diff --git a/ggml-vulkan.cpp b/ggml-vulkan.cpp
index 88159452f..67acf2237 100644
--- a/ggml-vulkan.cpp
+++ b/ggml-vulkan.cpp
@@ -730,30 +730,6 @@ static void ggml_vk_load_shaders() {
         vk_pipeline_dequant[GGML_TYPE_Q5_K] = ggml_vk_create_pipeline("dequant_q5_K", dequant_q5_K_len, dequant_q5_K_data, "main", 2, 4 * sizeof(int), {256 * 64, 1, 1}, {}, 1);
         vk_pipeline_dequant[GGML_TYPE_Q6_K] = ggml_vk_create_pipeline("dequant_q6_K", dequant_q6_K_len, dequant_q6_K_data, "main", 2, 4 * sizeof(int), {256 * 64, 1, 1}, {}, 1);
 
-        vk_pipeline_dequant_mul_mat_vec[GGML_TYPE_F16] = ggml_vk_create_pipeline("mul_mat_vec_f16", mul_mat_vec_f16_len, mul_mat_vec_f16_data, "main", 3, 1 * sizeof(int), {1, 1, 1}, {}, 1);
-        vk_pipeline_dequant_mul_mat_vec[GGML_TYPE_Q4_0] = ggml_vk_create_pipeline("mul_mat_vec_q4_0", mul_mat_vec_q4_0_len, mul_mat_vec_q4_0_data, "main", 3, 1 * sizeof(int), {1, 1, 1}, {}, 1);
-        vk_pipeline_dequant_mul_mat_vec[GGML_TYPE_Q4_1] = ggml_vk_create_pipeline("mul_mat_vec_q4_1", mul_mat_vec_q4_1_len, mul_mat_vec_q4_1_data, "main", 3, 1 * sizeof(int), {1, 1, 1}, {}, 1);
-        vk_pipeline_dequant_mul_mat_vec[GGML_TYPE_Q5_0] = ggml_vk_create_pipeline("mul_mat_vec_q5_0", mul_mat_vec_q5_0_len, mul_mat_vec_q5_0_data, "main", 3, 1 * sizeof(int), {1, 1, 1}, {}, 1);
-        vk_pipeline_dequant_mul_mat_vec[GGML_TYPE_Q5_1] = ggml_vk_create_pipeline("mul_mat_vec_q5_1", mul_mat_vec_q5_1_len, mul_mat_vec_q5_1_data, "main", 3, 1 * sizeof(int), {1, 1, 1}, {}, 1);
-        vk_pipeline_dequant_mul_mat_vec[GGML_TYPE_Q8_0] = ggml_vk_create_pipeline("mul_mat_vec_q8_0", mul_mat_vec_q8_0_len, mul_mat_vec_q8_0_data, "main", 3, 1 * sizeof(int), {1, 1, 1}, {}, 1);
-        vk_pipeline_dequant_mul_mat_vec[GGML_TYPE_Q2_K] = ggml_vk_create_pipeline("mul_mat_vec_q2_K", mul_mat_vec_q2_K_len, mul_mat_vec_q2_K_data, "main", 3, 1 * sizeof(int), {1, 1, 1}, {}, 1);
-        vk_pipeline_dequant_mul_mat_vec[GGML_TYPE_Q3_K] = ggml_vk_create_pipeline("mul_mat_vec_q3_K", mul_mat_vec_q3_K_len, mul_mat_vec_q3_K_data, "main", 3, 1 * sizeof(int), {1, 1, 1}, {}, 1);
-        vk_pipeline_dequant_mul_mat_vec[GGML_TYPE_Q4_K] = ggml_vk_create_pipeline("mul_mat_vec_q4_K", mul_mat_vec_q4_K_len, mul_mat_vec_q4_K_data, "main", 3, 1 * sizeof(int), {1, 1, 1}, {}, 1);
-        vk_pipeline_dequant_mul_mat_vec[GGML_TYPE_Q5_K] = ggml_vk_create_pipeline("mul_mat_vec_q5_K", mul_mat_vec_q5_K_len, mul_mat_vec_q5_K_data, "main", 3, 1 * sizeof(int), {1, 1, 1}, {}, 1);
-        vk_pipeline_dequant_mul_mat_vec[GGML_TYPE_Q6_K] = ggml_vk_create_pipeline("mul_mat_vec_q6_K", mul_mat_vec_q6_K_len, mul_mat_vec_q6_K_data, "main", 3, 1 * sizeof(int), {1, 1, 1}, {}, 1);
-
