Why add 1024*1024 to ctx_size

[Metal-joker]

ggml/examples/mnist/main.cpp

Lines 79 to 83 in 64d645a

	struct ggml_init_params params = {
	/*.mem_size =*/ ctx_size + 1024*1024,
	/*.mem_buffer =*/ NULL,
	/*.no_alloc =*/ false,
	};

new to cpp and ggml, could anyone help me understand this?

[ggerganov]

It should be n_tenors * ggml_tensor_overhead() instead of 1024*1024.
But I was just being lazy to count the number of tensor and decided to put a some extra space that should be enough

[Metal-joker]

It should be n_tenors * ggml_tensor_overhead() instead of 1024*1024. But I was just being lazy to count the number of tensor and decided to put a some extra space that should be enough

Thanks man, got it now