ReadonlybiasADDITIVE bias term (.bias, mlx nn.QuantizedLinear's optional bias —
qwen2 qkv, starcoder2, …). Distinct from biases, the quantization
zero-points. Applied after the matmul, before any LoRA residual,
exactly like mlx's QuantizedLinear.call.
Readonlydelta(in_features, out_features) from the quantized tensors (reference _infer_linear_shape: scales are [out, in/group_size]).
Atomically install a quantized payload and return the previous one. This is the sole mutation capability used by sensitivity sweeps.
Staticload
ADDITIVE bias term (
.bias, mlx nn.QuantizedLinear's optional bias — qwen2 qkv, starcoder2, …). Distinct frombiases, the quantization zero-points. Applied after the matmul, before any LoRA residual, exactly like mlx's QuantizedLinear.call.