Vulkan Mixture of Experts (MoE) support (llama/7628)

* Finish Vulkan mul_mat_id implementation

* Add Vulkan sum_rows and div ops

* Fix MUL_MAT_ID matrix matrix shader

* Fix MUL_MAT_ID matrix vector shader dispatch size

* Fix MUL_MAT_ID matrix vector shader and dispatch code

* Update Vulkan CPU offload for MUL_MAT_ID

* Fix crash when using split mode none and setting a main GPU

diff --git a/src/ggml-vulkan.cpp b/src/ggml-vulkan.cpp
index 92e622b04317720c452a68c55a6a3cd16bba8628..5e12ea9dde4d7b53d217451b2c3e040f686625a9 100644 (file)
--- a/src/ggml-vulkan.cpp
+++ b/src/ggml-vulkan.cpp
@@ -137,6 +137,7 @@ struct vk_device {
     vk_pipeline pipeline_get_rows[VK_NUM_TYPES];
     vk_pipeline pipeline_get_rows_f32[VK_NUM_TYPES];
     vk_pipeline pipeline_mul_f32;
+    vk_pipeline pipeline_div_f32;
     vk_pipeline pipeline_add_f32;
     vk_pipeline pipeline_scale_f32;
     vk_pipeline pipeline_sqr_f32;
@@ -152,6 +153,7 @@ struct vk_device {
     vk_pipeline pipeline_rope_f32, pipeline_rope_f16;
     vk_pipeline pipeline_rope_neox_f32, pipeline_rope_neox_f16;
     vk_pipeline pipeline_argsort_f32;
+    vk_pipeline pipeline_sum_rows_f32;
 
     std::vector<vk_pipeline_ref> pipelines;
 
@@ -226,17 +228,27 @@ typedef std::vector<vk_submission> vk_sequence;
 
 struct vk_mat_mat_push_constants {
     uint32_t M; uint32_t N; uint32_t K;
-    uint32_t stride_a; uint32_t stride_b; uint32_t stride_d; uint32_t k_split;
-    uint32_t ne02; uint32_t ne12; uint32_t broadcast2; uint32_t broadcast3;
+    uint32_t stride_a; uint32_t stride_b; uint32_t stride_d;
     uint32_t batch_stride_a; uint32_t batch_stride_b; uint32_t batch_stride_d;
-    uint32_t expert_stride_b; uint32_t expert_stride_d;
-    uint32_t idx; uint32_t nbi1; uint32_t n_as;
+    uint32_t k_split;
+    uint32_t ne02; uint32_t ne12; uint32_t broadcast2; uint32_t broadcast3;
 };
-
 struct vk_mat_vec_push_constants {
     uint32_t ncols; uint32_t stride_a; uint32_t stride_b; uint32_t stride_d;
+    uint32_t batch_stride_a; uint32_t batch_stride_b; uint32_t batch_stride_d;
     uint32_t ne02; uint32_t ne12; uint32_t broadcast2; uint32_t broadcast3;
+};
+
+struct vk_mat_mat_id_push_constants {
+    uint32_t M; uint32_t N; uint32_t K;
+    uint32_t stride_a; uint32_t stride_b; uint32_t stride_d;
+    uint32_t batch_stride_a; uint32_t batch_stride_b; uint32_t batch_stride_d;
+    uint32_t nei0; uint32_t nei1; uint32_t nbi1; uint32_t ne11;
+};
+struct vk_mat_vec_id_push_constants {
+    uint32_t ncols; uint32_t stride_a; uint32_t stride_b; uint32_t stride_d;
     uint32_t batch_stride_a; uint32_t batch_stride_b; uint32_t batch_stride_d;
+    uint32_t nei0; uint32_t ne11;
 };
 
 struct vk_op_push_constants {
@@ -1028,7 +1040,7 @@ static void ggml_vk_load_shaders(ggml_backend_vk_context * ctx) {
     ctx->device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q5_K] = std::make_shared<vk_matmul_pipeline_struct>();
     ctx->device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q6_K] = std::make_shared<vk_matmul_pipeline_struct>();
 
-    /*ctx->device->pipeline_matmul_id_f32 = std::make_shared<vk_matmul_pipeline_struct>();
+    ctx->device->pipeline_matmul_id_f32 = std::make_shared<vk_matmul_pipeline_struct>();
     ctx->device->pipeline_matmul_id_f16_f32 = std::make_shared<vk_matmul_pipeline_struct>();
     ctx->device->pipeline_matmul_id_f16 = std::make_shared<vk_matmul_pipeline_struct>();
     ctx->device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_0] = std::make_shared<vk_matmul_pipeline_struct>();
@@ -1040,7 +1052,7 @@ static void ggml_vk_load_shaders(ggml_backend_vk_context * ctx) {
     ctx->device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q3_K] = std::make_shared<vk_matmul_pipeline_struct>();
     ctx->device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_K] = std::make_shared<vk_matmul_pipeline_struct>();
     ctx->device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q5_K] = std::make_shared<vk_matmul_pipeline_struct>();
-    ctx->device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q6_K] = std::make_shared<vk_matmul_pipeline_struct>();*/
+    ctx->device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q6_K] = std::make_shared<vk_matmul_pipeline_struct>();
 
     if (device->fp16) {
         ggml_vk_create_pipeline(ctx, ctx->device->pipeline_matmul_f32->l, "matmul_f32_l", matmul_f32_len, matmul_f32_data, "main", 3, sizeof(vk_mat_mat_push_constants), l_wg_denoms, warptile_l, 1);
@@ -1078,12 +1090,12 @@ static void ggml_vk_load_shaders(ggml_backend_vk_context * ctx) {
         ggml_vk_create_pipeline(ctx, ctx->device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q4_0]->a_m, "matmul_q4_0_f32_aligned_m", matmul_q4_0_f32_aligned_len, matmul_q4_0_f32_aligned_data, "main", 3, sizeof(vk_mat_mat_push_constants), m_wg_denoms, warptile_mmq_m, m_align);
         ggml_vk_create_pipeline(ctx, ctx->device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q4_0]->a_s, "matmul_q4_0_f32_aligned_s", matmul_q4_0_f32_aligned_len, matmul_q4_0_f32_aligned_data, "main", 3, sizeof(vk_mat_mat_push_constants), s_wg_denoms, warptile_mmq_s, s_align);
 
