Record adopted seed1 PASS and actual two-worker launch with first-epoch cost gates

This commit is contained in:
MeltAstra committed 2026-11-18 14:46:16 +08:00
1 parent 8c5020c5cb
commit 8108e52cd5
28 files changed
+1230

No files matched your search

+10
View File
@@ -1,3 +1,13 @@
## 最新 seed1首两折成本门PASS(2026-11-18 14:45)
原唯一1365333队列/1365444与1365428 worker已进入epoch2;首epoch门均PASS,fold0投影19207.933567s、fold1为21651.438886s,合计投影115019.372453/118800GPU-s,含pending2/3/4各22320整帽和7200postfit预留,不是五折全实测。phase3原72421.315999一次,总投影187440.688452/612000。first-two-epoch-audit记录原始history/门/进程;ac45e2de/95d23536源码与Track A原23pins均PASS。无重启/重试/后处理,原动态队列继续,后续若成本门停止须保留原因,不自动放宽帽。
## 最新 seed1唯一训练已真实启动(2026-11-18 14:43,Asia/Shanghai)
新实际PASS79b55546原样freeze并validate,implementationac45e2de/contract95d23536/proof04592d0f/comparisoncba51291/38pins保持;旧BLOCKED不变。唯一service meltmind-phase3-seed1-20261118,coord1365333;fold0 guard1365362/worker1365444 GPU0,fold1 guard1365363/worker1365428 GPU1,真实epoch1进度147/320和180件快照,actual-start-audit已记录,assignment hash和三idle polls PASS。首epoch成本门尚未实观;不能把启动当成本PASS。pending2/3/4按动态释放卡调度,22320/fold、118800含7200postfit、phase3旧72421.315999一次。
启动前available RAM46347366400B、home free237850185728B,初始两worker+2GiB余量及100GiB磁盘通过。原Track A1262364/两worker在2/3继续,GPU2为其postfit优先,item1COMPLETE不重启。seed1只checkpoint冻结,不预测/评分;postfit独立审阅保留。不要因会话idle重启或修改固定源码。真实状态看新root complete/live-progress/每fold原始receipts。
## 最新 seed1同次轻审空闲分配修复待回(2026-11-18)
旧d2f552d1获BLOCKED ba4c62e8(唯一S1-DISPATCH-IDLE),原件/evidence308003c7/原源码与合同全部保留于artifacts/phase3-seed1-idle-fix-2026-11-18。只改common/queue/test三文件:eligible先实测显存≤128MiB且util0,再分配/写hash assignment/消费pending;busy未分配不算attempt,guard三polls/锁和分配后竞态STOP不变。18CPU合成PASS(busy0/idle1、全busy不分配等),科学worker/guard及预算不变。新implementationac45e2de82f1c336aba75c2d6908ce3776603fec/contract95d23536a653ef2d2598127cabe29c3b4bcab40bcc0f2fafee0a2931893395ff/proof04592d0fd72cc8a71201e5dea145bc8c618ff05d15a78b257f0d2dc60ab54431,comparisoncba51291和38map,待同次Assist复核;尚无review/started/fit,旧BLOCKED入口不得启动。item1已COMPLETE不重跑,Track A原训练和已采用postfit PASS983bba99不动。
+11
View File
@@ -0,0 +1,11 @@
# Seed1 actual launch — 2026-11-18
The exact revised PASS79b5554654f8bfe5a06be49bf7af5cfe7d6cf467bdf55f53994b342e00545874 was copied unchanged, frozen and validated against ac45e2de /95d23536 /04592d0f /cba51291 and all38code pins. Original BLOCKED ba4c62e8 remains immutable. The review confirms implementation only; it is not a training or cost result.
One service `meltmind-phase3-seed1-20261118.service` started at14:43 Asia/Shanghai, Restart=no. Actual coordinator1365333, fold0 guard1365362/worker1365444 onGPU0, fold1 guard1365363/worker1365428 onGPU1. Two workers reached epoch1. Assignment hashes match both command lines; independent guard idle polls passed. Original Track A coordinator1262364 and workers1262432/1262422 continue onGPU2/3. Item1 remains complete.
Startup audit found46347366400B available RAM and237850185728B free disk. It required two16GiB initial worker allowances plus2GiB overhead, and100GiB disk. This checks initial availability, not a guarantee about future host load. Later dynamic release and sampled aggregate guards remain active. GPU2 stays reserved for Track A postfit. No duplicate seed1 service/process existed.
Root `/home/frank14f/MeltMind-runs/phase3-seed1-2026-11-18` retains review-adoption, host-precheck, launch-attempt/service-launch, started, per-job assignments/dispatch/idle polls and actual-start-audit. Initial evidence is before the first-epoch cost gates; read subsequent receipts for their actual outcome. The original118800GPU-s cap,22320/fold and7200postfit reserve are unchanged; phase3 old72421.31599938683s counted once. No retries, no source mutation, no test predictions or scoring. Stop at frozen five endpoints; postfit has its own later pre-score review.
At14:45 the two initial first-epoch gates passed and both workers entered epoch2. Per-fold projections19207.933567/21651.438886s; total115019.372453/118800s includes all three pending full22320s caps plus7200postfit. Phase3 total projection187440.688452/612000s includes item1 once. The first-two-epoch-audit retains actual histories and receipts. This is not a five-fold measurement or final result; later every-epoch guards remain binding.