transceiver-db/sync/history/2026-05-09-magatama-all-lane-runpod-training-start.md
2026-05-10 01:11:21 +02:00

64 lines
1.5 KiB
Markdown

# MAGATAMA All-Lane RunPod Training Start
Date: 2026-05-09 23:09 UTC
## Scope
- Train all current MAGATAMA LLM lanes via RunPod:
- `magatamallm`
- `fo_blogllm`
- `tip_llm`
- `pulso_llm`
- `contact_llm`
## Preflight
- MAGATAMA services were online on Erik.
- Active RunPod endpoint reported by MAGATAMA: `0rmkf28w2g5gip`.
- RunPod worker kind: `custom-magatama`.
- Dataset source: URL-based lane export.
- Previous successful/adopted runs existed for:
- `magatamallm`
- `fo_blogllm`
- `tip_llm`
- No previous run existed yet for:
- `pulso_llm`
- `contact_llm`
## Runner Fix
- Fixed `scripts/trigger_lane_training_once.py` locally and on Erik.
- The script previously used stale API keys:
- `iterations`
- `seedOnly`
- The MAGATAMA training API expects:
- `iters`
- `seed_only`
- Added `all` mode to run all lanes sequentially.
- Added streamed SSE logging so progress is visible during long RunPod runs.
## Live Run
- Started on Erik:
- `python3 -u scripts/trigger_lane_training_once.py all 500 false`
- Log:
- `/opt/magatama/logs/runpod-all-lanes-20260509T230549Z.log`
- First active lane:
- `magatamallm`
- First RunPod job:
- `89627e7e-8533-45db-9fe8-eca994018aa6-e2`
- Initial `magatamallm` dataset:
- `1375 train`
- `153 eval`
- `1528 total`
## Success Rule
- Do not treat RunPod `COMPLETED` as success by itself.
- A lane is only successful when:
- the model artifact exists,
- MAGATAMA imports/adopts it locally,
- smoke checks pass,
- the active alias/version is updated.