-        ggml_vk_create_pipeline(ctx, ctx->device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q4_1]->l, "matmul_q4_0_f32_l", matmul_q4_1_f32_len, matmul_q4_1_f32_data, "main", 3, sizeof(vk_mat_mat_push_constants), l_wg_denoms, warptile_mmq_l, l_align);
-        ggml_vk_create_pipeline(ctx, ctx->device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q4_1]->m, "matmul_q4_0_f32_m", matmul_q4_1_f32_len, matmul_q4_1_f32_data, "main", 3, sizeof(vk_mat_mat_push_constants), m_wg_denoms, warptile_mmq_m, m_align);
-        ggml_vk_create_pipeline(ctx, ctx->device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q4_1]->s, "matmul_q4_0_f32_s", matmul_q4_1_f32_len, matmul_q4_1_f32_data, "main", 3, sizeof(vk_mat_mat_push_constants), s_wg_denoms, warptile_mmq_s, s_align);
-        ggml_vk_create_pipeline(ctx, ctx->device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q4_1]->a_l, "matmul_q4_0_f32_aligned_l", matmul_q4_1_f32_aligned_len, matmul_q4_1_f32_aligned_data, "main", 3, sizeof(vk_mat_mat_push_constants), l_wg_denoms, warptile_mmq_l, l_align);
-        ggml_vk_create_pipeline(ctx, ctx->device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q4_1]->a_m, "matmul_q4_0_f32_aligned_m", matmul_q4_1_f32_aligned_len, matmul_q4_1_f32_aligned_data, "main", 3, sizeof(vk_mat_mat_push_constants), m_wg_denoms, warptile_mmq_m, m_align);
-        ggml_vk_create_pipeline(ctx, ctx->device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q4_1]->a_s, "matmul_q4_0_f32_aligned_s", matmul_q4_1_f32_aligned_len, matmul_q4_1_f32_aligned_data, "main", 3, sizeof(vk_mat_mat_push_constants), s_wg_denoms, warptile_mmq_s, s_align);
+        ggml_vk_create_pipeline(ctx, ctx->device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q4_1]->l, "matmul_q4_1_f32_l", matmul_q4_1_f32_len, matmul_q4_1_f32_data, "main", 3, sizeof(vk_mat_mat_push_constants), l_wg_denoms, warptile_mmq_l, l_align);
+        ggml_vk_create_pipeline(ctx, ctx->device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q4_1]->m, "matmul_q4_1_f32_m", matmul_q4_1_f32_len, matmul_q4_1_f32_data, "main", 3, sizeof(vk_mat_mat_push_constants), m_wg_denoms, warptile_mmq_m, m_align);
+        ggml_vk_create_pipeline(ctx, ctx->device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q4_1]->s, "matmul_q4_1_f32_s", matmul_q4_1_f32_len, matmul_q4_1_f32_data, "main", 3, sizeof(vk_mat_mat_push_constants), s_wg_denoms, warptile_mmq_s, s_align);
+        ggml_vk_create_pipeline(ctx, ctx->device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q4_1]->a_l, "matmul_q4_1_f32_aligned_l", matmul_q4_1_f32_aligned_len, matmul_q4_1_f32_aligned_data, "main", 3, sizeof(vk_mat_mat_push_constants), l_wg_denoms, warptile_mmq_l, l_align);
+        ggml_vk_create_pipeline(ctx, ctx->device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q4_1]->a_m, "matmul_q4_1_f32_aligned_m", matmul_q4_1_f32_aligned_len, matmul_q4_1_f32_aligned_data, "main", 3, sizeof(vk_mat_mat_push_constants), m_wg_denoms, warptile_mmq_m, m_align);
+        ggml_vk_create_pipeline(ctx, ctx->device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q4_1]->a_s, "matmul_q4_1_f32_aligned_s", matmul_q4_1_f32_aligned_len, matmul_q4_1_f32_aligned_data, "main", 3, sizeof(vk_mat_mat_push_constants), s_wg_denoms, warptile_mmq_s, s_align);
 
         ggml_vk_create_pipeline(ctx, ctx->device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q5_0]->l, "matmul_q5_0_f32_l", matmul_q5_0_f32_len, matmul_q5_0_f32_data, "main", 3, sizeof(vk_mat_mat_push_constants), l_wg_denoms, warptile_mmq_l, l_align);
         ggml_vk_create_pipeline(ctx, ctx->device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q5_0]->m, "matmul_q5_0_f32_m", matmul_q5_0_f32_len, matmul_q5_0_f32_data, "main", 3, sizeof(vk_mat_mat_push_constants), m_wg_denoms, warptile_mmq_m, m_align);
@@ -1141,96 +1153,96 @@ static void ggml_vk_load_shaders(ggml_backend_vk_context * ctx) {
         ggml_vk_create_pipeline(ctx, ctx->device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q6_K]->a_m, "matmul_q6_k_f32_aligned_m", matmul_q6_k_f32_aligned_len, matmul_q6_k_f32_aligned_data, "main", 3, sizeof(vk_mat_mat_push_constants), m_wg_denoms, warptile_mmq_m, m_align);
         ggml_vk_create_pipeline(ctx, ctx->device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q6_K]->a_s, "matmul_q6_k_f32_aligned_s", matmul_q6_k_f32_aligned_len, matmul_q6_k_f32_aligned_data, "main", 3, sizeof(vk_mat_mat_push_constants), s_wg_denoms, warptile_mmq_s, s_align);
 
