Unable to export LLaMa2 model to bin file

[guoqingbao]

I have the following errors:

Exception has occurred: AttributeError
'GeneralQuantLinear' object has no attribute 'build'
  File "/root/llama2.rs/export.py", line 57, in export
    hidden_dim = model.layers[0].mlp.up_proj.build()[0].shape[1]
  File "/root/llama2.rs/export.py", line 127, in load_and_export
    export(model, output_path)
  File "/root/llama2.rs/export.py", line 135, in <module>
    load_and_export("", output_path)
AttributeError: 'GeneralQuantLinear' object has no attribute 'build'

My Python package versions:

Package Version

---

auto-gptq 0.3.1 huggingface-hub 0.16.4 torch 2.1.0a0+fe05266 transformer-engine 0.7.0 transformers 4.31.0

[guoqingbao]

I have managed to resolve the problem by revising the following code in "export" function, refer to PR #18

    def serialize(k):
        # w = None
        # if isinstance(k, torch.Tensor):
        #     w = k       
        # elif "GeneralQuantLinear" in str(k.__class__) and EXPAND:
        #     w = k.build()[0].T

        # elif "GeneralQuantLinear" not in str(k.__class__):
        #     w = k.weight

        if hasattr(k, "qweight"):
            for w in [k.qweight.type(torch.int32), k.qzeros.type(torch.int32), k.scales.type(torch.float32)]:
                print("Quant")
                print(w.shape)
                t = w.T.contiguous().view(-1).detach().cpu().numpy()
                f.write(memoryview(t))
        else:
            if hasattr(k, "weight"):
                w = k.weight
            else:
                w = k
            print("Regular")
            print(w.shape)
            t = w.contiguous().view(-1).detach().cpu().type(torch.float32).numpy()
            f.write(memoryview(t))

        # del state_dict[key]

    # first write out the header
    p['n_heads'] = model.layers[0].self_attn.num_heads
    # hidden_dim = model.layers[0].mlp.up_proj.build()[0].shape[1]
    # p['dim'] = model.layers[0].mlp.up_proj.build()[0].shape[0]
    hidden_dim = model.layers[0].mlp.up_proj.out_features
    p['dim'] = model.layers[0].mlp.up_proj.in_features