# Reproducible method — Laguna-S-2.1-Spark-Agentic pack

Independent reproducible method.

## Pins (card freeze)

| Item | Value |
|------|--------|
| Base model | `poolside/Laguna-S-2.1` |
| Base revision | `00af5a51782109b587a3b3bbf11875e566036fa7` |
| Official GGUF repo (upstream) | `poolside/Laguna-S-2.1-GGUF` |
| Optional byte-identical mirror | `hizrianraz/Laguna-S-2.1-GGUF` |
| GGUF revision (upstream pin) | `fc4e481289523cf7d0df668da6d1d391616141ca` |
| Default weight | `laguna-s-2.1-Q4_K_M.gguf` |
| Q4_K_M sha256 | `a8b55c75714ea73fd90ec85de5defdc0b8d88ca0ad2108343cdd8fc22f7583e4` |
| Engine | `github.com/poolsideai/llama.cpp` branch `laguna` |
| Hardware | DGX Spark GB10 only |

## Steps

1. Read OpenMDW-1.1 on the base model card/LICENSE. Accept constraints before redistribute.
2. Build engine per `SPARK.md` (CUDA `121` → `121a`, `llama-server`).
3. Record `git rev-parse HEAD` of the engine tree into `results/engine_sha.txt`.
4. Download official GGUF (not a DIY re-quant unless you will measure a real delta):

```bash
# preferred: fail-closed helper (reads pack SHA256SUMS + pinned GGUF rev)
./scripts/pull_official_gguf.sh

# or manual download + GNU 2-column check from the weight directory:
hf download poolside/Laguna-S-2.1-GGUF \
  laguna-s-2.1-Q4_K_M.gguf \
  --revision fc4e481289523cf7d0df668da6d1d391616141ca \
  --local-dir ~/models/laguna-s-2.1
# Optional byte-identical mirror:
# hf download hizrianraz/Laguna-S-2.1-GGUF \
#   laguna-s-2.1-Q4_K_M.gguf \
#   --local-dir ~/models/laguna-s-2.1
cp SHA256SUMS ~/models/laguna-s-2.1/
(cd ~/models/laguna-s-2.1 && sha256sum -c SHA256SUMS)
```

Engine pin before build (measured commit, not floating branch tip):

```bash
git clone https://github.com/poolsideai/llama.cpp ~/src/llama.cpp-laguna
cd ~/src/llama.cpp-laguna
git checkout 04b2b72cb54048ead292884adbe11f284e3ec950
# optional host patch: docs/BUILD_SPARK.md (isfinite / GNU 13.3)
```

5. Serve with last-green flags (match `results/LAST_GREEN_PIN.md` / `SPARK.md`):

```bash
./build/bin/llama-server \
  -m ~/models/laguna-s-2.1/laguna-s-2.1-Q4_K_M.gguf \
  --host 127.0.0.1 --port 8000 \
  --ctx-size 8192 -ngl -1 -fa on --jinja \
  --alias local-laguna --metrics
```

Confirm `curl -sf http://127.0.0.1:8000/v1/models` → 200.  
Smoke / bench **`--model` must equal `--alias`** → `local-laguna`.

6. `llama-bench` @ prompt 2048 and 8192; save stdout to `results/llama_bench_q4km.txt`.
7. `python scripts/bench_server.py --ctx-mark 2k --ctx-mark 8k`.
8. `python eval/agent_smoke/run_smoke.py --model local-laguna --out results/agent_smoke.json`.
8b. Optional Hermes-class: `python eval/hermes_agent_smoke/run_hermes_smoke.py --model local-laguna --out results/hermes_agent_smoke.json`.
9. Populate the tables in `SPARK.md` from those files only — never invent.
10. Publish pack (docs + results + scripts). **Host GGUF binary only if you have a measured unique artifact**; otherwise ship digests + download commands (S5).
11. Optional DFlash is **not** default. If tried on Spark, record with honesty — 2026-07-29 measure is **DO_NOT_PROMOTE** (`results/dflash_2026-07-29/`). Note: that trial used `-ngl 99` while baseline pin is `-ngl -1`; reject still holds (gen + prefill both slower).

## HF repository surface

- User: `hizrianraz` (personal)
- Repo id: `hizrianraz/Laguna-S-2.1-Spark-Agentic`
- Card framing: independent measurements by the repository owner; no org product claims

## DIY quant policy

Only if:

1. Official/Unsloth GGUF cannot load or is >10% worse on agent_smoke pass rate or tg tok/s at same ctx, **and**
2. You publish method + imatrix/source commit + full sha256 + smoke numbers

Otherwise: official GGUF + Spark delta docs only.