Operators¶
Use this hub when you install OddsFox Graph, keep a local model available, and run the inference pipeline.
Operator
Install¶
uv sync --frozen --extra dev
On Apple Silicon, install llama-cpp-python with Metal support:
CMAKE_ARGS="-DGGML_METAL=on" uv sync --frozen --extra dev
On Linux or any CPU-only machine, see
Linux / CPU-only setup to avoid compiling
llama-cpp-python from source.
Download the recommended GGUF model (see models/README.md):
huggingface-cli download Qwen/Qwen3-4B-GGUF Qwen3-4B-Q4_K_M.gguf \
--local-dir models \
--local-dir-use-symlinks False
mv models/Qwen3-4B-Q4_K_M.gguf models/qwen3-4b-q4_k_m.gguf
Place source data¶
Put a Pipeline golden mart export at the repository root or under data/:
polymarket_wc2026_market_hourly_odds_<timestamp>.parquetpolymarket_wc2026_market_hourly_odds_<timestamp>.schema.json
Run end to end¶
oddsgraph run
Prefer the server backend when you want concurrent residual inference on a large full-dataset run:
llama-server -m models/qwen3-4b-q4_k_m.gguf -ngl -1 -c 12288 -np 4 -cb -fa on \
--host 127.0.0.1 --port 8080
oddsgraph run --llm-backend server --concurrency 4
Backend choices¶
Prefer inprocess (default) or mlx for single-machine decode; use server
for concurrent pipelining. Full comparison, outlines notes, and MLX setup:
Inference backends.
Next pages¶
| Goal | Page |
|---|---|
| First successful run | Quickstart |
| Stage-by-stage CLI | Running the pipeline |
| Backend choice / MLX | Inference backends |
| Faster infer (server) | llama-server |
| Settings defaults | Configuration |
| Deterministic coverage | Deterministic topology |
| Official bracket | Official bracket |
| Common failures | Troubleshooting |
| Common questions | FAQ |