diff --git a/ggml/src/ggml-vulkan/vulkan-shaders/mul_mmq_cm1.comp b/ggml/src/ggml-vulkan/vulkan-shaders/mul_mmq_cm1.comp
index 64951ae8a9..4f0a364248 100644
--- a/ggml/src/ggml-vulkan/vulkan-shaders/mul_mmq_cm1.comp
+++ b/ggml/src/ggml-vulkan/vulkan-shaders/mul_mmq_cm1.comp
@@ -353,33 +353,61 @@ void main() {
                     USE_MAGIC_BIAS ? ACC_BIAS_BITS : 0);
             }
 
-            coopmat<int8_t, gl_ScopeSubgroup, TM, TK, gl_MatrixUseA> cache_a;
-            coopmat<int8_t, gl_ScopeSubgroup, TK, TN, gl_MatrixUseB> cache_b;
+            const uint K_SUB = BK / TK;
+            coopmat<int8_t, gl_ScopeSubgroup, TM, TK, gl_MatrixUseA> all_a[cms_per_row * K_SUB];
+            coopmat<int8_t, gl_ScopeSubgroup, TK, TN, gl_MatrixUseB> all_b[cms_per_col * K_SUB];
 
-            [[unroll]] for (uint i = 0; i < BK; i += TK) {
-                [[unroll]] for (uint cm_row = 0; cm_row < cms_per_row; cm_row++) {
-                    coopMatLoad(cache_a, buf_a_qs, (warp_r * WM + cm_row * TM) * QPITCH + ks * (BK / 4) + i / 4, QPITCH, gl_CooperativeMatrixLayoutRowMajor);
+            [[unroll]] for (uint cm_row = 0; cm_row < cms_per_row; cm_row++) {
+                [[unroll]] for (uint h = 0; h < K_SUB; h++) {
+                    coopMatLoad(all_a[cm_row * K_SUB + h], buf_a_qs, (warp_r * WM + cm_row * TM) * QPITCH + ks * (BK / 4) + h * (TK / 4), QPITCH, gl_CooperativeMatrixLayoutRowMajor);
+                }
+            }
+            [[unroll]] for (uint cm_col = 0; cm_col < cms_per_col; cm_col++) {
+                [[unroll]] for (uint h = 0; h < K_SUB; h++) {
+                    coopMatLoad(all_b[cm_col * K_SUB + h], buf_b_qs, (warp_c * WN + cm_col * TN) * QPITCH + ks * (BK / 4) + h * (TK / 4), QPITCH, gl_CooperativeMatrixLayoutColumnMajor);
+                }
+            }
 
-                    [[unroll]] for (uint cm_col = 0; cm_col < cms_per_col; cm_col++) {
-                        coopMatLoad(cache_b, buf_b_qs, (warp_c * WN + cm_col * TN) * QPITCH + ks * (BK / 4) + i / 4, QPITCH, gl_CooperativeMatrixLayoutColumnMajor);
-
-                        cm_result[cm_col * cms_per_row + cm_row] = coopMatMulAdd(cache_a, cache_b, cm_result[cm_col * cms_per_row + cm_row]);
+            [[unroll]] for (uint cm_row = 0; cm_row < cms_per_row; cm_row++) {
+                [[unroll]] for (uint cm_col = 0; cm_col < cms_per_col; cm_col++) {
+                    [[unroll]] for (uint h = 0; h < K_SUB; h++) {
+                        cm_result[cm_col * cms_per_row + cm_row] = coopMatMulAdd(all_a[cm_row * K_SUB + h], all_b[cm_col * K_SUB + h], cm_result[cm_col * cms_per_row + cm_row]);
                     }
                 }
             }
 
-            // Apply scales directly from coopmat elements
+            // Pre-load scales into registers
+            float scale_a[cms_per_row * CM_ELEMS];
+            float nbias_a[cms_per_row * CM_ELEMS];
+            float scale_b[cms_per_col * CM_ELEMS];
+            [[unroll]] for (uint r = 0; r < cms_per_row; r++) {
+                [[unroll]] for (uint e = 0; e < CM_ELEMS; e++) {
+                    scale_a[r * CM_ELEMS + e] = buf_a_d[ks * BM + warp_r * WM + r * TM + elem_row[e]];
+                    if (USE_MAGIC_BIAS) {
+                        nbias_a[r * CM_ELEMS + e] = -ACC_BIAS_F * scale_a[r * CM_ELEMS + e];
+                    }
+                }
+            }
+            [[unroll]] for (uint c = 0; c < cms_per_col; c++) {
+                [[unroll]] for (uint e = 0; e < CM_ELEMS; e++) {
+                    scale_b[c * CM_ELEMS + e] = buf_b_d[ks * BN + warp_c * WN + c * TN + elem_col[e]];
+                }
+            }
+
+            // Apply scales from registers
             [[unroll]] for (uint cm_row = 0; cm_row < cms_per_row; cm_row++) {
                 [[unroll]] for (uint cm_col = 0; cm_col < cms_per_col; cm_col++) {
                     const uint tile_idx = cm_col * cms_per_row + cm_row;
                     [[unroll]] for (uint e = 0; e < CM_ELEMS; e++) {
-                        const float da = buf_a_d[ks * BM + warp_r * WM + cm_row * TM + elem_row[e]];
-                        const float db = buf_b_d[ks * BN + warp_c * WN + cm_col * TN + elem_col[e]];
                         if (USE_MAGIC_BIAS) {
-                            const float t = fma(intBitsToFloat(int(cm_result[tile_idx][e])), da, -ACC_BIAS_F * da);
-                            sums[tile_idx * CM_ELEMS + e] = ACC_TYPE(fma(t, db, float(sums[tile_idx * CM_ELEMS + e])));
+                            const float t = fma(intBitsToFloat(int(cm_result[tile_idx][e])),
+                                                scale_a[cm_row * CM_ELEMS + e],
+                                                nbias_a[cm_row * CM_ELEMS + e]);
+                            sums[tile_idx * CM_ELEMS + e] = ACC_TYPE(fma(t, scale_b[cm_col * CM_ELEMS + e],
+                                                                float(sums[tile_idx * CM_ELEMS + e])));
                         } else {
-                            sums[tile_idx * CM_ELEMS + e] += ACC_TYPE(float(cm_result[tile_idx][e]) * da * db);
+                            sums[tile_idx * CM_ELEMS + e] += ACC_TYPE(float(cm_result[tile_idx][e])
+                                * scale_a[cm_row * CM_ELEMS + e] * scale_b[cm_col * CM_ELEMS + e]);
                         }
                     }
                 }