Repository navigation
Affon 0.3.0
Affon 0.3.0
Downloads
- Linux x64 archive: https://github.com/neucine/affon/releases/download/v0.3.0/affon-v0.3.0-linux-x64.tar.gz
- Linux x64 standalone binary: https://github.com/neucine/affon/releases/download/v0.3.0/affon-v0.3.0-linux-x64
- Archive SHA256:
a7b18d59c9f7afc03a05dbabdc8b0ecdf9dd7985b4f31336d11737242f86f6dd - Binary SHA256:
86d96c95cbca8dccaf956571026ee729409c72b376abb045cb3897df53971523 - macOS artifacts will be uploaded separately.
Highlights
- Added the Linux CUDA backend with runtime-selected NVIDIA devices, NVRTC kernels, and cuBLAS matmul.
- Added CUDA tensor, autograd, optimizer, graph-execution, loss, indexing, and decoder-LM training coverage.
- Added CUDA memory telemetry and device-neutral decoder training metrics.
- Separated compiled training from eager validation so validation observes current parameters.
- Added platform-aware Linux CUDA and macOS Metal release tests.
Changelog
- Prepare v0.3.0 CUDA release (707845a)
- Separate compiled training from validation forward (4da05f7)
- Expand Metal matmul benchmark coverage (8d5c2af)
- Add Metal matmul benchmark harness (09e67f2)
- Add CUDA training overrides and neutral pool metrics (7c8ce39)
- Integrate CUDA training, graph execution, and decoder support (9fc68cf)
- Attribute embedded Hao runtime memory (052e890)
- Check compute memory regions at shutdown (2dedacc)
- Refine decoder LM training workflow (937bbc6)
See the repository CHANGELOG.md for the full release notes.