-        /*ggml_vk_create_pipeline(ctx, ctx->device->pipeline_matmul_id_f32->l, "matmul_id_f32_l", matmul_id_f32_len, matmul_id_f32_data, "main", 4, sizeof(vk_mat_mat_push_constants), l_wg_denoms, warptile_l, 1);
-        ggml_vk_create_pipeline(ctx, ctx->device->pipeline_matmul_id_f32->m, "matmul_id_f32_m", matmul_id_f32_len, matmul_id_f32_data, "main", 4, sizeof(vk_mat_mat_push_constants), m_wg_denoms, warptile_m, 1);
-        ggml_vk_create_pipeline(ctx, ctx->device->pipeline_matmul_id_f32->s, "matmul_id_f32_s", matmul_id_f32_len, matmul_id_f32_data, "main", 4, sizeof(vk_mat_mat_push_constants), s_wg_denoms, warptile_s, 1);
-        ggml_vk_create_pipeline(ctx, ctx->device->pipeline_matmul_id_f32->a_l, "matmul_id_f32_aligned_l", matmul_id_f32_aligned_len, matmul_id_f32_aligned_data, "main", 4, sizeof(vk_mat_mat_push_constants), l_wg_denoms, warptile_l, l_align);
-        ggml_vk_create_pipeline(ctx, ctx->device->pipeline_matmul_id_f32->a_m, "matmul_id_f32_aligned_m", matmul_id_f32_aligned_len, matmul_id_f32_aligned_data, "main", 4, sizeof(vk_mat_mat_push_constants), m_wg_denoms, warptile_m, m_align);
-        ggml_vk_create_pipeline(ctx, ctx->device->pipeline_matmul_id_f32->a_s, "matmul_id_f32_aligned_s", matmul_id_f32_aligned_len, matmul_id_f32_aligned_data, "main", 4, sizeof(vk_mat_mat_push_constants), s_wg_denoms, warptile_s, s_align);
-
-        ggml_vk_create_pipeline(ctx, ctx->device->pipeline_matmul_id_f16->l, "matmul_id_f16_l", matmul_id_f16_len, matmul_id_f16_data, "main", 4, sizeof(vk_mat_mat_push_constants), l_wg_denoms, warptile_l, 1);
-        ggml_vk_create_pipeline(ctx, ctx->device->pipeline_matmul_id_f16->m, "matmul_id_f16_m", matmul_id_f16_len, matmul_id_f16_data, "main", 4, sizeof(vk_mat_mat_push_constants), m_wg_denoms, warptile_m, 1);
-        ggml_vk_create_pipeline(ctx, ctx->device->pipeline_matmul_id_f16->s, "matmul_id_f16_s", matmul_id_f16_len, matmul_id_f16_data, "main", 4, sizeof(vk_mat_mat_push_constants), s_wg_denoms, warptile_s, 1);
-        ggml_vk_create_pipeline(ctx, ctx->device->pipeline_matmul_id_f16->a_l, "matmul_id_f16_aligned_l", matmul_id_f16_aligned_len, matmul_id_f16_aligned_data, "main", 4, sizeof(vk_mat_mat_push_constants), l_wg_denoms, warptile_l, l_align);
-        ggml_vk_create_pipeline(ctx, ctx->device->pipeline_matmul_id_f16->a_m, "matmul_id_f16_aligned_m", matmul_id_f16_aligned_len, matmul_id_f16_aligned_data, "main", 4, sizeof(vk_mat_mat_push_constants), m_wg_denoms, warptile_m, m_align);
-        ggml_vk_create_pipeline(ctx, ctx->device->pipeline_matmul_id_f16->a_s, "matmul_id_f16_aligned_s", matmul_id_f16_aligned_len, matmul_id_f16_aligned_data, "main", 4, sizeof(vk_mat_mat_push_constants), s_wg_denoms, warptile_s, s_align);
-
-        ggml_vk_create_pipeline(ctx, ctx->device->pipeline_matmul_id_f16_f32->l, "matmul_id_f16_f32_l", matmul_id_f16_f32_len, matmul_id_f16_f32_data, "main", 4, sizeof(vk_mat_mat_push_constants), l_wg_denoms, warptile_l, 1);
-        ggml_vk_create_pipeline(ctx, ctx->device->pipeline_matmul_id_f16_f32->m, "matmul_id_f16_f32_m", matmul_id_f16_f32_len, matmul_id_f16_f32_data, "main", 4, sizeof(vk_mat_mat_push_constants), m_wg_denoms, warptile_m, 1);
-        ggml_vk_create_pipeline(ctx, ctx->device->pipeline_matmul_id_f16_f32->s, "matmul_id_f16_f32_s", matmul_id_f16_f32_len, matmul_id_f16_f32_data, "main", 4, sizeof(vk_mat_mat_push_constants), s_wg_denoms, warptile_s, 1);
-        ggml_vk_create_pipeline(ctx, ctx->device->pipeline_matmul_id_f16_f32->a_l, "matmul_id_f16_f32_aligned_l", matmul_id_f16_f32_aligned_len, matmul_id_f16_f32_aligned_data, "main", 4, sizeof(vk_mat_mat_push_constants), l_wg_denoms, warptile_l, l_align);
-        ggml_vk_create_pipeline(ctx, ctx->device->pipeline_matmul_id_f16_f32->a_m, "matmul_id_f16_f32_aligned_m", matmul_id_f16_f32_aligned_len, matmul_id_f16_f32_aligned_data, "main", 4, sizeof(vk_mat_mat_push_constants), m_wg_denoms, warptile_m, m_align);
-        ggml_vk_create_pipeline(ctx, ctx->device->pipeline_matmul_id_f16_f32->a_s, "matmul_id_f16_f32_aligned_s", matmul_id_f16_f32_aligned_len, matmul_id_f16_f32_aligned_data, "main", 4, sizeof(vk_mat_mat_push_constants), s_wg_denoms, warptile_s, s_align);
-
-        ggml_vk_create_pipeline(ctx, ctx->device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_0]->l, "matmul_id_q4_0_f32_l", matmul_id_q4_0_f32_len, matmul_id_q4_0_f32_data, "main", 4, sizeof(vk_mat_mat_push_constants), l_wg_denoms, warptile_mmq_l, l_align);
-        ggml_vk_create_pipeline(ctx, ctx->device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_0]->m, "matmul_id_q4_0_f32_m", matmul_id_q4_0_f32_len, matmul_id_q4_0_f32_data, "main", 4, sizeof(vk_mat_mat_push_constants), m_wg_denoms, warptile_mmq_m, m_align);
-        ggml_vk_create_pipeline(ctx, ctx->device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_0]->s, "matmul_id_q4_0_f32_s", matmul_id_q4_0_f32_len, matmul_id_q4_0_f32_data, "main", 4, sizeof(vk_mat_mat_push_constants), s_wg_denoms, warptile_mmq_s, s_align);
-        ggml_vk_create_pipeline(ctx, ctx->device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_0]->a_l, "matmul_id_q4_0_f32_aligned_l", matmul_id_q4_0_f32_aligned_len, matmul_id_q4_0_f32_aligned_data, "main", 4, sizeof(vk_mat_mat_push_constants), l_wg_denoms, warptile_mmq_l, l_align);
-        ggml_vk_create_pipeline(ctx, ctx->device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_0]->a_m, "matmul_id_q4_0_f32_aligned_m", matmul_id_q4_0_f32_aligned_len, matmul_id_q4_0_f32_aligned_data, "main", 4, sizeof(vk_mat_mat_push_constants), m_wg_denoms, warptile_mmq_m, m_align);
-        ggml_vk_create_pipeline(ctx, ctx->device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_0]->a_s, "matmul_id_q4_0_f32_aligned_s", matmul_id_q4_0_f32_aligned_len, matmul_id_q4_0_f32_aligned_data, "main", 4, sizeof(vk_mat_mat_push_constants), s_wg_denoms, warptile_mmq_s, s_align);
-
-        ggml_vk_create_pipeline(ctx, ctx->device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_1]->l, "matmul_id_q4_0_f32_l", matmul_id_q4_1_f32_len, matmul_id_q4_1_f32_data, "main", 4, sizeof(vk_mat_mat_push_constants), l_wg_denoms, warptile_mmq_l, l_align);
-        ggml_vk_create_pipeline(ctx, ctx->device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_1]->m, "matmul_id_q4_0_f32_m", matmul_id_q4_1_f32_len, matmul_id_q4_1_f32_data, "main", 4, sizeof(vk_mat_mat_push_constants), m_wg_denoms, warptile_mmq_m, m_align);
-        ggml_vk_create_pipeline(ctx, ctx->device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_1]->s, "matmul_id_q4_0_f32_s", matmul_id_q4_1_f32_len, matmul_id_q4_1_f32_data, "main", 4, sizeof(vk_mat_mat_push_constants), s_wg_denoms, warptile_mmq_s, s_align);
-        ggml_vk_create_pipeline(ctx, ctx->device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_1]->a_l, "matmul_id_q4_0_f32_aligned_l", matmul_id_q4_1_f32_aligned_len, matmul_id_q4_1_f32_aligned_data, "main", 4, sizeof(vk_mat_mat_push_constants), l_wg_denoms, warptile_mmq_l, l_align);
-        ggml_vk_create_pipeline(ctx, ctx->device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_1]->a_m, "matmul_id_q4_0_f32_aligned_m", matmul_id_q4_1_f32_aligned_len, matmul_id_q4_1_f32_aligned_data, "main", 4, sizeof(vk_mat_mat_push_constants), m_wg_denoms, warptile_mmq_m, m_align);
-        ggml_vk_create_pipeline(ctx, ctx->device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_1]->a_s, "matmul_id_q4_0_f32_aligned_s", matmul_id_q4_1_f32_aligned_len, matmul_id_q4_1_f32_aligned_data, "main", 4, sizeof(vk_mat_mat_push_constants), s_wg_denoms, warptile_mmq_s, s_align);
-
-        ggml_vk_create_pipeline(ctx, ctx->device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q5_0]->l, "matmul_id_q5_0_f32_l", matmul_id_q5_0_f32_len, matmul_id_q5_0_f32_data, "main", 4, sizeof(vk_mat_mat_push_constants), l_wg_denoms, warptile_mmq_l, l_align);
-        ggml_vk_create_pipeline(ctx, ctx->device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q5_0]->m, "matmul_id_q5_0_f32_m", matmul_id_q5_0_f32_len, matmul_id_q5_0_f32_data, "main", 4, sizeof(vk_mat_mat_push_constants), m_wg_denoms, warptile_mmq_m, m_align);
-        ggml_vk_create_pipeline(ctx, ctx->device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q5_0]->s, "matmul_id_q5_0_f32_s", matmul_id_q5_0_f32_len, matmul_id_q5_0_f32_data, "main", 4, sizeof(vk_mat_mat_push_constants), s_wg_denoms, warptile_mmq_s, s_align);
-        ggml_vk_create_pipeline(ctx, ctx->device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q5_0]->a_l, "matmul_id_q5_0_f32_aligned_l", matmul_id_q5_0_f32_aligned_len, matmul_id_q5_0_f32_aligned_data, "main", 4, sizeof(vk_mat_mat_push_constants), l_wg_denoms, warptile_mmq_l, l_align);
