gliclass-v4-edge-q8-onnx
Weight-only int8 build of knowledgator/gliclass-v4-edge, quantized with ORT MatMulNBits (bits=8, block 128, asymmetric, fp32 scales).
GLiClass v4-edge (DeBERTa-v3, labels-first) weight-only q8: int8 MatMulNBits (block 128, asymmetric), fp32 head excluded. Gate: 0 flips, 6/6 golden agree.
Parity: worst 0.0899, 0 flips (6/6 golden agree).
Apache-2.0, as the source checkpoint.