Function llm_build_sparse_mul_mat

llama.cpp:4546–4597

}

static struct ggml_tensor * llm_build_sparse_mul_mat(
        struct ggml_context * ctx,
         struct ggml_tensor * up,
         struct ggml_tensor * inp,
         struct ggml_tensor * idx,
         struct ggml_tensor * up_gpu,
         struct ggml_tensor * gpu_index,
         struct ggml_tensor * gpu_bucket,
   const llm_build_cb_short & cb,
                 const char * name,
                         bool full_gpu) {
    std::string full_name = "ffn_" + std::string(name) + "_sparse";
    ggml_tensor * out = nullptr;

#ifdef GGML_USE_HIPBLAS
// WARNING: THIS IS A HACK! 
// if up_gpu->data is null
// inference fails when model exceeds 40B on rocm device
// so we just let up_gpu->data point to itself
    
    up_gpu->data = up_gpu;

#endif 

#ifdef GGML_USE_CUBLAS
    // Full offloading fast path
    if (full_gpu) {
        GGML_ASSERT(up_gpu && "full_gpu but no up_gpu");
        out = ggml_mul_mat_idx(ctx, up_gpu, inp, idx, NULL);
        ggml_cuda_assign_buffers_no_alloc(out);
        cb(out, (full_name).c_str());
        return out;
    }
#endif

    out = ggml_mul_mat_idx(ctx, up, inp, idx, gpu_index);
    cb(out, full_name.c_str());

#ifdef GGML_USE_CUBLAS
    if (up_gpu) {
        ggml_tensor * out_gpu = ggml_mul_mat_idx_upscale(ctx, up_gpu, inp, idx, gpu_bucket, out->ne[0]);
        ggml_cuda_assign_buffers_no_alloc(out_gpu);
        cb(out_gpu, (full_name + "_gpu").c_str());
        out = ggml_add(ctx, out, out_gpu);
        // We don't need to assign buffers here, as the output will be passed into Axpy,
        // which in this case, is also a hybrid operation.
        cb(out, (full_name + "_merged").c_str());
    }
#endif

    return out;
}

static struct ggml_tensor * llm_build_sparse_axpy(
        struct ggml_context * ctx,