-        vk_pipeline_dequant_mul_mat_vec_f32[GGML_TYPE_F16] = ggml_vk_create_pipeline("mul_mat_vec_f16_f32", mul_mat_vec_f16_f32_len, mul_mat_vec_f16_f32_data, "main", 3, 1 * sizeof(int), {1, 1, 1}, {}, 1);
-        vk_pipeline_dequant_mul_mat_vec_f32[GGML_TYPE_Q4_0] = ggml_vk_create_pipeline("mul_mat_vec_q4_0_f32", mul_mat_vec_q4_0_f32_len, mul_mat_vec_q4_0_f32_data, "main", 3, 1 * sizeof(int), {1, 1, 1}, {}, 1);
-        vk_pipeline_dequant_mul_mat_vec_f32[GGML_TYPE_Q4_1] = ggml_vk_create_pipeline("mul_mat_vec_q4_1_f32", mul_mat_vec_q4_1_f32_len, mul_mat_vec_q4_1_f32_data, "main", 3, 1 * sizeof(int), {1, 1, 1}, {}, 1);
-        vk_pipeline_dequant_mul_mat_vec_f32[GGML_TYPE_Q5_0] = ggml_vk_create_pipeline("mul_mat_vec_q5_0_f32", mul_mat_vec_q5_0_f32_len, mul_mat_vec_q5_0_f32_data, "main", 3, 1 * sizeof(int), {1, 1, 1}, {}, 1);
-        vk_pipeline_dequant_mul_mat_vec_f32[GGML_TYPE_Q5_1] = ggml_vk_create_pipeline("mul_mat_vec_q5_1_f32", mul_mat_vec_q5_1_f32_len, mul_mat_vec_q5_1_f32_data, "main", 3, 1 * sizeof(int), {1, 1, 1}, {}, 1);
-        vk_pipeline_dequant_mul_mat_vec_f32[GGML_TYPE_Q8_0] = ggml_vk_create_pipeline("mul_mat_vec_q8_0_f32", mul_mat_vec_q8_0_f32_len, mul_mat_vec_q8_0_f32_data, "main", 3, 1 * sizeof(int), {1, 1, 1}, {}, 1);
-        vk_pipeline_dequant_mul_mat_vec_f32[GGML_TYPE_Q2_K] = ggml_vk_create_pipeline("mul_mat_vec_q2_K_f32", mul_mat_vec_q2_K_f32_len, mul_mat_vec_q2_K_f32_data, "main", 3, 1 * sizeof(int), {1, 1, 1}, {}, 1);
-        vk_pipeline_dequant_mul_mat_vec_f32[GGML_TYPE_Q3_K] = ggml_vk_create_pipeline("mul_mat_vec_q3_K_f32", mul_mat_vec_q3_K_f32_len, mul_mat_vec_q3_K_f32_data, "main", 3, 1 * sizeof(int), {1, 1, 1}, {}, 1);
-        vk_pipeline_dequant_mul_mat_vec_f32[GGML_TYPE_Q4_K] = ggml_vk_create_pipeline("mul_mat_vec_q4_K_f32", mul_mat_vec_q4_K_f32_len, mul_mat_vec_q4_K_f32_data, "main", 3, 1 * sizeof(int), {1, 1, 1}, {}, 1);
-        vk_pipeline_dequant_mul_mat_vec_f32[GGML_TYPE_Q5_K] = ggml_vk_create_pipeline("mul_mat_vec_q5_K_f32", mul_mat_vec_q5_K_f32_len, mul_mat_vec_q5_K_f32_data, "main", 3, 1 * sizeof(int), {1, 1, 1}, {}, 1);
-        vk_pipeline_dequant_mul_mat_vec_f32[GGML_TYPE_Q6_K] = ggml_vk_create_pipeline("mul_mat_vec_q6_K_f32", mul_mat_vec_q6_K_f32_len, mul_mat_vec_q6_K_f32_data, "main", 3, 1 * sizeof(int), {1, 1, 1}, {}, 1);
-
         // get_rows
         vk_pipeline_get_rows[GGML_TYPE_F16] = ggml_vk_create_pipeline("get_rows_f16", get_rows_f16_len, get_rows_f16_data, "main", 3, sizeof(vk_op_push_constants), {512, 1, 1}, {}, 1);
         vk_pipeline_get_rows[GGML_TYPE_Q4_0] = ggml_vk_create_pipeline("get_rows_q4_0", get_rows_q4_0_len, get_rows_q4_0_data, "main", 3, sizeof(vk_op_push_constants), {512, 1, 1}, {}, 1);
@@ -817,30 +793,6 @@ static void ggml_vk_load_shaders() {