-        ggml_vk_create_pipeline(ctx, ctx->device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q5_0]->a_m, "matmul_id_q5_0_f32_aligned_m", matmul_id_q5_0_f32_aligned_len, matmul_id_q5_0_f32_aligned_data, "main", 4, sizeof(vk_mat_mat_push_constants), m_wg_denoms, warptile_mmq_m, m_align);
-        ggml_vk_create_pipeline(ctx, ctx->device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q5_0]->a_s, "matmul_id_q5_0_f32_aligned_s", matmul_id_q5_0_f32_aligned_len, matmul_id_q5_0_f32_aligned_data, "main", 4, sizeof(vk_mat_mat_push_constants), s_wg_denoms, warptile_mmq_s, s_align);
-
-        ggml_vk_create_pipeline(ctx, ctx->device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q5_1]->l, "matmul_id_q5_1_f32_l", matmul_id_q5_1_f32_len, matmul_id_q5_1_f32_data, "main", 4, sizeof(vk_mat_mat_push_constants), l_wg_denoms, warptile_mmq_l, l_align);
-        ggml_vk_create_pipeline(ctx, ctx->device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q5_1]->m, "matmul_id_q5_1_f32_m", matmul_id_q5_1_f32_len, matmul_id_q5_1_f32_data, "main", 4, sizeof(vk_mat_mat_push_constants), m_wg_denoms, warptile_mmq_m, m_align);
-        ggml_vk_create_pipeline(ctx, ctx->device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q5_1]->s, "matmul_id_q5_1_f32_s", matmul_id_q5_1_f32_len, matmul_id_q5_1_f32_data, "main", 4, sizeof(vk_mat_mat_push_constants), s_wg_denoms, warptile_mmq_s, s_align);
-        ggml_vk_create_pipeline(ctx, ctx->device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q5_1]->a_l, "matmul_id_q5_1_f32_aligned_l", matmul_id_q5_1_f32_aligned_len, matmul_id_q5_1_f32_aligned_data, "main", 4, sizeof(vk_mat_mat_push_constants), l_wg_denoms, warptile_mmq_l, l_align);
-        ggml_vk_create_pipeline(ctx, ctx->device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q5_1]->a_m, "matmul_id_q5_1_f32_aligned_m", matmul_id_q5_1_f32_aligned_len, matmul_id_q5_1_f32_aligned_data, "main", 4, sizeof(vk_mat_mat_push_constants), m_wg_denoms, warptile_mmq_m, m_align);
-        ggml_vk_create_pipeline(ctx, ctx->device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q5_1]->a_s, "matmul_id_q5_1_f32_aligned_s", matmul_id_q5_1_f32_aligned_len, matmul_id_q5_1_f32_aligned_data, "main", 4, sizeof(vk_mat_mat_push_constants), s_wg_denoms, warptile_mmq_s, s_align);
-
-        ggml_vk_create_pipeline(ctx, ctx->device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q8_0]->l, "matmul_id_q8_0_f32_l", matmul_id_q8_0_f32_len, matmul_id_q8_0_f32_data, "main", 4, sizeof(vk_mat_mat_push_constants), l_wg_denoms, warptile_mmq_l, l_align);
-        ggml_vk_create_pipeline(ctx, ctx->device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q8_0]->m, "matmul_id_q8_0_f32_m", matmul_id_q8_0_f32_len, matmul_id_q8_0_f32_data, "main", 4, sizeof(vk_mat_mat_push_constants), m_wg_denoms, warptile_mmq_m, m_align);
-        ggml_vk_create_pipeline(ctx, ctx->device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q8_0]->s, "matmul_id_q8_0_f32_s", matmul_id_q8_0_f32_len, matmul_id_q8_0_f32_data, "main", 4, sizeof(vk_mat_mat_push_constants), s_wg_denoms, warptile_mmq_s, s_align);
-        ggml_vk_create_pipeline(ctx, ctx->device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q8_0]->a_l, "matmul_id_q8_0_f32_aligned_l", matmul_id_q8_0_f32_aligned_len, matmul_id_q8_0_f32_aligned_data, "main", 4, sizeof(vk_mat_mat_push_constants), l_wg_denoms, warptile_mmq_l, l_align);
-        ggml_vk_create_pipeline(ctx, ctx->device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q8_0]->a_m, "matmul_id_q8_0_f32_aligned_m", matmul_id_q8_0_f32_aligned_len, matmul_id_q8_0_f32_aligned_data, "main", 4, sizeof(vk_mat_mat_push_constants), m_wg_denoms, warptile_mmq_m, m_align);
-        ggml_vk_create_pipeline(ctx, ctx->device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q8_0]->a_s, "matmul_id_q8_0_f32_aligned_s", matmul_id_q8_0_f32_aligned_len, matmul_id_q8_0_f32_aligned_data, "main", 4, sizeof(vk_mat_mat_push_constants), s_wg_denoms, warptile_mmq_s, s_align);
-
-        ggml_vk_create_pipeline(ctx, ctx->device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q2_K]->l, "matmul_id_q2_k_f32_l", matmul_id_q2_k_f32_len, matmul_id_q2_k_f32_data, "main", 4, sizeof(vk_mat_mat_push_constants), l_wg_denoms, warptile_mmq_l, l_align);
-        ggml_vk_create_pipeline(ctx, ctx->device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q2_K]->m, "matmul_id_q2_k_f32_m", matmul_id_q2_k_f32_len, matmul_id_q2_k_f32_data, "main", 4, sizeof(vk_mat_mat_push_constants), m_wg_denoms, warptile_mmq_m, m_align);
-        ggml_vk_create_pipeline(ctx, ctx->device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q2_K]->s, "matmul_id_q2_k_f32_s", matmul_id_q2_k_f32_len, matmul_id_q2_k_f32_data, "main", 4, sizeof(vk_mat_mat_push_constants), s_wg_denoms, warptile_mmq_s, s_align);
-        ggml_vk_create_pipeline(ctx, ctx->device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q2_K]->a_l, "matmul_id_q2_k_f32_aligned_l", matmul_id_q2_k_f32_aligned_len, matmul_id_q2_k_f32_aligned_data, "main", 4, sizeof(vk_mat_mat_push_constants), l_wg_denoms, warptile_mmq_l, l_align);
-        ggml_vk_create_pipeline(ctx, ctx->device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q2_K]->a_m, "matmul_id_q2_k_f32_aligned_m", matmul_id_q2_k_f32_aligned_len, matmul_id_q2_k_f32_aligned_data, "main", 4, sizeof(vk_mat_mat_push_constants), m_wg_denoms, warptile_mmq_m, m_align);
-        ggml_vk_create_pipeline(ctx, ctx->device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q2_K]->a_s, "matmul_id_q2_k_f32_aligned_s", matmul_id_q2_k_f32_aligned_len, matmul_id_q2_k_f32_aligned_data, "main", 4, sizeof(vk_mat_mat_push_constants), s_wg_denoms, warptile_mmq_s, s_align);
-
-        ggml_vk_create_pipeline(ctx, ctx->device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q3_K]->l, "matmul_id_q3_k_f32_l", matmul_id_q3_k_f32_len, matmul_id_q3_k_f32_data, "main", 4, sizeof(vk_mat_mat_push_constants), l_wg_denoms, warptile_mmq_l, l_align);
-        ggml_vk_create_pipeline(ctx, ctx->device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q3_K]->m, "matmul_id_q3_k_f32_m", matmul_id_q3_k_f32_len, matmul_id_q3_k_f32_data, "main", 4, sizeof(vk_mat_mat_push_constants), m_wg_denoms, warptile_mmq_m, m_align);
-        ggml_vk_create_pipeline(ctx, ctx->device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q3_K]->s, "matmul_id_q3_k_f32_s", matmul_id_q3_k_f32_len, matmul_id_q3_k_f32_data, "main", 4, sizeof(vk_mat_mat_push_constants), s_wg_denoms, warptile_mmq_s, s_align);
-        ggml_vk_create_pipeline(ctx, ctx->device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q3_K]->a_l, "matmul_id_q3_k_f32_aligned_l", matmul_id_q3_k_f32_aligned_len, matmul_id_q3_k_f32_aligned_data, "main", 4, sizeof(vk_mat_mat_push_constants), l_wg_denoms, warptile_mmq_l, l_align);
-        ggml_vk_create_pipeline(ctx, ctx->device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q3_K]->a_m, "matmul_id_q3_k_f32_aligned_m", matmul_id_q3_k_f32_aligned_len, matmul_id_q3_k_f32_aligned_data, "main", 4, sizeof(vk_mat_mat_push_constants), m_wg_denoms, warptile_mmq_m, m_align);
-        ggml_vk_create_pipeline(ctx, ctx->device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q3_K]->a_s, "matmul_id_q3_k_f32_aligned_s", matmul_id_q3_k_f32_aligned_len, matmul_id_q3_k_f32_aligned_data, "main", 4, sizeof(vk_mat_mat_push_constants), s_wg_denoms, warptile_mmq_s, s_align);
-
-        ggml_vk_create_pipeline(ctx, ctx->device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_K]->l, "matmul_id_q4_k_f32_l", matmul_id_q4_k_f32_len, matmul_id_q4_k_f32_data, "main", 4, sizeof(vk_mat_mat_push_constants), l_wg_denoms, warptile_mmq_l, l_align);
-        ggml_vk_create_pipeline(ctx, ctx->device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_K]->m, "matmul_id_q4_k_f32_m", matmul_id_q4_k_f32_len, matmul_id_q4_k_f32_data, "main", 4, sizeof(vk_mat_mat_push_constants), m_wg_denoms, warptile_mmq_m, m_align);
-        ggml_vk_create_pipeline(ctx, ctx->device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_K]->s, "matmul_id_q4_k_f32_s", matmul_id_q4_k_f32_len, matmul_id_q4_k_f32_data, "main", 4, sizeof(vk_mat_mat_push_constants), s_wg_denoms, warptile_mmq_s, s_align);
-        ggml_vk_create_pipeline(ctx, ctx->device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_K]->a_l, "matmul_id_q4_k_f32_aligned_l", matmul_id_q4_k_f32_aligned_len, matmul_id_q4_k_f32_aligned_data, "main", 4, sizeof(vk_mat_mat_push_constants), l_wg_denoms, warptile_mmq_l, l_align);
-        ggml_vk_create_pipeline(ctx, ctx->device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_K]->a_m, "matmul_id_q4_k_f32_aligned_m", matmul_id_q4_k_f32_aligned_len, matmul_id_q4_k_f32_aligned_data, "main", 4, sizeof(vk_mat_mat_push_constants), m_wg_denoms, warptile_mmq_m, m_align);
