vulkan: reduce iq1 shared memory usage for mul_mm (#24287)

2026-06-09 13:27:38 +02:00
parent b4e3dc613b
commit d6d0ce8215
3 changed files with 11 additions and 2 deletions
@@ -3394,7 +3394,9 @@ static bool ggml_vk_matmul_shmem_support(const vk_device& device, const std::vec
    switch (src0_type) {
    case GGML_TYPE_IQ1_S:
    case GGML_TYPE_IQ1_M:
-        lut_size = 2*2048 + 4*2048;
+        // Regular matmul uses the compact uint16_t IQ1 grid; the expanded
+        // uint32_t grid is only enabled for the q8_1/int-dot vector path.
+        lut_size = 2*2048;
        break;
    case GGML_TYPE_IQ2_XXS:
        lut_size = 8*256;