         vk_pipeline_dequant[GGML_TYPE_Q5_K] = ggml_vk_create_pipeline("dequant_q5_K", dequant_q5_K_fp32_len, dequant_q5_K_fp32_data, "main", 2, 4 * sizeof(int), {256 * 32, 1, 1}, {}, 1);
         vk_pipeline_dequant[GGML_TYPE_Q6_K] = ggml_vk_create_pipeline("dequant_q6_K", dequant_q6_K_fp32_len, dequant_q6_K_fp32_data, "main", 2, 4 * sizeof(int), {256 * 32, 1, 1}, {}, 1);
 
-        vk_pipeline_dequant_mul_mat_vec[GGML_TYPE_F16] = ggml_vk_create_pipeline("mul_mat_vec_f16", mul_mat_vec_f16_fp32_len, mul_mat_vec_f16_fp32_data, "main", 3, 1 * sizeof(int), {1, 1, 1}, {}, 1);
-        vk_pipeline_dequant_mul_mat_vec[GGML_TYPE_Q4_0] = ggml_vk_create_pipeline("mul_mat_vec_q4_0", mul_mat_vec_q4_0_fp32_len, mul_mat_vec_q4_0_fp32_data, "main", 3, 1 * sizeof(int), {1, 1, 1}, {}, 1);
-        vk_pipeline_dequant_mul_mat_vec[GGML_TYPE_Q4_1] = ggml_vk_create_pipeline("mul_mat_vec_q4_1", mul_mat_vec_q4_1_fp32_len, mul_mat_vec_q4_1_fp32_data, "main", 3, 1 * sizeof(int), {1, 1, 1}, {}, 1);
-        vk_pipeline_dequant_mul_mat_vec[GGML_TYPE_Q5_0] = ggml_vk_create_pipeline("mul_mat_vec_q5_0", mul_mat_vec_q5_0_fp32_len, mul_mat_vec_q5_0_fp32_data, "main", 3, 1 * sizeof(int), {1, 1, 1}, {}, 1);
-        vk_pipeline_dequant_mul_mat_vec[GGML_TYPE_Q5_1] = ggml_vk_create_pipeline("mul_mat_vec_q5_1", mul_mat_vec_q5_1_fp32_len, mul_mat_vec_q5_1_fp32_data, "main", 3, 1 * sizeof(int), {1, 1, 1}, {}, 1);
-        vk_pipeline_dequant_mul_mat_vec[GGML_TYPE_Q8_0] = ggml_vk_create_pipeline("mul_mat_vec_q8_0", mul_mat_vec_q8_0_fp32_len, mul_mat_vec_q8_0_fp32_data, "main", 3, 1 * sizeof(int), {1, 1, 1}, {}, 1);
-        vk_pipeline_dequant_mul_mat_vec[GGML_TYPE_Q2_K] = ggml_vk_create_pipeline("mul_mat_vec_q2_K", mul_mat_vec_q2_K_fp32_len, mul_mat_vec_q2_K_fp32_data, "main", 3, 1 * sizeof(int), {1, 1, 1}, {}, 1);
-        vk_pipeline_dequant_mul_mat_vec[GGML_TYPE_Q3_K] = ggml_vk_create_pipeline("mul_mat_vec_q3_K", mul_mat_vec_q3_K_fp32_len, mul_mat_vec_q3_K_fp32_data, "main", 3, 1 * sizeof(int), {1, 1, 1}, {}, 1);
-        vk_pipeline_dequant_mul_mat_vec[GGML_TYPE_Q4_K] = ggml_vk_create_pipeline("mul_mat_vec_q4_K", mul_mat_vec_q4_K_fp32_len, mul_mat_vec_q4_K_fp32_data, "main", 3, 1 * sizeof(int), {1, 1, 1}, {}, 1);
-        vk_pipeline_dequant_mul_mat_vec[GGML_TYPE_Q5_K] = ggml_vk_create_pipeline("mul_mat_vec_q5_K", mul_mat_vec_q5_K_fp32_len, mul_mat_vec_q5_K_fp32_data, "main", 3, 1 * sizeof(int), {1, 1, 1}, {}, 1);
-        vk_pipeline_dequant_mul_mat_vec[GGML_TYPE_Q6_K] = ggml_vk_create_pipeline("mul_mat_vec_q6_K", mul_mat_vec_q6_K_fp32_len, mul_mat_vec_q6_K_fp32_data, "main", 3, 1 * sizeof(int), {1, 1, 1}, {}, 1);
-
