Class Linear

ops.py:231–244

    Dequantize weights on the fly before doing the compute
    """
    class Linear(GGMLLayer, comfy.ops.manual_cast.Linear):
        def __init__(self, in_features, out_features, bias=True, device=None, dtype=None):
            torch.nn.Module.__init__(self)
            # TODO: better workaround for reserved memory spike on windows
            # Issue is with `torch.empty` still reserving the full memory for the layer
            # Windows doesn't over-commit memory so without this 24GB+ of pagefile is used
            self.in_features = in_features
            self.out_features = out_features
            self.weight = None
            self.bias = None

        def forward_ggml_cast_weights(self, input):
            weight, bias = self.cast_bias_weight(input)
            return torch.nn.functional.linear(input, weight, bias)

    class Conv2d(GGMLLayer, comfy.ops.manual_cast.Conv2d):
        def forward_ggml_cast_weights(self, input):