-        ggml_vk_create_pipeline(ctx, ctx->device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_K]->a_s, "matmul_id_q4_k_f32_aligned_s", matmul_id_q4_k_f32_aligned_len, matmul_id_q4_k_f32_aligned_data, "main", 4, sizeof(vk_mat_mat_push_constants), s_wg_denoms, warptile_mmq_s, s_align);
-
-        ggml_vk_create_pipeline(ctx, ctx->device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q5_K]->l, "matmul_id_q5_k_f32_l", matmul_id_q5_k_f32_len, matmul_id_q5_k_f32_data, "main", 4, sizeof(vk_mat_mat_push_constants), l_wg_denoms, warptile_mmq_l, l_align);
-        ggml_vk_create_pipeline(ctx, ctx->device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q5_K]->m, "matmul_id_q5_k_f32_m", matmul_id_q5_k_f32_len, matmul_id_q5_k_f32_data, "main", 4, sizeof(vk_mat_mat_push_constants), m_wg_denoms, warptile_mmq_m, m_align);
-        ggml_vk_create_pipeline(ctx, ctx->device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q5_K]->s, "matmul_id_q5_k_f32_s", matmul_id_q5_k_f32_len, matmul_id_q5_k_f32_data, "main", 4, sizeof(vk_mat_mat_push_constants), s_wg_denoms, warptile_mmq_s, s_align);
-        ggml_vk_create_pipeline(ctx, ctx->device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q5_K]->a_l, "matmul_id_q5_k_f32_aligned_l", matmul_id_q5_k_f32_aligned_len, matmul_id_q5_k_f32_aligned_data, "main", 4, sizeof(vk_mat_mat_push_constants), l_wg_denoms, warptile_mmq_l, l_align);
-        ggml_vk_create_pipeline(ctx, ctx->device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q5_K]->a_m, "matmul_id_q5_k_f32_aligned_m", matmul_id_q5_k_f32_aligned_len, matmul_id_q5_k_f32_aligned_data, "main", 4, sizeof(vk_mat_mat_push_constants), m_wg_denoms, warptile_mmq_m, m_align);
-        ggml_vk_create_pipeline(ctx, ctx->device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q5_K]->a_s, "matmul_id_q5_k_f32_aligned_s", matmul_id_q5_k_f32_aligned_len, matmul_id_q5_k_f32_aligned_data, "main", 4, sizeof(vk_mat_mat_push_constants), s_wg_denoms, warptile_mmq_s, s_align);
-
-        ggml_vk_create_pipeline(ctx, ctx->device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q6_K]->l, "matmul_id_q6_k_f32_l", matmul_id_q6_k_f32_len, matmul_id_q6_k_f32_data, "main", 4, sizeof(vk_mat_mat_push_constants), l_wg_denoms, warptile_mmq_l, l_align);
-        ggml_vk_create_pipeline(ctx, ctx->device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q6_K]->m, "matmul_id_q6_k_f32_m", matmul_id_q6_k_f32_len, matmul_id_q6_k_f32_data, "main", 4, sizeof(vk_mat_mat_push_constants), m_wg_denoms, warptile_mmq_m, m_align);
-        ggml_vk_create_pipeline(ctx, ctx->device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q6_K]->s, "matmul_id_q6_k_f32_s", matmul_id_q6_k_f32_len, matmul_id_q6_k_f32_data, "main", 4, sizeof(vk_mat_mat_push_constants), s_wg_denoms, warptile_mmq_s, s_align);
-        ggml_vk_create_pipeline(ctx, ctx->device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q6_K]->a_l, "matmul_id_q6_k_f32_aligned_l", matmul_id_q6_k_f32_aligned_len, matmul_id_q6_k_f32_aligned_data, "main", 4, sizeof(vk_mat_mat_push_constants), l_wg_denoms, warptile_mmq_l, l_align);
-        ggml_vk_create_pipeline(ctx, ctx->device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q6_K]->a_m, "matmul_id_q6_k_f32_aligned_m", matmul_id_q6_k_f32_aligned_len, matmul_id_q6_k_f32_aligned_data, "main", 4, sizeof(vk_mat_mat_push_constants), m_wg_denoms, warptile_mmq_m, m_align);
-        ggml_vk_create_pipeline(ctx, ctx->device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q6_K]->a_s, "matmul_id_q6_k_f32_aligned_s", matmul_id_q6_k_f32_aligned_len, matmul_id_q6_k_f32_aligned_data, "main", 4, sizeof(vk_mat_mat_push_constants), s_wg_denoms, warptile_mmq_s, s_align);*/
+        ggml_vk_create_pipeline(ctx, ctx->device->pipeline_matmul_id_f32->l, "matmul_id_f32_l", matmul_id_f32_len, matmul_id_f32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), l_wg_denoms, warptile_l, 1);
+        ggml_vk_create_pipeline(ctx, ctx->device->pipeline_matmul_id_f32->m, "matmul_id_f32_m", matmul_id_f32_len, matmul_id_f32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), m_wg_denoms, warptile_m, 1);
+        ggml_vk_create_pipeline(ctx, ctx->device->pipeline_matmul_id_f32->s, "matmul_id_f32_s", matmul_id_f32_len, matmul_id_f32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), s_wg_denoms, warptile_s, 1);
+        ggml_vk_create_pipeline(ctx, ctx->device->pipeline_matmul_id_f32->a_l, "matmul_id_f32_aligned_l", matmul_id_f32_aligned_len, matmul_id_f32_aligned_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), l_wg_denoms, warptile_l, l_align);
+        ggml_vk_create_pipeline(ctx, ctx->device->pipeline_matmul_id_f32->a_m, "matmul_id_f32_aligned_m", matmul_id_f32_aligned_len, matmul_id_f32_aligned_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), m_wg_denoms, warptile_m, m_align);
+        ggml_vk_create_pipeline(ctx, ctx->device->pipeline_matmul_id_f32->a_s, "matmul_id_f32_aligned_s", matmul_id_f32_aligned_len, matmul_id_f32_aligned_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), s_wg_denoms, warptile_s, s_align);
+
+        ggml_vk_create_pipeline(ctx, ctx->device->pipeline_matmul_id_f16->l, "matmul_id_f16_l", matmul_id_f16_len, matmul_id_f16_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), l_wg_denoms, warptile_l, 1);
+        ggml_vk_create_pipeline(ctx, ctx->device->pipeline_matmul_id_f16->m, "matmul_id_f16_m", matmul_id_f16_len, matmul_id_f16_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), m_wg_denoms, warptile_m, 1);
+        ggml_vk_create_pipeline(ctx, ctx->device->pipeline_matmul_id_f16->s, "matmul_id_f16_s", matmul_id_f16_len, matmul_id_f16_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), s_wg_denoms, warptile_s, 1);
+        ggml_vk_create_pipeline(ctx, ctx->device->pipeline_matmul_id_f16->a_l, "matmul_id_f16_aligned_l", matmul_id_f16_aligned_len, matmul_id_f16_aligned_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), l_wg_denoms, warptile_l, l_align);
+        ggml_vk_create_pipeline(ctx, ctx->device->pipeline_matmul_id_f16->a_m, "matmul_id_f16_aligned_m", matmul_id_f16_aligned_len, matmul_id_f16_aligned_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), m_wg_denoms, warptile_m, m_align);
+        ggml_vk_create_pipeline(ctx, ctx->device->pipeline_matmul_id_f16->a_s, "matmul_id_f16_aligned_s", matmul_id_f16_aligned_len, matmul_id_f16_aligned_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), s_wg_denoms, warptile_s, s_align);
+
+        ggml_vk_create_pipeline(ctx, ctx->device->pipeline_matmul_id_f16_f32->l, "matmul_id_f16_f32_l", matmul_id_f16_f32_len, matmul_id_f16_f32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), l_wg_denoms, warptile_l, 1);
+        ggml_vk_create_pipeline(ctx, ctx->device->pipeline_matmul_id_f16_f32->m, "matmul_id_f16_f32_m", matmul_id_f16_f32_len, matmul_id_f16_f32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), m_wg_denoms, warptile_m, 1);
+        ggml_vk_create_pipeline(ctx, ctx->device->pipeline_matmul_id_f16_f32->s, "matmul_id_f16_f32_s", matmul_id_f16_f32_len, matmul_id_f16_f32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), s_wg_denoms, warptile_s, 1);
+        ggml_vk_create_pipeline(ctx, ctx->device->pipeline_matmul_id_f16_f32->a_l, "matmul_id_f16_f32_aligned_l", matmul_id_f16_f32_aligned_len, matmul_id_f16_f32_aligned_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), l_wg_denoms, warptile_l, l_align);
+        ggml_vk_create_pipeline(ctx, ctx->device->pipeline_matmul_id_f16_f32->a_m, "matmul_id_f16_f32_aligned_m", matmul_id_f16_f32_aligned_len, matmul_id_f16_f32_aligned_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), m_wg_denoms, warptile_m, m_align);
+        ggml_vk_create_pipeline(ctx, ctx->device->pipeline_matmul_id_f16_f32->a_s, "matmul_id_f16_f32_aligned_s", matmul_id_f16_f32_aligned_len, matmul_id_f16_f32_aligned_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), s_wg_denoms, warptile_s, s_align);
+
+        ggml_vk_create_pipeline(ctx, ctx->device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_0]->l, "matmul_id_q4_0_f32_l", matmul_id_q4_0_f32_len, matmul_id_q4_0_f32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), l_wg_denoms, warptile_mmq_l, l_align);
+        ggml_vk_create_pipeline(ctx, ctx->device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_0]->m, "matmul_id_q4_0_f32_m", matmul_id_q4_0_f32_len, matmul_id_q4_0_f32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), m_wg_denoms, warptile_mmq_m, m_align);
+        ggml_vk_create_pipeline(ctx, ctx->device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_0]->s, "matmul_id_q4_0_f32_s", matmul_id_q4_0_f32_len, matmul_id_q4_0_f32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), s_wg_denoms, warptile_mmq_s, s_align);