-        vk_pipeline_dequant_mul_mat_vec_f32[GGML_TYPE_F16] = ggml_vk_create_pipeline("mul_mat_vec_f16_f32", mul_mat_vec_f16_f32_fp32_len, mul_mat_vec_f16_f32_fp32_data, "main", 3, 1 * sizeof(int), {1, 1, 1}, {}, 1);
-        vk_pipeline_dequant_mul_mat_vec_f32[GGML_TYPE_Q4_0] = ggml_vk_create_pipeline("mul_mat_vec_q4_0_f32", mul_mat_vec_q4_0_f32_fp32_len, mul_mat_vec_q4_0_f32_fp32_data, "main", 3, 1 * sizeof(int), {1, 1, 1}, {}, 1);
-        vk_pipeline_dequant_mul_mat_vec_f32[GGML_TYPE_Q4_1] = ggml_vk_create_pipeline("mul_mat_vec_q4_1_f32", mul_mat_vec_q4_1_f32_fp32_len, mul_mat_vec_q4_1_f32_fp32_data, "main", 3, 1 * sizeof(int), {1, 1, 1}, {}, 1);
-        vk_pipeline_dequant_mul_mat_vec_f32[GGML_TYPE_Q5_0] = ggml_vk_create_pipeline("mul_mat_vec_q5_0_f32", mul_mat_vec_q5_0_f32_fp32_len, mul_mat_vec_q5_0_f32_fp32_data, "main", 3, 1 * sizeof(int), {1, 1, 1}, {}, 1);
-        vk_pipeline_dequant_mul_mat_vec_f32[GGML_TYPE_Q5_1] = ggml_vk_create_pipeline("mul_mat_vec_q5_1_f32", mul_mat_vec_q5_1_f32_fp32_len, mul_mat_vec_q5_1_f32_fp32_data, "main", 3, 1 * sizeof(int), {1, 1, 1}, {}, 1);
-        vk_pipeline_dequant_mul_mat_vec_f32[GGML_TYPE_Q8_0] = ggml_vk_create_pipeline("mul_mat_vec_q8_0_f32", mul_mat_vec_q8_0_f32_fp32_len, mul_mat_vec_q8_0_f32_fp32_data, "main", 3, 1 * sizeof(int), {1, 1, 1}, {}, 1);
-        vk_pipeline_dequant_mul_mat_vec_f32[GGML_TYPE_Q2_K] = ggml_vk_create_pipeline("mul_mat_vec_q2_K_f32", mul_mat_vec_q2_K_f32_fp32_len, mul_mat_vec_q2_K_f32_fp32_data, "main", 3, 1 * sizeof(int), {1, 1, 1}, {}, 1);
-        vk_pipeline_dequant_mul_mat_vec_f32[GGML_TYPE_Q3_K] = ggml_vk_create_pipeline("mul_mat_vec_q3_K_f32", mul_mat_vec_q3_K_f32_fp32_len, mul_mat_vec_q3_K_f32_fp32_data, "main", 3, 1 * sizeof(int), {1, 1, 1}, {}, 1);
-        vk_pipeline_dequant_mul_mat_vec_f32[GGML_TYPE_Q4_K] = ggml_vk_create_pipeline("mul_mat_vec_q4_K_f32", mul_mat_vec_q4_K_f32_fp32_len, mul_mat_vec_q4_K_f32_fp32_data, "main", 3, 1 * sizeof(int), {1, 1, 1}, {}, 1);
-        vk_pipeline_dequant_mul_mat_vec_f32[GGML_TYPE_Q5_K] = ggml_vk_create_pipeline("mul_mat_vec_q5_K_f32", mul_mat_vec_q5_K_f32_fp32_len, mul_mat_vec_q5_K_f32_fp32_data, "main", 3, 1 * sizeof(int), {1, 1, 1}, {}, 1);
-        vk_pipeline_dequant_mul_mat_vec_f32[GGML_TYPE_Q6_K] = ggml_vk_create_pipeline("mul_mat_vec_q6_K_f32", mul_mat_vec_q6_K_f32_fp32_len, mul_mat_vec_q6_K_f32_fp32_data, "main", 3, 1 * sizeof(int), {1, 1, 1}, {}, 1);
-
         // get_rows
         vk_pipeline_get_rows[GGML_TYPE_F16] = ggml_vk_create_pipeline("get_rows_f16", get_rows_f16_fp32_len, get_rows_f16_fp32_data, "main", 3, sizeof(vk_op_push_constants), {512, 1, 1}, {}, 1);
         vk_pipeline_get_rows[GGML_TYPE_Q4_0] = ggml_vk_create_pipeline("get_rows_q4_0", get_rows_q4_0_fp32_len, get_rows_q4_0_fp32_data, "main", 3, sizeof(vk_op_push_constants), {512, 1, 1}, {}, 1);
