build: fix build error when build source code on Windows

[zhouwg]

I know nothing about Windows programming, but it seems there is a minor build error when build latest source code on Windows after verified twice:

pls help to close this PR accordingly if this is a misunderstanding, thanks.

[ggerganov]

Suggested change

	/*.mem_size =*/ static_cast<size_t>(gguf_get_n_tensors(base_model.ctx_gguf)*ggml_tensor_overhead()),
	/*.mem_size =*/ ggml_tensor_overhead()*gguf_get_n_tensors(base_model.ctx_gguf),

[zhouwg]

thanks for your comment. this modification seems more elegant(implicit type conversion), validation passed on Linux and Android, but still failed on Windows(this Windows10 ISO was downloaded from MS's official website) and the compiler suggest an explicit cast is needed:

cmake --preset x64-windows-llvm-release -DGGML_OPENMP=OFF

cmake --build build-x64-windows-llvm-release

after check the details:

struct ggml_init_params {
        // memory pool
        size_t  mem_size;   // bytes
        void * mem_buffer; // if NULL, memory will be allocated internally
        bool   no_alloc;   // don't allocate memory for the tensor data
    };

size_t ggml_tensor_overhead(void) {
    return GGML_OBJECT_SIZE + GGML_TENSOR_SIZE;
}

int64_t gguf_get_n_tensors(const struct gguf_context * ctx) {
    return ctx->info.size();
}

should I modify to:

 /*.mem_size   =*/ static_cast<size_t>(ggml_tensor_overhead()*gguf_get_n_tensors(base_model.ctx_gguf)),

based on your suggestion,

or keep my original modification:

 /*.mem_size   =*/ static_cast<size_t>(gguf_get_n_tensors(base_model.ctx_gguf)*ggml_tensor_overhead()),

I guess this is a compiler issue because your suggestion works fine on Linux and Android.

[ggerganov]

I just noticed that for some reason, on your system size_t is 32-bit unsigned int. This means that likely this is a 32-bit Windows. It's better to switch to a 64-bit version, because it is very likely that llama.cpp will not run correctly on this system.

[zhouwg]

I can confirm that this is a 64-bit Windows through VMware Player, pls refer to:

I guess this is a compiler issue:

• your suggestion code can build fine with x86-64 Linux toolchain and Android's NDK
• your suggestion code contains an implicit type conversion which should works fine
  

or this issue because of this is Windows VM?