ggml_init() return nullptr

[walker-ai]

const uint32_t n_embd_k_gqa = 4096;
const uint32_t n_embd_v_gqa = 4096;
const uint32_t kv_size = 4096;

std::vector<struct ggml_context *> ctxs;
int n_layers = 32;
for (int i = 0; i < n_layers; i ++ ) {
    
    struct ggml_init_params params = {
        // /*.mem_size   =*/ 2u*n_layers*ggml_tensor_overhead() + 4096 * 4096 * GGML_TYPE_F16,
        /*.mem_size   =*/ 10 * ((size_t)4096 * 4096 * 2 + ggml_tensor_overhead()), 
        /*.mem_buffer =*/ NULL,
        /*.no_alloc   =*/ false,
    };

    ggml_context * ctx = ggml_init(params);
    ctxs.push_back(ctx);
}

for (int i = 0; i < layer_num; ++ i) {
    struct ggml_tensor * k = ggml_new_tensor_1d(ctxs[i], GGML_TYPE_F16, n_embd_k_gqa * kv_size);
    struct ggml_tensor * v = ggml_new_tensor_1d(ctxs[i], GGML_TYPE_F16, n_embd_k_gqa * kv_size);
    
    k_buffer.push_back(k);
    v_buffer.push_back(v);
}

When I execute code above, I have encountered a nullptr error like this, it's so weird. If I alloc too little memory space, it should alert me to an out of memory error instead of this :(

[ngxson]

for (int i = 0; i < n_layers; i ++ ) {

for (int i = 0; i < layer_num; ++ i) {

These 2 for loop use different max value, n_layers and layer_num, and they maybe different.

[walker-ai]

I have already fixed that, the reason is the GGML_MAX_CONTEXTS = 64. I set it to 128 and it's done.