Pull requests / #15
#15 generate: residual control-vector projection from GGUFs
closed · @j-luwierski · 0 comments · View on GitHub
Multi-GPUModels & quantsSecurity
Description
## What Adds residual control-vector projection to the engine: per-layer control vectors loaded from GGUF files are projected onto the residual stream during generation, so model behaviour can be steered at inference time without modifying the weights. ## New options - `--control-vector-scaled` - enable the projection with a given strength - `--control-vector-layer-range` - apply the vector only to a range of layers - `--cvec-mode project` - projection mode - `--cvec-dir` - directory with the per-layer GGUFs ## Ready-to-use control vector A refusal-projection control vector for Qwen3.8-Flash-Next can be downloaded from: https://huggingface.co/Cudecnik/Qwen3.8-Flash-Next-refusal-projection
Related on strata.com
Editorial links to help you install, pick models, or read release notes — not part of the upstream thread.