vllm - 💡(How to fix) Fix [CPU Backend] Refactor CPU FusedMoE to MK flow [1 participants]

import mk_flow

# Define the FusedMoE model
class FusedMoE(mk_flow.Model):
    def __init__(self, num_experts, num_layers):
        super(FusedMoE, self).__init__()
        self.experts = mk_flow.ModuleList([mk_flow.Linear(128, 128) for _ in range(num_experts)])
        self.layers = mk_flow.ModuleList([mk_flow.Linear(128, 128) for _ in range(num_layers)])

    def forward(self, x):
        # Use MK flow's parallelize function to parallelize the experts
        expert_outputs = mk_flow.parallelize(self.experts, x)
        # Use MK flow's parallelize function to parallelize the layers
        layer_outputs = mk_flow.parallelize(self.layers, expert_outputs)
        return layer_outputs

# Initialize the FusedMoE model
model = FusedMoE(num_experts=8, num_layers=4)