Lesson:pivot-b200-20260717-portability-preflight-a01
| 제목 | B200 target-generation portability boundary |
|---|---|
| 궁금했던 점 | Which pre-final artifacts can be completed on B200-2 and safely migrated to the L40S/RTX PRO 5000 final experiment? |
| 해본 것 | At git 8f61aa84212dc232a1392180fdbe4c0f8333d8fb, audited final_v1 protocol, run-plan inventory, generation/latency/trace validators, then ran a read-only B200-2 preflight with ssh -p 49001: GPU inventory/process query, git state, and .venv torch/CUDA/vLLM version checks. |
| 당시 조건 | Repo dw-kv; protocol weighted_slo_final_v1 sha256=528c4a83e8dac516ab6898f9b73b652c348c14d9eb5c8a2eea1e965e50949c10; host B200-2; 8x NVIDIA B200 all 0 MiB/0% and no compute processes; runtime torch 2.11.0+cu130, CUDA 13.0, vLLM 0.21.0; pinned Qwen and Mistral snapshots available; global concurrency remains 1. |
| 실제 결과 | The inventory declares target_generation::primary and target_generation::architecture_replication with hardware=null. Promotable generation validates pinned model/tokenizer, clean git, native extension hash, backend/runtime versions, and records the producing GPU, but does not bind the artifact to L40S or RTX. Latency, regime selection, native baseline, traces, policy bindings, host preflight, and results are target hardware/config bound. B200 final cells are rejected. |
| 왜 그랬는지 | migration-only |
| 다음에 기억할 것 | Complete only the two deterministic target-model generation calibrations on one isolated B200 GPU at a time. Preserve them as structurally promotable, content-addressed migration candidates; never treat B200 latency, throughput, load, policy, or comparative outcomes as final evidence. |
| 언제 맞는지 | Applies to frozen final_v1 model IDs/revisions, deterministic temperature=0/top_p=1/seed=0, BF16, TP1, max_model_len=4096, output cap=256, disjoint 1,536-prompt calibration split, clean source commit, and current hardware-null target_generation inventory contract. |
| 신뢰도 | 높음 |
| 관련 자료 | research/EXPERIMENT_PROTOCOL_V1.md; research/CALIBRATION_PIPELINE_V1.md; scripts/pivot_calibration/generation.py; scripts/pivot_experiment/run_plan.py; scripts/pivot_experiment/cell_worker.py; B200-2 read-only preflight on 2026-07-17 |
| 자료 출처 | 우리 기록 |
| 작성자 | S3ResearchAgent |
| 처음 작성한 시각 (UTC) | 2026-07-17T05:39:32.530793Z |
| 마지막 수정 시각 (UTC) | 2026-07-17T05:39:55.936872Z |
근거 protocol-and-calibration-contract-20260717: experiments/protocols/final_v1.json sha256=528c4a83e8dac516ab6898f9b73b652c348c14d9eb5c8a2eea1e965e50949c10; scripts/pivot_calibration/generation.py sha256=edbb06672e00f08c5501410168739b8997109600e98c6ac387b29ac528d44ff8; scripts/pivot_experiment/run_plan.py sha256=4331d3de24ca909048a414cb51b98f512351406c65626d54ad419bd4f2b9e0f4; scripts/pivot_experiment/cell_worker.py sha256=4cfeaf7ce6569763d900c9760a26ca54c7cdbba8207879f53147465ba8c14867
코드 · 확인 범위: 기록 안 됨 · S3ResearchAgent · 2026-07-17T05:39:55.936872Z
Machine-readable evidence that target-generation inventory slots are hardware-null while target-bound calibration and final cells enforce hardware identity.