+        ggml_vk_create_pipeline(ctx, ctx->device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_0]->a_l, "matmul_id_q4_0_f32_aligned_l", matmul_id_q4_0_f32_aligned_len, matmul_id_q4_0_f32_aligned_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), l_wg_denoms, warptile_mmq_l, l_align);
+        ggml_vk_create_pipeline(ctx, ctx->device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_0]->a_m, "matmul_id_q4_0_f32_aligned_m", matmul_id_q4_0_f32_aligned_len, matmul_id_q4_0_f32_aligned_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), m_wg_denoms, warptile_mmq_m, m_align);
+        ggml_vk_create_pipeline(ctx, ctx->device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_0]->a_s, "matmul_id_q4_0_f32_aligned_s", matmul_id_q4_0_f32_aligned_len, matmul_id_q4_0_f32_aligned_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), s_wg_denoms, warptile_mmq_s, s_align);
+
+        ggml_vk_create_pipeline(ctx, ctx->device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_1]->l, "matmul_id_q4_1_f32_l", matmul_id_q4_1_f32_len, matmul_id_q4_1_f32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), l_wg_denoms, warptile_mmq_l, l_align);
+        ggml_vk_create_pipeline(ctx, ctx->device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_1]->m, "matmul_id_q4_1_f32_m", matmul_id_q4_1_f32_len, matmul_id_q4_1_f32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), m_wg_denoms, warptile_mmq_m, m_align);
+        ggml_vk_create_pipeline(ctx, ctx->device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_1]->s, "matmul_id_q4_1_f32_s", matmul_id_q4_1_f32_len, matmul_id_q4_1_f32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), s_wg_denoms, warptile_mmq_s, s_align);
+        ggml_vk_create_pipeline(ctx, ctx->device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_1]->a_l, "matmul_id_q4_1_f32_aligned_l", matmul_id_q4_1_f32_aligned_len, matmul_id_q4_1_f32_aligned_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), l_wg_denoms, warptile_mmq_l, l_align);
+        ggml_vk_create_pipeline(ctx, ctx->device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_1]->a_m, "matmul_id_q4_1_f32_aligned_m", matmul_id_q4_1_f32_aligned_len, matmul_id_q4_1_f32_aligned_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), m_wg_denoms, warptile_mmq_m, m_align);
+        ggml_vk_create_pipeline(ctx, ctx->device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_1]->a_s, "matmul_id_q4_1_f32_aligned_s", matmul_id_q4_1_f32_aligned_len, matmul_id_q4_1_f32_aligned_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), s_wg_denoms, warptile_mmq_s, s_align);
+
+        ggml_vk_create_pipeline(ctx, ctx->device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q5_0]->l, "matmul_id_q5_0_f32_l", matmul_id_q5_0_f32_len, matmul_id_q5_0_f32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), l_wg_denoms, warptile_mmq_l, l_align);
+        ggml_vk_create_pipeline(ctx, ctx->device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q5_0]->m, "matmul_id_q5_0_f32_m", matmul_id_q5_0_f32_len, matmul_id_q5_0_f32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), m_wg_denoms, warptile_mmq_m, m_align);
+        ggml_vk_create_pipeline(ctx, ctx->device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q5_0]->s, "matmul_id_q5_0_f32_s", matmul_id_q5_0_f32_len, matmul_id_q5_0_f32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), s_wg_denoms, warptile_mmq_s, s_align);
+        ggml_vk_create_pipeline(ctx, ctx->device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q5_0]->a_l, "matmul_id_q5_0_f32_aligned_l", matmul_id_q5_0_f32_aligned_len, matmul_id_q5_0_f32_aligned_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), l_wg_denoms, warptile_mmq_l, l_align);
+        ggml_vk_create_pipeline(ctx, ctx->device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q5_0]->a_m, "matmul_id_q5_0_f32_aligned_m", matmul_id_q5_0_f32_aligned_len, matmul_id_q5_0_f32_aligned_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), m_wg_denoms, warptile_mmq_m, m_align);
+        ggml_vk_create_pipeline(ctx, ctx->device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q5_0]->a_s, "matmul_id_q5_0_f32_aligned_s", matmul_id_q5_0_f32_aligned_len, matmul_id_q5_0_f32_aligned_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), s_wg_denoms, warptile_mmq_s, s_align);
+
+        ggml_vk_create_pipeline(ctx, ctx->device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q5_1]->l, "matmul_id_q5_1_f32_l", matmul_id_q5_1_f32_len, matmul_id_q5_1_f32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), l_wg_denoms, warptile_mmq_l, l_align);
+        ggml_vk_create_pipeline(ctx, ctx->device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q5_1]->m, "matmul_id_q5_1_f32_m", matmul_id_q5_1_f32_len, matmul_id_q5_1_f32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), m_wg_denoms, warptile_mmq_m, m_align);
+        ggml_vk_create_pipeline(ctx, ctx->device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q5_1]->s, "matmul_id_q5_1_f32_s", matmul_id_q5_1_f32_len, matmul_id_q5_1_f32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), s_wg_denoms, warptile_mmq_s, s_align);
+        ggml_vk_create_pipeline(ctx, ctx->device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q5_1]->a_l, "matmul_id_q5_1_f32_aligned_l", matmul_id_q5_1_f32_aligned_len, matmul_id_q5_1_f32_aligned_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), l_wg_denoms, warptile_mmq_l, l_align);
+        ggml_vk_create_pipeline(ctx, ctx->device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q5_1]->a_m, "matmul_id_q5_1_f32_aligned_m", matmul_id_q5_1_f32_aligned_len, matmul_id_q5_1_f32_aligned_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), m_wg_denoms, warptile_mmq_m, m_align);
+        ggml_vk_create_pipeline(ctx, ctx->device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q5_1]->a_s, "matmul_id_q5_1_f32_aligned_s", matmul_id_q5_1_f32_aligned_len, matmul_id_q5_1_f32_aligned_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), s_wg_denoms, warptile_mmq_s, s_align);
+
+        ggml_vk_create_pipeline(ctx, ctx->device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q8_0]->l, "matmul_id_q8_0_f32_l", matmul_id_q8_0_f32_len, matmul_id_q8_0_f32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), l_wg_denoms, warptile_mmq_l, l_align);
+        ggml_vk_create_pipeline(ctx, ctx->device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q8_0]->m, "matmul_id_q8_0_f32_m", matmul_id_q8_0_f32_len, matmul_id_q8_0_f32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), m_wg_denoms, warptile_mmq_m, m_align);
+        ggml_vk_create_pipeline(ctx, ctx->device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q8_0]->s, "matmul_id_q8_0_f32_s", matmul_id_q8_0_f32_len, matmul_id_q8_0_f32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), s_wg_denoms, warptile_mmq_s, s_align);
+        ggml_vk_create_pipeline(ctx, ctx->device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q8_0]->a_l, "matmul_id_q8_0_f32_aligned_l", matmul_id_q8_0_f32_aligned_len, matmul_id_q8_0_f32_aligned_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), l_wg_denoms, warptile_mmq_l, l_align);
+        ggml_vk_create_pipeline(ctx, ctx->device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q8_0]->a_m, "matmul_id_q8_0_f32_aligned_m", matmul_id_q8_0_f32_aligned_len, matmul_id_q8_0_f32_aligned_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), m_wg_denoms, warptile_mmq_m, m_align);
+        ggml_vk_create_pipeline(ctx, ctx->device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q8_0]->a_s, "matmul_id_q8_0_f32_aligned_s", matmul_id_q8_0_f32_aligned_len, matmul_id_q8_0_f32_aligned_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), s_wg_denoms, warptile_mmq_s, s_align);
+
+        ggml_vk_create_pipeline(ctx, ctx->device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q2_K]->l, "matmul_id_q2_k_f32_l", matmul_id_q2_k_f32_len, matmul_id_q2_k_f32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), l_wg_denoms, warptile_mmq_l, l_align);
+        ggml_vk_create_pipeline(ctx, ctx->device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q2_K]->m, "matmul_id_q2_k_f32_m", matmul_id_q2_k_f32_len, matmul_id_q2_k_f32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), m_wg_denoms, warptile_mmq_m, m_align);
+        ggml_vk_create_pipeline(ctx, ctx->device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q2_K]->s, "matmul_id_q2_k_f32_s", matmul_id_q2_k_f32_len, matmul_id_q2_k_f32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), s_wg_denoms, warptile_mmq_s, s_align);
+        ggml_vk_create_pipeline(ctx, ctx->device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q2_K]->a_l, "matmul_id_q2_k_f32_aligned_l", matmul_id_q2_k_f32_aligned_len, matmul_id_q2_k_f32_aligned_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), l_wg_denoms, warptile_mmq_l, l_align);
