Maintaining Weight Equivalence

[kimjoohyungsd]

In above code below

def smooth_and_quant_temporary(model, args, isllama):
    if args.let:
        with torch.no_grad():
            for name, module in model.named_parameters():
                if "smooth_scale" in name:
                    module.data = truncate_number(module)
        if isllama:
            smooth_ln_fcs_temporary(model.input_layernorm,[model.self_attn.q_proj, model.self_attn.k_proj, model.self_attn.v_proj],
                                    model.qkv_smooth_scale,model.qkv_smooth_shift)
            smooth_ln_fcs_temporary(model.post_attention_layernorm,[model.mlp.up_proj,model.mlp.gate_proj],
                                    model.fc1_smooth_scale,model.fc1_smooth_shift)
            smooth_fc_fc_temporary(model.self_attn.v_proj,model.self_attn.o_proj,
                                model.out_smooth_scale, model.out_smooth_shift)
            smooth_q_k_temporary(model.self_attn.q_proj, model.self_attn.k_proj,
                                model.qkt_smooth_scale)
            model.mlp.down_proj.temp_weight = model.mlp.down_proj.weight
        else:
            smooth_ln_fcs_temporary(model.self_attn_layer_norm,[model.self_attn.q_proj, model.self_attn.k_proj, model.self_attn.v_proj],
                                    model.qkv_smooth_scale,model.qkv_smooth_shift)
            smooth_ln_fcs_temporary(model.final_layer_norm,[model.fc1],
                                    model.fc1_smooth_scale,model.fc1_smooth_shift)
            smooth_ln_fcs_temporary(model.self_attn.v_proj,model.self_attn.out_proj,
                                model.out_smooth_scale, model.out_smooth_shift)
            smooth_q_k_temporary(model.self_attn.q_proj, model.self_attn.k_proj,
                                model.qkt_smooth_scale)
            model.fc2.temp_weight = model.fc2.weight
    else:
        for name, module in model.named_modules():
            if isinstance(module, QuantLinear):
                module.temp_weight = module.weight
    # quant
    for name, module in model.named_modules():
        if isinstance(module, QuantLinear):
            if hasattr(module, "temp_weight"):
                module.temp_weight = module.weight_quantizer(module.temp_weight)
            else:
                module.temp_weight = module.weight_quantizer(module.weight)
            if not hasattr(module, "temp_bias"):
                module.temp_bias = module.bias
            module.use_temporary_parameter=True

Doesn't model.mlp.down_proj.temp_weight = model.mlp.down_proj.weight this single line of code breaks weight equivalence? If so what was the reason for breaking weight equivalence?