windows11下使用deepspeed步驟

git clone https://github.com/microsoft/DeepSpeed.git

然後需要修改一下代碼：

1、打開目錄下文件：csrc/quantization/pt_binding.cpp

將其中第 203 行的

std::vector<long int>

改為：

std::vector<long long>

2、打開目錄下的文件，csrc/transformer/inference/csrc/pt_binding.cpp
其中第534行：

auto prev_key = torch::from_blob(workspace + offset,
                                     {bsz, heads, all_tokens, k},
                                     {hidden_dim * InferenceContext::Instance().GetMaxTokenLength(),
                                      k * InferenceContext::Instance().GetMaxTokenLength(),
                                      k,
                                      1},
                                     options);

添加上unsigned，即改為:

auto prev_key = torch::from_blob(workspace + offset,
                                     {bsz, heads, all_tokens, k},
                                     {hidden_dim * (unsigned)InferenceContext::Instance().GetMaxTokenLength(),
                                      k * (unsigned)InferenceContext::Instance().GetMaxTokenLength(),
                                      k,
                                      1},
                                     options);

並將1570行

auto intermediate_gemm =
        at::from_blob(intermediate_ptr, {input.size(0), input.size(1), mlp_1_out_neurons}, options);

改為：

auto intermediate_gemm =
        at::from_blob(intermediate_ptr, {input.size(0), input.size(1), (int)mlp_1_out_neurons}, options);

3、菜單中找到x64 Native Tools Command Prompt for VS 2022，鼠標右鍵，使用管理員運行build_win.bat