when will AVX512_BF16 be supported?

[marsggbo]

Hi authors,

Thanks for sharing this codebase! In your paper you mention:

"we design a specialized computation kernel for expert processing on the CPU using the AVX512_BF16 instruction set, which is not supported in the native PyTorch implementation."

However, in the released repository I only see the following Python function:

def run_expert_at_cpu(self, i_layer, i_expert, inps, routing_weights):
    """Run the expert at CPU"""
    return self.model.layers[i_layer].block_sparse_moe.experts[i_expert](
        inps, routing_weights
    )

May I ask when will AVX512_BF16 be supported?