''
+    Write-Host 'Press Ctrl+C to close the monitor. Training will not be stopped.' -ForegroundColor DarkGray
+
+    if (-not $procInfo -and (Test-Path -LiteralPath $SummaryPath)) {
+        Write-Host ''
+        Write-Host 'Training complete. Final report is available.' -ForegroundColor Green
+        Write-Host 'Monitor is staying on the final result. Press Ctrl+C to close.' -ForegroundColor DarkGray
+        Start-Sleep -Milliseconds ([int]([math]::Max(100, $RefreshSeconds * 1000)))
+        continue
+    }
+    Start-Sleep -Milliseconds ([int]([math]::Max(100, $RefreshSeconds * 1000)))
+}