Packaged GGUF benchmarks run again with the corrected llama-bench sidecar and current b9833 result parsing.
Backend.AI GO v1.10.2
4 commits since v1.10.0. (1,383 lines added, 160 lines deleted)
New Features
None
Improvements
None
Bug Fixes
- Fixed full, quick, and custom GGUF benchmarks failing with invalid argument: -o because the package contained a duplicate llama-server binary at the llama-bench path .
- Accepted the signed -1 all-layer sentinel and current backend and flash-attention fields emitted by llama.cpp b9833, preserving them through shared DTO and SQLite boundaries .
- Escaped benchmark CSV cells and neutralized spreadsheet formulas from untrusted model filenames and backend labels .
CI/CD Improvements
- Added package identity, safe archive, checksum, immutable provenance, and native llama-bench -o json --help validation .
- Pinned application packages to baiengine 26.7.22 .
Technical Details
- Preserved the v1.10 application source and continuum-router v1.12.0 while replacing only the defective llama.cpp engine package and benchmark compatibility boundary.
- Recorded the source workflow run and upstream llama.cpp revision in the engine release process.
Dependencies
- llama.cpp baiengine 26.7.22, upstream b9833.
Breaking Changes
None
Known Issues
None