Record independent research parallelism and withdrawn unstarted throughput draft

This commit is contained in:
MeltAstra committed 2026-11-18 21:03:33 +08:00
1 parent ff0033d733
commit 5b6a953fc5
10 files changed
+1812

No files matched your search

@@ -729,3 +729,11 @@ source8a8ebd1f/contract4b21dedd/proof8baec54f/comparison828530e4/27pins不变,
原1650006/1650109/1650207队列和worker,fold4首轮57.020449043秒,单折投影14437.426020≤30960;pending fold3仍保留原30960整帽,含post7200及old47943.947195一次,累计投影100541.373214≤162000。fold4已真实进入epoch2;fold3尚未实测,不称全两折成本PASS。快照/首轮原history/门在artifacts/track-a-recovery-start-evidence-2026-11-18。无重复启动/改源/新计算范围,当前状态看live/terminal。
## 最新用户纠正:研究事项并行,吞吐benchmark撤回且未启动(2026-11-18)
用户通过TRON明确要不同研究工作的并行,不是同一训练1/2/3副本吞吐或算子优化;以后意见先由Claude领导裁量。Claude1795006714/6766已实读:撤回6592的benchmark,入口按c→1b→b→Track B构建,各自审后可并行;seed1只约束效果/噪声判决,不一律阻塞开工。每项训练入口一次审阅,test前独立审阅保持。GPU0–2最多3任务、GPU3严格1,当前frozen训练和postfit不改。现有pending时保留一张0–2不给新入口,不能停现有worker制造空卡或重派旧fold。
Melt只做25.5685s六次只读采样并写过两份未完成草稿;草稿已移出tools、逐字归档withdrawn-throughput-draft,从未运行benchmark/测试/冻结/读标签,无probe root/process。原seed1 PID1476593和三worker、Track A recovery PID1650006和fold4worker1650207均核在,78code pins全匹配;两队列无STOP/complete。资源及撤回证据artifacts/gpu-concurrency-assessment-2026-11-18;说明docs/independent-research-concurrency-2026-11-18.md。新共享capacity调度仅待新入口实现/审阅,没有已执行收益或新研究结果。
@@ -0,0 +1,28 @@
# Independent research concurrency: corrected scope, observed resources
User corrections relayed in native TRON messages and Slack1795006553/6743: parallelize distinct research ideas; do not optimize operators or run repeated-copy throughput benchmarks. User ideas go to Claude for decisions. Claude1795006714 withdraws the earlier1795006592 benchmark and orders new entry construction c→1b→b→Track B, with independent reviewed entries allowed to run concurrently. Seed1 is needed for decisions about effect versus noise, not universally before starting their training. Each entry still needs implementation review and its own pre-score review before test reads.
**No throughput benchmark was ever launched.** Only6 read-only resource samples over25.568506s and2 incomplete new source drafts were produced. Drafts were moved out of tools into withdrawn-throughput-draft with byte hashes; no test, preparation root, freeze, GPU workload or label access occurred. withdrawal-process-audit confirms no probe process/root, original seed1/Track A processes continue and both code maps match. Earlier inference to a1/2/3-copy experiment is explicitly withdrawn, not a completed result.
## Actual resource snapshot
| GPU | sampled VRAM MiB | utilization range % | temperature range C |
|---|---:|---:|---:|
| 0 | 1717–1717 | 0–27 | 51–51 |
| 1 | 1719–1719 | 7–28 | 51–52 |
| 2 | 1839–1839 | 18–38 | 55–56 |
| 3 | 1839–1839 | 2–21 | 67–68 |
Excluding the initial CPU-percent priming sample, system CPU was12.8–14.5% across32 logical CPUs. Each of4 workers was about100% of one CPU core; affinity sets0–1,2–3,4–5,8 are distinct. Observed worker RSS around1.27–1.40GB is a snapshot, not a guarantee for other geometries or objectives. Available RAM stayed at least57.525GB. Completed-epoch recorded allocator peaks were1.388–1.516GB; these are allocator peaks, not complete device-process VRAM. Raw per-process read_bytes/read_chars and system disk counters are retained; cache-backed read_chars must not be called physical disk traffic. No data-loader wait instrumentation or comparative throughput experiment was run, and no speedup is inferred.
