vllm - 💡(How to fix) Fix [Bug]: In_proj_ba of GDN in Qwen3Next use MergeColumnParallelLinear may cause accuracy decrease? [1 participants]

# Assuming in_proj_ba is a MergeColumnParallelLinear module
in_proj_ba = MergeColumnParallelLinear(...)

# Load weights for in_proj_ba
def load_weights(state_dict, in_proj_ba):
    # Get the weight and bias for the merged column
    weight = state_dict['in_proj_ba.weight']
    bias = state_dict['in_proj_ba.bias']
    
    # Split the weight and bias into separate columns
    num_columns = in_proj_ba.num_columns
    weight_list = torch.split(weight, num_columns, dim=0)
    bias_list = torch.split(bias, num_columns, dim=0)
    
    # Load the weights and biases into the in_proj_ba module
    for i, (weight_col, bias_col) in enumerate(zip(weight_list, bias_list)):
        in_proj_ba.columns[i].weight.data.copy_(weight_col)
        in_proj_ba.columns[i].bias.data.copy_(bias_col)

# Call the load_weights function
load_weights(state_dict, in_proj_ba)