| 1776 | ds4_gpu_tensor_free(fused_comp); |
| 1777 | ds4_gpu_tensor_free(ref_comp); |
| 1778 | ds4_gpu_tensor_free(sc); |
| 1779 | ds4_gpu_tensor_free(kv); |
| 1780 | } |
| 1781 | |
| 1782 | static void test_metal_compressor_ratio4_replay_pack_exact(void) { |
| 1783 | test_metal_compressor_ratio4_replay_pack_exact_case(128, 4, 47); |
| 1784 | test_metal_compressor_ratio4_replay_pack_exact_case(512, 8, 53); |
| 1785 | } |
| 1786 | |
| 1787 | static void test_metal_compressor_ratio4_direct_pool_exact_case( |
| 1788 | uint32_t head_dim, |
| 1789 | uint32_t pos0, |
| 1790 | uint32_t n_tokens, |
| 1791 | uint32_t ape_type, |
| 1792 | bool replay, |
| 1793 | uint32_t seed) { |
| 1794 | const uint32_t ratio = 4u; |
| 1795 | const uint32_t width = 2u * head_dim; |
| 1796 | const uint32_t state_rows = 8u; |
| 1797 | const uint32_t n_comp = n_tokens / ratio; |
| 1798 | const uint64_t input_count = (uint64_t)n_tokens * width; |
| 1799 | const uint64_t state_count = (uint64_t)state_rows * width; |
no outgoing calls
no test coverage detected