Project a unit refusal direction out of the residual-stream writes
(attention, linear-attention and MLP outputs, and token embeddings) at
runtime, enabled by VLLM_ABLITERATION pointing to a JSON or safetensors
file with a refusal_direction vector. A bare name resolves to
VLLM_ABLITERATION_DIR, default /opt/vllm/abliteration. For one direction
this equals the weight edit W' = W - r(r^T W), so a base checkpoint
reproduces its abliterated release with no extra weights. Unset, the
path loads nothing and adds no work.
Co-Authored-By: Hinata <hinata@idihore.id>
Signed-off-by: MOVZX <movzx@yahoo.com>
MOVZXandHinata
8151c7b158