@@ -866,6 +818,30 @@ static void ggml_vk_load_shaders() {
         vk_pipeline_scale_f32 = ggml_vk_create_pipeline("scale_f32", scale_f32_fp32_len, scale_f32_fp32_data, "main", 2, sizeof(vk_op_push_constants), {512, 1, 1}, {}, 1);
     }
 
+    vk_pipeline_dequant_mul_mat_vec[GGML_TYPE_F16] = ggml_vk_create_pipeline("mul_mat_vec_f16", mul_mat_vec_f16_len, mul_mat_vec_f16_data, "main", 3, 1 * sizeof(int), {1, 1, 1}, {}, 1);
+    vk_pipeline_dequant_mul_mat_vec[GGML_TYPE_Q4_0] = ggml_vk_create_pipeline("mul_mat_vec_q4_0", mul_mat_vec_q4_0_len, mul_mat_vec_q4_0_data, "main", 3, 1 * sizeof(int), {1, 1, 1}, {}, 1);
+    vk_pipeline_dequant_mul_mat_vec[GGML_TYPE_Q4_1] = ggml_vk_create_pipeline("mul_mat_vec_q4_1", mul_mat_vec_q4_1_len, mul_mat_vec_q4_1_data, "main", 3, 1 * sizeof(int), {1, 1, 1}, {}, 1);
+    vk_pipeline_dequant_mul_mat_vec[GGML_TYPE_Q5_0] = ggml_vk_create_pipeline("mul_mat_vec_q5_0", mul_mat_vec_q5_0_len, mul_mat_vec_q5_0_data, "main", 3, 1 * sizeof(int), {1, 1, 1}, {}, 1);
+    vk_pipeline_dequant_mul_mat_vec[GGML_TYPE_Q5_1] = ggml_vk_create_pipeline("mul_mat_vec_q5_1", mul_mat_vec_q5_1_len, mul_mat_vec_q5_1_data, "main", 3, 1 * sizeof(int), {1, 1, 1}, {}, 1);
+    vk_pipeline_dequant_mul_mat_vec[GGML_TYPE_Q8_0] = ggml_vk_create_pipeline("mul_mat_vec_q8_0", mul_mat_vec_q8_0_len, mul_mat_vec_q8_0_data, "main", 3, 1 * sizeof(int), {1, 1, 1}, {}, 1);
+    vk_pipeline_dequant_mul_mat_vec[GGML_TYPE_Q2_K] = ggml_vk_create_pipeline("mul_mat_vec_q2_K", mul_mat_vec_q2_K_len, mul_mat_vec_q2_K_data, "main", 3, 1 * sizeof(int), {1, 1, 1}, {}, 1);
+    vk_pipeline_dequant_mul_mat_vec[GGML_TYPE_Q3_K] = ggml_vk_create_pipeline("mul_mat_vec_q3_K", mul_mat_vec_q3_K_len, mul_mat_vec_q3_K_data, "main", 3, 1 * sizeof(int), {1, 1, 1}, {}, 1);
+    vk_pipeline_dequant_mul_mat_vec[GGML_TYPE_Q4_K] = ggml_vk_create_pipeline("mul_mat_vec_q4_K", mul_mat_vec_q4_K_len, mul_mat_vec_q4_K_data, "main", 3, 1 * sizeof(int), {1, 1, 1}, {}, 1);
+    vk_pipeline_dequant_mul_mat_vec[GGML_TYPE_Q5_K] = ggml_vk_create_pipeline("mul_mat_vec_q5_K", mul_mat_vec_q5_K_len, mul_mat_vec_q5_K_data, "main", 3, 1 * sizeof(int), {1, 1, 1}, {}, 1);
+    vk_pipeline_dequant_mul_mat_vec[GGML_TYPE_Q6_K] = ggml_vk_create_pipeline("mul_mat_vec_q6_K", mul_mat_vec_q6_K_len, mul_mat_vec_q6_K_data, "main", 3, 1 * sizeof(int), {1, 1, 1}, {}, 1);
+
+    vk_pipeline_dequant_mul_mat_vec_f32[GGML_TYPE_F16] = ggml_vk_create_pipeline("mul_mat_vec_f16_f32", mul_mat_vec_f16_f32_len, mul_mat_vec_f16_f32_data, "main", 3, 1 * sizeof(int), {1, 1, 1}, {}, 1);
+    vk_pipeline_dequant_mul_mat_vec_f32[GGML_TYPE_Q4_0] = ggml_vk_create_pipeline("mul_mat_vec_q4_0_f32", mul_mat_vec_q4_0_f32_len, mul_mat_vec_q4_0_f32_data, "main", 3, 1 * sizeof(int), {1, 1, 1}, {}, 1);
