QuantResidualAdd¶

Bases: QuantBase

Quantized residual block with two parallel layer streams.

The input flows through two parallel paths (layers_a and layers_b), whose outputs are added element-wise. An optional ReLU activation and output fake-quantization are applied after the add. An empty layers_b means the skip path is the identity.

Parameters:

Name	Type	Description	Default
`layers_a`	`list[QuantBase]`	Quantized layers for stream A (main path).	required
`layers_b`	`list[QuantBase]`	Quantized layers for stream B (empty = identity skip).	required
`act_func`	`str \| None`	Either `"relu"` or `None`.	`None`
`ema_constant`	`float`	EMA smoothing factor for the output observer.	`0.01`

build_testcase_outputs ¶

build_testcase_outputs(
    x: Tensor,
) -> tuple[torch.Tensor, list[dict]]

Capture every sub-layer output plus the final combined output.