+        ggml_vk_create_pipeline(ctx, ctx->device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q2_K]->a_m, "matmul_id_q2_k_f32_aligned_m", matmul_id_q2_k_f32_aligned_len, matmul_id_q2_k_f32_aligned_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), m_wg_denoms, warptile_mmq_m, m_align);
+        ggml_vk_create_pipeline(ctx, ctx->device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q2_K]->a_s, "matmul_id_q2_k_f32_aligned_s", matmul_id_q2_k_f32_aligned_len, matmul_id_q2_k_f32_aligned_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), s_wg_denoms, warptile_mmq_s, s_align);
+
+        ggml_vk_create_pipeline(ctx, ctx->device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q3_K]->l, "matmul_id_q3_k_f32_l", matmul_id_q3_k_f32_len, matmul_id_q3_k_f32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), l_wg_denoms, warptile_mmq_l, l_align);
+        ggml_vk_create_pipeline(ctx, ctx->device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q3_K]->m, "matmul_id_q3_k_f32_m", matmul_id_q3_k_f32_len, matmul_id_q3_k_f32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), m_wg_denoms, warptile_mmq_m, m_align);
+        ggml_vk_create_pipeline(ctx, ctx->device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q3_K]->s, "matmul_id_q3_k_f32_s", matmul_id_q3_k_f32_len, matmul_id_q3_k_f32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), s_wg_denoms, warptile_mmq_s, s_align);
+        ggml_vk_create_pipeline(ctx, ctx->device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q3_K]->a_l, "matmul_id_q3_k_f32_aligned_l", matmul_id_q3_k_f32_aligned_len, matmul_id_q3_k_f32_aligned_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), l_wg_denoms, warptile_mmq_l, l_align);
+        ggml_vk_create_pipeline(ctx, ctx->device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q3_K]->a_m, "matmul_id_q3_k_f32_aligned_m", matmul_id_q3_k_f32_aligned_len, matmul_id_q3_k_f32_aligned_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), m_wg_denoms, warptile_mmq_m, m_align);
+        ggml_vk_create_pipeline(ctx, ctx->device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q3_K]->a_s, "matmul_id_q3_k_f32_aligned_s", matmul_id_q3_k_f32_aligned_len, matmul_id_q3_k_f32_aligned_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), s_wg_denoms, warptile_mmq_s, s_align);
+
+        ggml_vk_create_pipeline(ctx, ctx->device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_K]->l, "matmul_id_q4_k_f32_l", matmul_id_q4_k_f32_len, matmul_id_q4_k_f32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), l_wg_denoms, warptile_mmq_l, l_align);
+        ggml_vk_create_pipeline(ctx, ctx->device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_K]->m, "matmul_id_q4_k_f32_m", matmul_id_q4_k_f32_len, matmul_id_q4_k_f32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), m_wg_denoms, warptile_mmq_m, m_align);
+        ggml_vk_create_pipeline(ctx, ctx->device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_K]->s, "matmul_id_q4_k_f32_s", matmul_id_q4_k_f32_len, matmul_id_q4_k_f32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), s_wg_denoms, warptile_mmq_s, s_align);
+        ggml_vk_create_pipeline(ctx, ctx->device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_K]->a_l, "matmul_id_q4_k_f32_aligned_l", matmul_id_q4_k_f32_aligned_len, matmul_id_q4_k_f32_aligned_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), l_wg_denoms, warptile_mmq_l, l_align);
+        ggml_vk_create_pipeline(ctx, ctx->device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_K]->a_m, "matmul_id_q4_k_f32_aligned_m", matmul_id_q4_k_f32_aligned_len, matmul_id_q4_k_f32_aligned_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), m_wg_denoms, warptile_mmq_m, m_align);
+        ggml_vk_create_pipeline(ctx, ctx->device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_K]->a_s, "matmul_id_q4_k_f32_aligned_s", matmul_id_q4_k_f32_aligned_len, matmul_id_q4_k_f32_aligned_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), s_wg_denoms, warptile_mmq_s, s_align);
+
+        ggml_vk_create_pipeline(ctx, ctx->device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q5_K]->l, "matmul_id_q5_k_f32_l", matmul_id_q5_k_f32_len, matmul_id_q5_k_f32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), l_wg_denoms, warptile_mmq_l, l_align);
+        ggml_vk_create_pipeline(ctx, ctx->device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q5_K]->m, "matmul_id_q5_k_f32_m", matmul_id_q5_k_f32_len, matmul_id_q5_k_f32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), m_wg_denoms, warptile_mmq_m, m_align);
+        ggml_vk_create_pipeline(ctx, ctx->device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q5_K]->s, "matmul_id_q5_k_f32_s", matmul_id_q5_k_f32_len, matmul_id_q5_k_f32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), s_wg_denoms, warptile_mmq_s, s_align);
+        ggml_vk_create_pipeline(ctx, ctx->device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q5_K]->a_l, "matmul_id_q5_k_f32_aligned_l", matmul_id_q5_k_f32_aligned_len, matmul_id_q5_k_f32_aligned_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), l_wg_denoms, warptile_mmq_l, l_align);
+        ggml_vk_create_pipeline(ctx, ctx->device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q5_K]->a_m, "matmul_id_q5_k_f32_aligned_m", matmul_id_q5_k_f32_aligned_len, matmul_id_q5_k_f32_aligned_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), m_wg_denoms, warptile_mmq_m, m_align);
+        ggml_vk_create_pipeline(ctx, ctx->device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q5_K]->a_s, "matmul_id_q5_k_f32_aligned_s", matmul_id_q5_k_f32_aligned_len, matmul_id_q5_k_f32_aligned_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), s_wg_denoms, warptile_mmq_s, s_align);
+
+        ggml_vk_create_pipeline(ctx, ctx->device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q6_K]->l, "matmul_id_q6_k_f32_l", matmul_id_q6_k_f32_len, matmul_id_q6_k_f32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), l_wg_denoms, warptile_mmq_l, l_align);
+        ggml_vk_create_pipeline(ctx, ctx->device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q6_K]->m, "matmul_id_q6_k_f32_m", matmul_id_q6_k_f32_len, matmul_id_q6_k_f32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), m_wg_denoms, warptile_mmq_m, m_align);
+        ggml_vk_create_pipeline(ctx, ctx->device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q6_K]->s, "matmul_id_q6_k_f32_s", matmul_id_q6_k_f32_len, matmul_id_q6_k_f32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), s_wg_denoms, warptile_mmq_s, s_align);
+        ggml_vk_create_pipeline(ctx, ctx->device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q6_K]->a_l, "matmul_id_q6_k_f32_aligned_l", matmul_id_q6_k_f32_aligned_len, matmul_id_q6_k_f32_aligned_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), l_wg_denoms, warptile_mmq_l, l_align);
+        ggml_vk_create_pipeline(ctx, ctx->device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q6_K]->a_m, "matmul_id_q6_k_f32_aligned_m", matmul_id_q6_k_f32_aligned_len, matmul_id_q6_k_f32_aligned_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), m_wg_denoms, warptile_mmq_m, m_align);
+        ggml_vk_create_pipeline(ctx, ctx->device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q6_K]->a_s, "matmul_id_q6_k_f32_aligned_s", matmul_id_q6_k_f32_aligned_len, matmul_id_q6_k_f32_aligned_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), s_wg_denoms, warptile_mmq_s, s_align);
     } else {
         ggml_vk_create_pipeline(ctx, ctx->device->pipeline_matmul_f32->l, "matmul_f32_l", matmul_f32_fp32_len, matmul_f32_fp32_data, "main", 3, sizeof(vk_mat_mat_push_constants), l_wg_denoms, warptile_l, 1);
         ggml_vk_create_pipeline(ctx, ctx->device->pipeline_matmul_f32->m, "matmul_f32_m", matmul_f32_fp32_len, matmul_f32_fp32_data, "main", 3, sizeof(vk_mat_mat_push_constants), m_wg_denoms, warptile_m, 1);
@@ -1330,99 +1342,100 @@ static void ggml_vk_load_shaders(ggml_backend_vk_context * ctx) {
         ggml_vk_create_pipeline(ctx, ctx->device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q6_K]->a_m, "matmul_q6_k_f32_aligned_m", matmul_q6_k_f32_aligned_fp32_len, matmul_q6_k_f32_aligned_fp32_data, "main", 3, sizeof(vk_mat_mat_push_constants), m_wg_denoms, warptile_mmq_m, m_align);
         ggml_vk_create_pipeline(ctx, ctx->device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q6_K]->a_s, "matmul_q6_k_f32_aligned_s", matmul_q6_k_f32_aligned_fp32_len, matmul_q6_k_f32_aligned_fp32_data, "main", 3, sizeof(vk_mat_mat_push_constants), s_wg_denoms, warptile_mmq_s, s_align);
 