+    vk_pipeline_dequant_mul_mat_vec_f32[GGML_TYPE_Q4_1] = ggml_vk_create_pipeline("mul_mat_vec_q4_1_f32", mul_mat_vec_q4_1_f32_len, mul_mat_vec_q4_1_f32_data, "main", 3, 1 * sizeof(int), {1, 1, 1}, {}, 1);
+    vk_pipeline_dequant_mul_mat_vec_f32[GGML_TYPE_Q5_0] = ggml_vk_create_pipeline("mul_mat_vec_q5_0_f32", mul_mat_vec_q5_0_f32_len, mul_mat_vec_q5_0_f32_data, "main", 3, 1 * sizeof(int), {1, 1, 1}, {}, 1);
+    vk_pipeline_dequant_mul_mat_vec_f32[GGML_TYPE_Q5_1] = ggml_vk_create_pipeline("mul_mat_vec_q5_1_f32", mul_mat_vec_q5_1_f32_len, mul_mat_vec_q5_1_f32_data, "main", 3, 1 * sizeof(int), {1, 1, 1}, {}, 1);
+    vk_pipeline_dequant_mul_mat_vec_f32[GGML_TYPE_Q8_0] = ggml_vk_create_pipeline("mul_mat_vec_q8_0_f32", mul_mat_vec_q8_0_f32_len, mul_mat_vec_q8_0_f32_data, "main", 3, 1 * sizeof(int), {1, 1, 1}, {}, 1);
+    vk_pipeline_dequant_mul_mat_vec_f32[GGML_TYPE_Q2_K] = ggml_vk_create_pipeline("mul_mat_vec_q2_K_f32", mul_mat_vec_q2_K_f32_len, mul_mat_vec_q2_K_f32_data, "main", 3, 1 * sizeof(int), {1, 1, 1}, {}, 1);
+    vk_pipeline_dequant_mul_mat_vec_f32[GGML_TYPE_Q3_K] = ggml_vk_create_pipeline("mul_mat_vec_q3_K_f32", mul_mat_vec_q3_K_f32_len, mul_mat_vec_q3_K_f32_data, "main", 3, 1 * sizeof(int), {1, 1, 1}, {}, 1);
+    vk_pipeline_dequant_mul_mat_vec_f32[GGML_TYPE_Q4_K] = ggml_vk_create_pipeline("mul_mat_vec_q4_K_f32", mul_mat_vec_q4_K_f32_len, mul_mat_vec_q4_K_f32_data, "main", 3, 1 * sizeof(int), {1, 1, 1}, {}, 1);
+    vk_pipeline_dequant_mul_mat_vec_f32[GGML_TYPE_Q5_K] = ggml_vk_create_pipeline("mul_mat_vec_q5_K_f32", mul_mat_vec_q5_K_f32_len, mul_mat_vec_q5_K_f32_data, "main", 3, 1 * sizeof(int), {1, 1, 1}, {}, 1);
+    vk_pipeline_dequant_mul_mat_vec_f32[GGML_TYPE_Q6_K] = ggml_vk_create_pipeline("mul_mat_vec_q6_K_f32", mul_mat_vec_q6_K_f32_len, mul_mat_vec_q6_K_f32_data, "main", 3, 1 * sizeof(int), {1, 1, 1}, {}, 1);
+
     vk_pipeline_rms_norm_f32 = ggml_vk_create_pipeline("rms_norm_f32", rms_norm_f32_len, rms_norm_f32_data, "main", 2, sizeof(vk_op_push_constants), {1, 1, 1}, {}, 1);
 }
 
@@ -2585,8 +2561,8 @@ void ggml_vk_preallocate_buffers_graph(ggml_tensor * node){
 #endif
     node->extra = nullptr;
 
-    const bool src0_gpu = node->src[0] != nullptr && node->src[0]->ne[1] > 32 && node->src[0]->extra != nullptr && node->src[0]->backend == GGML_BACKEND_CPU;
-    const bool src1_gpu = node->src[1] != nullptr && node->src[1]->ne[1] > 32 && node->src[1]->extra != nullptr && node->src[1]->backend == GGML_BACKEND_CPU;
+    const bool src0_gpu = false;  // node->src[0] != nullptr && node->src[0]->ne[1] > 32 && node->src[0]->extra != nullptr && node->src[0]->backend == GGML_BACKEND_CPU;
