mirror of
https://github.com/pewdiepie-archdaemon/odysseus.git
synced 2026-08-09 15:08:44 -04:00
Alexandre Teixeira
86 lines
3.0 KiB
Markdown
86 lines
3.0 KiB
Markdown
# Cookbook
|
|
|
|
Features in this document are generated from [`../feature-catalog.json`](../feature-catalog.json), the canonical inventory.
|
|
|
|
## COOKBOOK-001 — Local Model Download & Recipe Lifecycle Management
|
|
|
|
- **Domain**: `cookbook`
|
|
- **Status**: `verified`
|
|
- **Evidence Maturity**: `E0`
|
|
- **Commit Verified**: `d8a2059df8e53bc7275c45339849d14c8651e73c`
|
|
- **Runtime Validation**: not-required — No separate environment-dependent runtime validation was identified during this documentation pass.
|
|
|
|
### Purpose
|
|
|
|
Downloads HuggingFace models, configures execution parameters, and manages local GGUF/MLX model servers.
|
|
|
|
### Evidence summary
|
|
|
|
- `routes/cookbook_routes.py` — `setup_cookbook_routes` — Exposes model downloading and process serving endpoints.
|
|
- `static/js/cookbook.js` — `initCookbook` — UI manager for local model library.
|
|
|
|
### Unknowns
|
|
|
|
- Disk space exhaustion during multi-gigabyte GGUF weights downloads.
|
|
|
|
## COOKBOOK-002 — Hardware Model Fitting ('What Fits?') Analysis Engine
|
|
|
|
- **Domain**: `cookbook`
|
|
- **Status**: `verified`
|
|
- **Evidence Maturity**: `E0`
|
|
- **Commit Verified**: `d8a2059df8e53bc7275c45339849d14c8651e73c`
|
|
- **Runtime Validation**: not-required — No separate environment-dependent runtime validation was identified during this documentation pass.
|
|
|
|
### Purpose
|
|
|
|
Calculates RAM/VRAM requirements, quantized size, and context overhead to determine model compatibility.
|
|
|
|
### Evidence summary
|
|
|
|
- `routes/hwfit_routes.py` — `setup_hwfit_routes` — Calculates hardware model compatibility.
|
|
- `services/hwfit/fit.py` — `calculate_fit` — Performs parameter and memory fit calculations.
|
|
|
|
### Unknowns
|
|
|
|
- Inaccurate VRAM estimation for non-standard KV-cache quantization.
|
|
|
|
## COOKBOOK-003 — HuggingFace & MLX Model Discovery Services
|
|
|
|
- **Domain**: `cookbook`
|
|
- **Status**: `verified`
|
|
- **Evidence Maturity**: `E0`
|
|
- **Commit Verified**: `d8a2059df8e53bc7275c45339849d14c8651e73c`
|
|
- **Runtime Validation**: not-required — No separate environment-dependent runtime validation was identified during this documentation pass.
|
|
|
|
### Purpose
|
|
|
|
Searches HuggingFace Hub and local MLX model repositories for compatible GGUF and MLX weights.
|
|
|
|
### Evidence summary
|
|
|
|
- `services/hwfit/hf_discovery.py` — `search_hf_models` — Queries HuggingFace API for model tags and files.
|
|
|
|
### Unknowns
|
|
|
|
- HuggingFace API rate limits when searching without an API token.
|
|
|
|
## COOKBOOK-004 — Host Docker Access for Inference Container Runtimes
|
|
|
|
- **Domain**: `cookbook`
|
|
- **Status**: `verified`
|
|
- **Evidence Maturity**: `E0`
|
|
- **Commit Verified**: `d8a2059df8e53bc7275c45339849d14c8651e73c`
|
|
- **Runtime Validation**: pending — Requires Docker access and supported physical GPU hardware.
|
|
|
|
### Purpose
|
|
|
|
Detects and connects to host Docker engine to launch containerized Ollama, vLLM, or SGLang runtimes.
|
|
|
|
### Evidence summary
|
|
|
|
- `src/host_docker_access.py` — `HostDockerAccess` — Interacts with host docker daemon.
|
|
|
|
### Unknowns
|
|
|
|
- Permission denied accessing docker socket on non-root setups.
|