GPU3 was hotter in this window, consistent with treating the user's reported cooling issue conservatively; the sampling does not diagnose its cause. GPU3 remains strictly one task. GPU0–2 maximum3 means a capacity ceiling including existing jobs, not a requirement to fill slots or permission to introduce unapproved research.
## Concrete scheduling boundaries for Claude's revised plan
- Existing seed1 v2, Track A recovery and already reviewed postfits retain their frozen idle rules and sources. Pending work remains owned by its existing coordinator; do not manually launch or transfer those folds.
- New entries alone use reviewed shared capacity admission: total slots, measured/appropriate peak memory plus2GiB, host RSS headroom, CPU assignments, per-job accounting and global phase3 ledger. Memory demand must be scoped to the new objective/model; the present small H2 footprint does not certify GINO or other graphs.
- Preserve one of GPU0–2 for legacy idle-only dispatch while legacy folds are pending. All three currently have old jobs, so this means excluding a designated card from **new** admissions until old work releases it, not stopping an old worker to manufacture an idle card. Track A fold3 is fixed to GPU3 and cannot be moved to that spare card without separate changes.
- New entry construction order is c,1b,b,Track B; execution can overlap once each entry has passed review and resource/dependency gates. Scientific settings remain their fixed specs; existing612000 phase3 ledger, item caps and per-job wall-GPU accounting remain explicit. Shared-mode cost projections use Claude's specified1.5 margin and total-projection stops; do not retroactively change old counters/caps.
- Validation/model decisions that compare effects with seed noise await seed1 results. Preparing independent entries need not wait for that result. Exact item contracts, admission locks and allocation/reservation handling remain to be implemented and reviewed, not claimed as running code.
Sources actually read: Claude1795006714.455999 and1795006766.279919; native user corrections; existing seed1/recovery source and receipts. Five-line MeltAstra update delivered in original Slack thread. No new scientific execution from this assessment.
+6
View File
@@ -1,3 +1,9 @@
## 最新用户纠正:研究事项并行,吞吐benchmark撤回且未启动(2026-11-18)
用户通过TRON明确要不同研究工作的并行,不是同一训练1/2/3副本吞吐或算子优化;以后意见先由Claude领导裁量。Claude1795006714/6766已实读:撤回6592的benchmark,入口按c→1b→b→Track B构建,各自审后可并行;seed1只约束效果/噪声判决,不一律阻塞开工。每项训练入口一次审阅,test前独立审阅保持。GPU0–2最多3任务、GPU3严格1,当前frozen训练和postfit不改。现有pending时保留一张0–2不给新入口,不能停现有worker制造空卡或重派旧fold。
Melt只做25.5685s六次只读采样并写过两份未完成草稿;草稿已移出tools、逐字归档withdrawn-throughput-draft,从未运行benchmark/测试/冻结/读标签,无probe root/process。原seed1 PID1476593和三worker、Track A recovery PID1650006和fold4worker1650207均核在,78code pins全匹配;两队列无STOP/complete。资源及撤回证据artifacts/gpu-concurrency-assessment-2026-11-18;说明docs/independent-research-concurrency-2026-11-18.md。新共享capacity调度仅待新入口实现/审阅,没有已执行收益或新研究结果。
## 最新恢复fold4首轮成本门PASS,fold3仍待派(2026-11-18)
原1650006/1650109/1650207队列和worker,fold4首轮57.020449043秒,单折投影14437.426020≤30960;pending fold3仍保留原30960整帽,含post7200及old47943.947195一次,累计投影100541.373214≤162000。fold4已真实进入epoch2;fold3尚未实测,不称全两折成本PASS。快照/首轮原history/门在artifacts/track-a-recovery-start-evidence-2026-11-18。无重复启动/改源/新计算范围,当前状态看live/terminal。