+    const bool src1_gpu = false;  // node->src[1] != nullptr && node->src[1]->ne[1] > 32 && node->src[1]->extra != nullptr && node->src[1]->backend == GGML_BACKEND_CPU;
 
     const bool any_on_device = node->backend == GGML_BACKEND_GPU
         || (node->src[0] != nullptr && (node->src[0]->backend == GGML_BACKEND_GPU || node->src[0]->backend == GGML_BACKEND_GPU_SPLIT || src0_gpu))
diff --git a/ggml_vk_generate_shaders.py b/ggml_vk_generate_shaders.py
index ec1fd55f3..e474d9537 100644
--- a/ggml_vk_generate_shaders.py
+++ b/ggml_vk_generate_shaders.py
@@ -1653,39 +1653,6 @@ async def main():
 
             tasks.append(string_to_spv(f"dequant_{type_names[i]}", "".join(stream), {"D_TYPE": "float16_t"}, fp16))
 
-        # mul mat vec
-        for i in range(0, VK_NUM_TYPES):
-            stream.clear()
-            stream.extend((mul_mat_vec_head, shader_int8_ext, shader_float_type))
-
-            if i == GGML_TYPE_F16:
-                stream.extend((shader_f16_defines, shader_f16_dequant_func_compat if not fp16 else shader_f16_dequant_func, mul_mat_vec_body))
-            elif i == GGML_TYPE_Q4_0:
-                stream.extend((shader_q4_0_defines, shader_q4_0_dequant_func_compat if not fp16 else shader_q4_0_dequant_func, mul_mat_vec_body))
-            elif i == GGML_TYPE_Q4_1:
-                stream.extend((shader_q4_1_defines, shader_q4_1_dequant_func_compat if not fp16 else shader_q4_1_dequant_func, mul_mat_vec_body))
-            elif i == GGML_TYPE_Q5_0:
-                stream.extend((shader_q5_0_defines, shader_q5_0_dequant_func_compat if not fp16 else shader_q5_0_dequant_func, mul_mat_vec_body))
-            elif i == GGML_TYPE_Q5_1:
-                stream.extend((shader_q5_1_defines, shader_q5_1_dequant_func_compat if not fp16 else shader_q5_1_dequant_func, mul_mat_vec_body))
-            elif i == GGML_TYPE_Q8_0:
-                stream.extend((shader_q8_0_defines, shader_q8_0_dequant_func_compat if not fp16 else shader_q8_0_dequant_func, mul_mat_vec_body))
-            elif i == GGML_TYPE_Q2_K:
-                stream.extend((shader_q2_K_defines, mul_mat_vec_q2_K_body))
-            elif i == GGML_TYPE_Q3_K:
-                stream.extend((shader_q3_K_defines, mul_mat_vec_q3_K_body))
-            elif i == GGML_TYPE_Q4_K:
-                stream.extend((shader_q4_K_defines, mul_mat_vec_q4_K_body))
-            elif i == GGML_TYPE_Q5_K:
-                stream.extend((shader_q5_K_defines, mul_mat_vec_q5_K_body))
-            elif i == GGML_TYPE_Q6_K:
-                stream.extend((shader_q6_K_defines, mul_mat_vec_q6_K_body))
-            else:
-                continue
-
-            tasks.append(string_to_spv(f"mul_mat_vec_{type_names[i]}", "".join(stream), {"B_TYPE": "float", "D_TYPE": "float16_t", "K_QUANTS_PER_ITERATION": K_QUANTS_PER_ITERATION}, fp16))
