Check group size is divisible by 32

Summary: Currently when checking the per_channel_group quantization parameters we don't check that the group_size must be a multiple of 32. This constraint was added after we implemented the original checks here. Let's add multiple of 32 here.

Reviewed By: digantdesai

Differential Revision: D66131456

Author: mcr229

backends/xnnpack/operators/node_visitor.py

@@ -318,6 +318,9 @@ def _check_per_channel_group_params(
         ), "For per_channel_group quant, num_groups and group_size must be > 0, but got num_groups: {num_groups}, group_size: {quant_params.group_size}"
         output_channels = dims[quant_params.axis]
         input_channels = dims[quant_params.axis ^ 1]
+        assert (
+            quant_params.group_size % 32 == 0
+        ), "Delegation to XNNPACK requires group_size to be a multiple of 32, but got {quant_params.group_size}"
         assert (
             output_channels == cast(torch.Tensor, quant_params.scale).shape[0]
         ), "For per_channel_group quant, expecting output channels to match scale.shape[0], gut got: {output_channels}, scale.shape[0]: {quant_params.scale.shape[0]}"