| # UP2 low-cap identity reconciliation |
|
|
| Coordinator ledger (binding brief 2026-10-24): sparkhead `~/dsv41-deploy` |
| main `21bed31`, tag `prod-k1-327-20261023`. No Spark access by this lane. |
| MEASURED CPU source inspection: `git show 21bed31:{dsv41.env,serve.sh,start.sh}`. |
| MEASURED x86host **file reads only** over SSH: `current` resolves to |
| `/home/operator/dsv41-offload/prod/dspark-worker/fb68c52`; SOURCE_COMMIT is |
| `fb68c5291792ab90c0f468bd94010917967799f3`. The installed TP2 unit reads |
| that tree, uses the original checkpoint, max length1048576 and the production |
| NVLink pair. Status file reports TP2 and no communication errors at the read. |
| This is file evidence, not proof of process bytes or native UP2 qualification. |
| |
| `UP2-low-k1.env` derives from the composed UP2 wrapper; it does **not** change |
| the worker/checkpoint identity. Original identity be2c1436… is still production's |
| identity, not an obsolete worker revision. Worker *source revision* fb68c52 and |
| HELLO *checkpoint identity* are different fields. Inherited noncanonical K1 |
| values are refused. No config/worker overwrite is permitted on a mismatch. |
| |
| | Setting, equal across every arm | Pinned value / verification | |
| |---|---| |
| | Production rollback | main21bed31, tag prod-k1-327-20261023; save actual argv/env/hashes before opening | |
| | Target/checkpoint | /srv/dsv41/legal-v2a; candidate manifest, quant plan, config/tokenizer hashes identical both ranks and saved production | |
| | Base/quant revision | 1bae0e1088993b9b82de08f59de5be5a34e42170 / 86f852533ec94d1a537fea41fd9c01d6b0f6eb50 | |
| | Image/engine/kernels | v5; exact running image SHA and engine/JIT/code hashes from saved production; tag alone insufficient | |
| | Target TP / fabric | 2; enp1s0f1np1 / rocep1s0f1; NCCL both f1 functions; EXL3 ring rocep1s0f1 | |
| | Context / KV bytes / pool | 327680 / 6442450944 per rank / 2502600 tokens | |
| | K1 | pair3, drop-MTP1, transient1, alias1, indexer-factor4; all13 K1 bytes match manifest | |
| | Admission / batching | 4 sequences, MAX_BATCH4096, threshold1024, priority scheduler, throttle min5 / every24 | |
| | Speculation | dspark5; K5 at B1–2, K4 at B3–4; adaptive margin1/min2/hi1.6/maxreq4; no K5 override | |
| | Remote worker | fb68c5291792ab90c0f468bd94010917967799f3; TP2, maximum1048576, checkpoint/adapters unchanged | |
| | Worker GPUs (identity only) | GPU-c4e7a371-d8a8-ebf0-918c-da3e38c87a93 / GPU-7744b44b-8a13-b06d-a67e-f958fd74b8fc; never changed by runner | |
| | HELLO identities | be2c1436fa9f9e66b398e1b1da2e140e6fddc82a11939e2ac4240a0ca79543bd / Markov0b6e71e638abd5abce2db433965bb5ff523cbbe92ed7ae54324a50ccaeffa19a | |
| | Worker transport/cost | o1rdma://10.202.28.1:18620; bind10.202.28.2; timeout50; round28.79,31.86,34.44,37.76,41.56,43.68; draft4.4ms | |
| | Proxy | capacity, async1/process, spin50us/nap20us, backoff5000ms, respawns3/reset3600s; no local drafter, no proxy torch | |
| | Engram | original O1 shim/JIT; both RDMA links, timeout100/retry5000/log60; REQUIRED1 in **all experiment arms** | |
| | Engram production difference | production REQUIRED0 allows fallback; recipe requires REQUIRED1 to exclude fallback from replay ratios; verify remote active both ranks | |
| | Other experiments | split0, VOR0, dual-ring0, CPU-pin0, SO1/SO2 absent; no connector/PF1/capture overlays | |
| | Clocks / thermal / loading | existing2150MHz cap, LAZY; same worker clocks; record drift, never adjust worker clocks/units | |
| | Timing requests | exact150000/300000 complete chat; output1500, thinking true, high/75, T1/top_p1/seed42/priority10; same corpus/order/header | |
| | Cache / repetitions | cold cached_tokens0, 3 per size per arm; one fixed fresh-window salt prefix; point/rep prompt and salt SHA equal | |
| | Allowed differences | replay OFF/ON/OFF; audit1 on both diagnostic starts, audit0 on all measured starts; label/run/cache reset/start time | |
|
|
| UP2's additional composed engine bytes are identical across OFF/ON. Runtime |
| `verify_runtime --emit` checks all42 bytes on **both ranks**. Any source/image, |
| worker, target, K1 or policy mismatch stops; it never authorizes worker staging. |
| All live gates remain NOT RUN by this lane. Required-native missing apparatus |
| must be reported NOT RUN/HOLD. The newer lane brief sets the resource floor to |
| 3GiB; the older recipe's5GiB is superseded only for this low-cap runner. |
|
|