-            tasks.append(string_to_spv(f"mul_mat_vec_{type_names[i]}_f32", "".join(stream), {"B_TYPE": "float", "D_TYPE": "float", "K_QUANTS_PER_ITERATION": K_QUANTS_PER_ITERATION}, fp16))
-
         # get_rows
         for i in range(0, VK_NUM_TYPES):
             stream.clear()
@@ -1722,6 +1689,41 @@ async def main():
         tasks.append(string_to_spv("scale_f32", f"{generic_head}\n{shader_float_type}\n{scale_body}", {"A_TYPE": "float", "D_TYPE": "float"}, fp16))
 
     # Shaders where precision is needed, so no fp16 version
+
+    # mul mat vec
+    for i in range(0, VK_NUM_TYPES):
+        stream.clear()
+        stream.extend((mul_mat_vec_head, shader_int8_ext, shader_f32))
+
+        if i == GGML_TYPE_F16:
+            stream.extend((shader_f16_defines, shader_f16_dequant_func_compat, mul_mat_vec_body))
+        elif i == GGML_TYPE_Q4_0:
+            stream.extend((shader_q4_0_defines, shader_q4_0_dequant_func_compat, mul_mat_vec_body))
+        elif i == GGML_TYPE_Q4_1:
+            stream.extend((shader_q4_1_defines, shader_q4_1_dequant_func_compat, mul_mat_vec_body))
+        elif i == GGML_TYPE_Q5_0:
+            stream.extend((shader_q5_0_defines, shader_q5_0_dequant_func_compat, mul_mat_vec_body))
+        elif i == GGML_TYPE_Q5_1:
+            stream.extend((shader_q5_1_defines, shader_q5_1_dequant_func_compat, mul_mat_vec_body))
+        elif i == GGML_TYPE_Q8_0:
+            stream.extend((shader_q8_0_defines, shader_q8_0_dequant_func_compat, mul_mat_vec_body))
+        elif i == GGML_TYPE_Q2_K:
+            stream.extend((shader_q2_K_defines, mul_mat_vec_q2_K_body))
+        elif i == GGML_TYPE_Q3_K:
+            stream.extend((shader_q3_K_defines, mul_mat_vec_q3_K_body))
+        elif i == GGML_TYPE_Q4_K:
+            stream.extend((shader_q4_K_defines, mul_mat_vec_q4_K_body))
+        elif i == GGML_TYPE_Q5_K:
+            stream.extend((shader_q5_K_defines, mul_mat_vec_q5_K_body))
+        elif i == GGML_TYPE_Q6_K:
+            stream.extend((shader_q6_K_defines, mul_mat_vec_q6_K_body))
+        else:
+            continue
+
+        tasks.append(string_to_spv(f"mul_mat_vec_{type_names[i]}", "".join(stream), {"B_TYPE": "float", "D_TYPE": "float16_t", "K_QUANTS_PER_ITERATION": K_QUANTS_PER_ITERATION}, fp16))
+        tasks.append(string_to_spv(f"mul_mat_vec_{type_names[i]}_f32", "".join(stream), {"B_TYPE": "float", "D_TYPE": "float", "K_QUANTS_PER_ITERATION": K_QUANTS_PER_ITERATION}, fp16))
+
+    # RMS Norm
     tasks.append(string_to_spv("rms_norm_f32", f"{generic_head}\n{shader_f32}\n{rms_norm_body}", {"A_TYPE": "float", "D_TYPE": "float"}, True))
 
     await asyncio.gather(*tasks)