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 +1,109 @@
+"""One-off migration: repoint feature-bank paths from D:\\ to H:\\Memory\\nm_cache.
+
+The feature banks were moved off D: because that volume ran short of space.  Only
+*live* references are rewritten -- defaults, run scripts and documentation.  Historical
+run records (``metrics.jsonl``, ``training_stdout.log``, ``router_v5_training.json``,
+``feature_bank_verification.json``) are deliberately left untouched: they are evidence of
+what actually ran and where the data lived at that moment, and rewriting them would
+falsify the record.
+
+Run with ``--apply`` to write; without it, the script only reports what would change.
+"""
+
+from __future__ import annotations
+
+import argparse
+import json
+from pathlib import Path
+
+ROOT = Path(__file__).resolve().parent
+
+NEW_ROOT = r"H:\Memory\nm_cache"
+
+#: old path fragment -> new path fragment (with the escaped form handled separately)
+MOVES = {
+    r"D:\nm_router_v6\feature_cache": NEW_ROOT + r"\nm_router_v6\feature_cache",
+    r"D:\nm_router_v5\feature_cache": NEW_ROOT + r"\nm_router_v5\feature_cache",
+    r"D:\nm_zero_overlap\feature_cache": NEW_ROOT + r"\nm_zero_overlap\feature_cache",
+    r"D:\nm_replay_v7\feature_cache": NEW_ROOT + r"\nm_replay_v7\feature_cache",
+    r"D:\nm_probe_bf16": NEW_ROOT + r"\nm_probe_bf16",
+}
+
+#: Live files only.  Historical logs and per-run training records are excluded on purpose.
+TARGETS = [
+    "bench_router_latency.py",
+    "build_replay_corpus.py",
+    "compare_record_rankers.py",
+    "chain_v5.ps1",
+    "chain_v6.ps1",
+    "eval_router_v5.py",
+    "probe_encoder_precision.py",
+    "train_router_v5.py",
+    "train_and_score_v6_128.ps1",
+    "train_and_score_v6.ps1",
+    "run_router_v6.ps1",
+    "run_router_v5.ps1",
+    "verify_feature_bank.py",
+    "README_FORK.md",
+    "V6_FINAL_REPORT.md",
+    "ZERO_OVERLAP_FINDINGS.md",
+    "ZERO_OVERLAP_PHASE.md",
+    "zero_overlap_phase.md",
+]
+
+
+def main() -> int:
+    parser = argparse.ArgumentParser(description=__doc__)
+    parser.add_argument("--apply", action="store_true")
+    args = parser.parse_args()
+
+    changes = []
+    for relative in TARGETS:
+        path = ROOT / relative
+        if not path.exists():
+            continue
+        original = path.read_text(encoding="utf-8")
+        text = original
+        for old, new in MOVES.items():
+            text = text.replace(old, new)
+            # Escaped form, e.g. inside Python docstrings and PowerShell here-strings.
+            escaped_old = old.replace("\\", "\\\\")
+            escaped_new = new.replace("\\", "\\\\")
+            text = text.replace(escaped_old, escaped_new)
+        if text != original:
+            count = sum(original.count(old) + original.count(old.replace("\\", "\\\\"))
+                        for old in MOVES)
+            changes.append({"file": relative, "replacements": count})
+            if args.apply:
+                path.write_text(text, encoding="utf-8")
+
+    report = {
+        "applied": args.apply,
+        "new_root": NEW_ROOT,
+        "files_changed": changes,
+        "total_replacements": sum(item["replacements"] for item in changes),
+        "note": ("historical run records were intentionally left unmodified: "
+                 "checkpoints/*/metrics.jsonl, checkpoints/*/training_stdout.log, "
+                 "checkpoints/*/router_v5_training.json, feature_bank_verification.json"),
+    }
+    print(json.dumps(report, ensure_ascii=False, indent=2))
+
+    # Verify no live reference remains.
+    stale = []
+    for relative in TARGETS:
+        path = ROOT / relative
+        if not path.exists():
+            continue
+        text = path.read_text(encoding="utf-8")
+        hits = [old for old in MOVES if old in text or old.replace("\\", "\\\\") in text]
+        if hits:
+            stale.append({"file": relative, "remaining": hits})
+    if stale:
+        print(json.dumps({"remaining_d_references": stale}, ensure_ascii=False, indent=2))
+        return 1
+    print(json.dumps({"remaining_d_references": []}, ensure_ascii=False))
+    return 0
+
+
+if __name__ == "__main__":
+    raise SystemExit(main())