Legacy: foundry_wsl

foundry_wsl/ is the first tool in this repository: a toolkit for making Microsoft Foundry Local usable from WSL2, built during the evaluation described in the research notes. It is kept for reference, is not part of the installable prism-local package, and is run from a checkout. New work goes into Prism.

python -m foundry_wsl.cli --help          # doctor, proxy, status, inject

Command	What it does
doctor	Reports NVML GPU detection, whether CUDA libraries resolve, and the Foundry daemon state
status	Shows Foundry Local status information
proxy [--host H] [--port P]	A stable reverse proxy (default 127.0.0.1:5272) in front of Foundry's ephemeral daemon port, resolved from ~/.foundry/daemon.json
inject CONFIG [--device-id N]	Writes a CUDA provider_options entry into a model's genai_config.json

Known issue with inject

inject writes the provider options under a top-level session_options key. In the current model configs inspected while writing these docs, the decoder's provider list lives under model.decoder.session_options, and it is not established that ONNX Runtime GenAI honors the top-level key. Prefer Prism's --device cuda, which selects the provider through the runtime API instead of by editing files.

Windows host bridge

windows_bridge/ holds two helpers for the alternative architecture of running Foundry Local on the Windows host and reaching it from WSL2:

• setup_windows_gateway.ps1: run in an elevated PowerShell on Windows; configures the firewall and starts Foundry Local listening on all interfaces.
• test_connection.sh: run in WSL2; probes http://<windows-host>:${FOUNDRY_PORT:-5272}/v1/models.

The trade-offs are discussed in the feasibility analysis.