Skip to content

Add Windows CUDA worker support#44

Merged
evilsocket merged 1 commit intoevilsocket:mainfrom
icryo:windows-cuda-support
Mar 5, 2026
Merged

Add Windows CUDA worker support#44
evilsocket merged 1 commit intoevilsocket:mainfrom
icryo:windows-cuda-support

Conversation

@icryo
Copy link
Contributor

@icryo icryo commented Mar 5, 2026

Summary

  • Add Windows broadcast address detection via ipconfig parsing in discovery
  • Add Windows system memory detection via PowerShell Get-CimInstance
  • Add CUDA_PATH env var fallback for nvcc on Windows
  • Replace hardcoded /tmp with std::env::temp_dir() in cache path fallbacks
  • Add Windows CUDA section to install docs

All changes are behind #[cfg(target_os = "windows")] — zero impact on Linux/macOS builds.

Tested

Windows RTX 5090 (CUDA 13.1) ↔ Linux RTX 4090 (CUDA 12.4) over 10GbE:

  • Windows as worker, Linux as master — discovery + inference working
  • Windows as master, Linux as worker — discovery + inference working
  • Zero-config model push at ~175 MiB/s
  • 123 tok/s on Qwen3.5-0.8B split across both machines

Files changed

File Change
cake-core/src/cake/discovery.rs Windows broadcast, memory, nvcc detection
cake-core/src/utils/models.rs /tmptemp_dir()
cake-cli/src/main.rs /tmptemp_dir()
docs/install.md Windows CUDA build instructions

- Add Windows broadcast address detection via ipconfig (discovery.rs)
- Add Windows system memory detection via PowerShell/CIM (discovery.rs)
- Add CUDA_PATH fallback for nvcc on Windows (discovery.rs)
- Replace hardcoded /tmp with std::env::temp_dir() (models.rs, main.rs)
- Add Windows CUDA build instructions to install docs

Tested: Windows RTX 5090 worker ↔ Linux RTX 4090 master, both directions,
zero-config discovery + inference working over 10GbE.
@evilsocket
Copy link
Owner

AMAZING! Thanks!

@evilsocket evilsocket merged commit bff9b93 into evilsocket:main Mar 5, 2026
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

2 participants