SeaBiscuit Documentation
Reference for installing, operating, and tuning SeaBiscuit on‑prem.
System requirements (validated tiers)
The table lists the lowest configurations we have tested end‑to‑end. Hardware below the Baseline tier may still work for smaller or quantised models, but it is not yet covered by support SLAs.
| Tier | Use case | CPU | RAM | GPU / VRAM | NIC |
|---|---|---|---|---|---|
| Baseline | Models ≤ 13 B params, 4‑bit or FP16 | 4 cores (AVX2) | 16 GB | Optional; any PCIe GPU | 10 GbE |
| Large‑model | 30 B–70 B param splits | 8 cores (AVX‑512) | 32 GB | ≥1 GPU with ≥16 GB VRAM (A100/H100, MI‑300, etc.) | 25 GbE RDMA or HDR InfiniBand |
Both tiers assume at least 50 GB of free NVMe for model files and logs.
Network guidelines
SeaBiscuit adapts its partitioning to the available fabric. To split one model across hosts you need sustained bandwidth ≥ 3 GB/s (25 GbE) or InfiniBand FDR/HDR.
| Fabric | Minimum | Preferred |
|---|---|---|
| Ethernet | 25 GbE RDMA/ROCE | 100 GbE |
| InfiniBand | FDR 56 Gb/s | HDR 200 Gb/s |
10 GbE networks are supported but the planner keeps tensors >1 MB on the same host.
Placement config
# /etc/seabiscuit/policy.yaml
prefer_same_rack: true
max_block_memory_gb: 16
cpu_block_allow:
- layer_norm
- softmax
Monitoring
Metrics at /metrics (Prometheus). Grafana dashboard JSON in dashboards/.
Glossary
| Term | Definition |
|---|---|
| AVX2 / AVX‑512 | CPU SIMD extensions; speed up FP16 math on CPUs. |
| GbE | Ethernet rated in gigabits per second. |
| FDR / HDR | InfiniBand generations at 56 Gb/s and 200 Gb/s. |
| Min‑cut | Graph partition algorithm SeaBiscuit uses to split models. |
| ROCE | RDMA over Converged Ethernet. |
FAQ
Can I run on a single laptop?
Yes for small (≤7 B) models, but performance will be limited and not covered by support.
Does SeaBiscuit support training?
Not yet – inference only.