feat: bootstrap jetson-ai-ollama from agent-swarm template
Some checks failed
gates / quality (push) Has been cancelled
Some checks failed
gates / quality (push) Has been cancelled
Leader-Bootstrap (Planung) aus Template + Jetson-Meta-Prompt:
- ROADMAP-Board W1-W6 als Abhaengigkeits-DAG (Ready-Set = {W1})
- STATE mit disjunkter Datei-Ownership W2-W6
- CONVENTIONS Projekt-Vertrag (Profile, Pfade, Ports, Daemon-Env, num_ctx, DVC, Scope)
- manifests/requirements.md = kalt verifizierbares Review-Manifest (R-01..R-10)
- REVIEW.md Kalt-Review-Einstieg; ADR 0002 (Ollama-nativ, kein Swapper/Gateway)
- session-prompts/W1..W6 dispatch-fertig; .gitignore um Jetson-Ignores erweitert
- examples/walkthrough entfernt (Projekt hat eigenes gefuelltes Board)
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
This commit is contained in:
44
session-prompts/W2.md
Normal file
44
session-prompts/W2.md
Normal file
@@ -0,0 +1,44 @@
|
||||
# Session-Prompt — W2 Ollama-Profile (parallel, nach W1)
|
||||
|
||||
```
|
||||
Modell: Sonnet
|
||||
|
||||
Du bist WORKER für Paket W2 in /pfad/zu/jetson-ai-ollama, Branch `feat/w2-profiles`
|
||||
(von `main`, nach W1-Merge). Lies AGENTS.md → STATE.md → CONVENTIONS.md (Projekt-
|
||||
Vertrag = Single Source of Truth für Profilnamen, Pfade, Env, num_ctx).
|
||||
|
||||
## Aufgabe
|
||||
Die drei Ollama-Profile reproduzierbar bauen lassen — über Modelfiles, systemd-
|
||||
Daemon-Env und ein Build-Skript, mit DVC als Modell-Source-of-Truth.
|
||||
|
||||
## Deliverables (deine Dateien — Ownership: `ollama/`, `scripts/build-models.sh`, `.dvc/`)
|
||||
- `ollama/Modelfile.light` — `FROM /opt/jetson-ai/models/qwen3.5-4b-q4_k_m.gguf`,
|
||||
`PARAMETER num_ctx 16384`, `PARAMETER temperature 0.7`.
|
||||
- `ollama/Modelfile.gemma` — Gemma-4-E4B GGUF, `num_ctx 16384`.
|
||||
- `ollama/Modelfile.heavy` — Qwen3.5-9B GGUF, `num_ctx 8192`.
|
||||
- `ollama/systemd/ollama.service.d/override.conf` — exakt die Env aus CONVENTIONS
|
||||
(FLASH_ATTENTION=1, KV_CACHE_TYPE=q8_0, MAX_LOADED_MODELS=1, KEEP_ALIVE=30s,
|
||||
MODELS=/opt/ollama/models, HOST=0.0.0.0:11434).
|
||||
- `scripts/build-models.sh` — `dvc pull` → `ollama create qwen-light -f …light`,
|
||||
`ollama create gemma -f …gemma`, `ollama create qwen-heavy -f …heavy`. Idempotent.
|
||||
- `.dvc/config` finalisieren (Remote bleibt env-/platzhaltergesteuert).
|
||||
|
||||
## Akzeptanzkriterien
|
||||
- [ ] `pytest tests/test_modelfiles.py` grün (W5 liefert den Test; bis dahin lokal:
|
||||
`grep -q 'num_ctx 16384' ollama/Modelfile.light`). (R-02, R-04)
|
||||
- [ ] `grep -Eq 'q8_0' ollama/systemd/ollama.service.d/override.conf` und
|
||||
`grep -q 'MAX_LOADED_MODELS=1'` (R-03).
|
||||
- [ ] `bash -n scripts/build-models.sh` ok; baut alle drei Profile (R-08).
|
||||
- [ ] num_ctx in JEDEM Modelfile explizit gesetzt.
|
||||
|
||||
## Nicht anfassen
|
||||
`exporter/`, `client/`, `tests/`, `scripts/setup-jetson.sh`, `scripts/benchmark.sh`,
|
||||
`.github/`, `docs/`. Nur deine Ownership-Pfade.
|
||||
|
||||
## Abschluss
|
||||
1. `manifests/requirements.md`: R-02, R-03, R-04, R-08 → `done`.
|
||||
2. `STATE.md` + Session-Log aktualisieren.
|
||||
3. Commit(s) `feat(w2): …` + Trailer `Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>`.
|
||||
4. `bash scripts/gate-session-hygiene.sh && bash scripts/gate-ownership.sh`.
|
||||
5. PR (Base `main`). Dann Review-Gate (Opus).
|
||||
```
|
||||
Reference in New Issue
Block a user