This private-fork branch preserves the DFlash2 implementation from llama.cpp pull request 27342 and adds reproducible scripts and measurements for Ornith 1.5 35B A3B Abliterated.
- Full experiment report
- CUDA build script
- One-GPU server launcher
- Benchmark script
- Public proxy check
- Raw benchmark data
- Static script validation
- Standalone runtime acceptance test
The tested DFlash2 source base is commit 2f3923bc81346046aa5765dda15fb28497f49ac6.