Why force_fp32 is set to true for conv weight and biases on XNNPack backend

[onurtore]

Is there a specific reason force_fp32 set to true for convolution weight and biases for the XNNPack backend? I would like to use fp16 operations but default does not allow me.

    self.define_tensor(
        kernel_node,
        xnn_graph,
        vals_to_ids,
        ...,
        force_fp32=True,
    )
    ...
    self.define_tensor(
        get_input_node(node, 2),
        xnn_graph,
        vals_to_ids,
        ...,
        quant_params=bias_quant_params,
        force_fp32=True,
    )

[Gasoonjia]

@GregoryComer can you take a look?

[GregoryComer]

I'll take a look. There was some historical reason for this but I forget why. I expect we can relax it.

[onurtore]

@GregoryComer , I would like to mention that I changed force_fp32 to false, but the model did not work. Unfortunately I did not get the chance to inspect it, but if you think my help would be benefical, I am happy to inspect.

Here is the error: