diff --git a/STATE.md b/STATE.md index 49cfde1..866ce0d 100644 --- a/STATE.md +++ b/STATE.md @@ -4,21 +4,26 @@ ## 现在在跑什么 -**两档都停着** —— 2026-10-09 11:30 业主叫停，先修 bug 再重启；进程已按 PID 全部清掉，GPU encoder 0 %。 -**重启前先读「第 10 个 bug：MemoryError」和「8 个 `.new.mp4` 残留」两节**，并先腾内存。 - -原设计：两个 HEVC 档在跑，各自独立的锁（`run-x265.lock` / `run-nvenc.lock`；x265 = E 核，nvenc = GPU ASIC，互不争抢）。 - -**它们不在 Hermes 的进程管理里** → `process_manage` 看不到，也没有退出通知：只能看日志、按 PID 杀。 +**只有 x265 一档在跑**（2026-10-09 20:51:41 起，单档启动器 `run/relaunch_x265.sh`）。 +**nvenc 档故意不启动** —— GPU 被 TubeSync 的 58 个文件修复批次占着（20:36 起，与铁律一致：同一颗编码器不能两个任务）。 ```bash -# 断点续跑（幂等：已转好的自动跳过，失败项文件未动，下一轮自动重做），每档最多重试 8 次 -bash /d/tmp/xcode/relaunch_tiers.sh - -# 看进度 -tail -f /d/tmp/xcode/run_x265.log /d/tmp/xcode/run_nvenc.log +# 只起 x265（E 核，纯 CPU）—— 仓库源 run/relaunch_x265.sh 拷到 /d/tmp/xcode/ 再跑 +powershell -NoProfile -Command 'Start-Process -FilePath "C:\Program Files\Git\bin\bash.exe" -ArgumentList "D:/tmp/xcode/relaunch_x265.sh" -WindowStyle Hidden' +# ⚠ 参数必须只有脚本路径；写成 -ArgumentList "-lc","bash <脚本>" 会起一个交互式 bash，什么都不跑 +tail -f /d/tmp/xcode/run_x265.log ``` +### 2026-10-09 16:51:58：两档是**整机重启**停的，不是 bug + +（两档都不在 Hermes 的进程管理里 → `process_manage` 看不到，也没有退出通知：只能看日志，按 PID 杀。） +`LastBootUpTime = 2026-10-09 16:51:58`；x265 日志停在 16:50:42（404/589）、nvenc 停在 16:51:30 +（308/589）。11:48 由 relauncher 重新拉起、跑到重启为止。锁文件里那两个 PID（23792 / 50128）已是死号 —— +`transcode.Lock` 是**过期感知**的（age > 300 s 且 pid 已死即接管），不需要手工清锁。 + +重启后内存正常：物理空闲 12.9–13 GB、提交 50.3/55.9 GB（11:19 那次 MemoryError 时是 96.5%）。 +⚠ 但提交上限从 79.6 GB 缩到 55.9 GB：D: 已 92% 满，pagefile 长不上去。 + ### 为什么改成"脱管"启动 —— 10:37 那次全停的教训（第 9 个 bug） 上一轮的 x265 / nvenc 是 Hermes 后台进程。**用户停会话时它们被一起杀了**：两个都停在编码到一半 diff --git a/run/relaunch_x265.sh b/run/relaunch_x265.sh new file mode 100644 index 0000000..e867aa9 --- /dev/null +++ b/run/relaunch_x265.sh @@ -0,0 +1,45 @@ +#!/usr/bin/env bash +# Start ONLY the x265 tier of the 樊登读书会 conversion, DETACHED from Hermes. +# +# Why a separate launcher: run/relaunch_tiers.sh hardcodes `for t in x265 nvenc`, and the nvenc +# tier must NOT start while another job owns the GPU -- the TubeSync 58-file repair batch has held +# the GPU encoder since 2026-10-09 20:36. x265 is pure CPU pinned to the E-cores (0x3FF0000), so it +# shares nothing with the GPU job; the only shared resource is the RaiDrive mount, and 2 concurrent +# share jobs is the documented safe limit (3 produced `[Errno 9] Bad file descriptor`). +# +# Detached for the reason recorded in STATE.md: a Hermes-scoped background job is killed when the +# session ends (measured 2026-10-08 10:37, both tiers killed mid-encode), so the launcher is started +# by `powershell Start-Process ... -WindowStyle Hidden` and its parent exits immediately. No exit +# notification -- watch the log; stop by PID only, never by image name. +# +# Idempotent: already-converted files are skipped, a failure leaves the live file untouched and the +# next attempt retries it. Lock: C:\AI\transcode-run\run-x265.lock (stale-aware: age > 300 s AND a +# dead pid is taken over automatically, so the lock left by the 16:51 reboot needs no cleanup). +# +# Started correctly (the argument form matters): the script path goes in -ArgumentList ALONE. +# Start-Process -FilePath "C:\Program Files\Git\bin\bash.exe" -ArgumentList "" -WindowStyle Hidden +# NOT "-lc","bash