| 180 | } |
| 181 | |
| 182 | bool test_reset_functionality() { |
| 183 | KVCache cache; |
| 184 | cache.init(2, 1024, std::vector<size_t>(2, 64), std::vector<size_t>(2, 8), Precision::FP16); |
| 185 | cache.set_window_size(16, 4); |
| 186 | |
| 187 | CactusGraph graph; |
| 188 | graph.input({1}, Precision::FP32); |
| 189 | vector<size_t> k_nodes, v_nodes; |
| 190 | |
| 191 | for (size_t layer = 0; layer < 2; layer++) { |
| 192 | size_t k_node = graph.input({10, 8, 64}, Precision::FP16); |
| 193 | size_t v_node = graph.input({10, 8, 64}, Precision::FP16); |
| 194 | |
| 195 | vector<uint8_t> data; |
| 196 | fill_fp16(data, 10 * 8 * 64, 1.0f); |
| 197 | graph.set_input(k_node, data.data(), Precision::FP16); |
| 198 | graph.set_input(v_node, data.data(), Precision::FP16); |
| 199 | |
| 200 | k_nodes.push_back(k_node); |
| 201 | v_nodes.push_back(v_node); |
| 202 | } |
| 203 | |
| 204 | graph.execute(); |
| 205 | cache.update_from_graph(&graph, k_nodes, v_nodes, 10, 2); |
| 206 | |
| 207 | assert(cache.get_effective_seq_len() == 10); |
| 208 | assert(cache.get_total_seq_len() == 10); |
| 209 | |
| 210 | cache.reset(); |
| 211 | |
| 212 | assert(cache.get_effective_seq_len() == 0); |
| 213 | assert(cache.get_total_seq_len() == 0); |
| 214 | assert(cache.get_keys_int8(0) == nullptr); |
| 215 | assert(cache.get_values_int8(0) == nullptr); |
| 216 | |
| 217 | return true; |
| 218 | } |
| 219 | |
| 220 | bool test_large_window() { |
| 221 | const size_t num_layers = 4; |
no test coverage detected