-        /*ggml_vk_create_pipeline(ctx, ctx->device->pipeline_matmul_id_f32->l, "matmul_id_f32_l", matmul_id_f32_fp32_len, matmul_id_f32_fp32_data, "main", 4, sizeof(vk_mat_mat_push_constants), l_wg_denoms, warptile_l, 1);
-        ggml_vk_create_pipeline(ctx, ctx->device->pipeline_matmul_id_f32->m, "matmul_id_f32_m", matmul_id_f32_fp32_len, matmul_id_f32_fp32_data, "main", 4, sizeof(vk_mat_mat_push_constants), m_wg_denoms, warptile_m, 1);
-        ggml_vk_create_pipeline(ctx, ctx->device->pipeline_matmul_id_f32->s, "matmul_id_f32_s", matmul_id_f32_fp32_len, matmul_id_f32_fp32_data, "main", 4, sizeof(vk_mat_mat_push_constants), s_wg_denoms, warptile_s, 1);
-        ggml_vk_create_pipeline(ctx, ctx->device->pipeline_matmul_id_f32->a_l, "matmul_id_f32_aligned_l", matmul_id_f32_aligned_fp32_len, matmul_id_f32_aligned_fp32_data, "main", 4, sizeof(vk_mat_mat_push_constants), l_wg_denoms, warptile_l, l_align);
-        ggml_vk_create_pipeline(ctx, ctx->device->pipeline_matmul_id_f32->a_m, "matmul_id_f32_aligned_m", matmul_id_f32_aligned_fp32_len, matmul_id_f32_aligned_fp32_data, "main", 4, sizeof(vk_mat_mat_push_constants), m_wg_denoms, warptile_m, m_align);
-        ggml_vk_create_pipeline(ctx, ctx->device->pipeline_matmul_id_f32->a_s, "matmul_id_f32_aligned_s", matmul_id_f32_aligned_fp32_len, matmul_id_f32_aligned_fp32_data, "main", 4, sizeof(vk_mat_mat_push_constants), s_wg_denoms, warptile_s, s_align);
-
-        ggml_vk_create_pipeline(ctx, ctx->device->pipeline_matmul_id_f16->l, "matmul_id_f16_l", matmul_id_f16_fp32_len, matmul_id_f16_fp32_data, "main", 4, sizeof(vk_mat_mat_push_constants), l_wg_denoms, warptile_l, 1);
-        ggml_vk_create_pipeline(ctx, ctx->device->pipeline_matmul_id_f16->m, "matmul_id_f16_m", matmul_id_f16_fp32_len, matmul_id_f16_fp32_data, "main", 4, sizeof(vk_mat_mat_push_constants), m_wg_denoms, warptile_m, 1);
-        ggml_vk_create_pipeline(ctx, ctx->device->pipeline_matmul_id_f16->s, "matmul_id_f16_s", matmul_id_f16_fp32_len, matmul_id_f16_fp32_data, "main", 4, sizeof(vk_mat_mat_push_constants), s_wg_denoms, warptile_s, 1);
-        ggml_vk_create_pipeline(ctx, ctx->device->pipeline_matmul_id_f16->a_l, "matmul_id_f16_aligned_l", matmul_id_f16_aligned_fp32_len, matmul_id_f16_aligned_fp32_data, "main", 4, sizeof(vk_mat_mat_push_constants), l_wg_denoms, warptile_l, l_align);
-        ggml_vk_create_pipeline(ctx, ctx->device->pipeline_matmul_id_f16->a_m, "matmul_id_f16_aligned_m", matmul_id_f16_aligned_fp32_len, matmul_id_f16_aligned_fp32_data, "main", 4, sizeof(vk_mat_mat_push_constants), m_wg_denoms, warptile_m, m_align);
-        ggml_vk_create_pipeline(ctx, ctx->device->pipeline_matmul_id_f16->a_s, "matmul_id_f16_aligned_s", matmul_id_f16_aligned_fp32_len, matmul_id_f16_aligned_fp32_data, "main", 4, sizeof(vk_mat_mat_push_constants), s_wg_denoms, warptile_s, s_align);
-
-        ggml_vk_create_pipeline(ctx, ctx->device->pipeline_matmul_id_f16_f32->l, "matmul_id_f16_f32_l", matmul_id_f16_f32_fp32_len, matmul_id_f16_f32_fp32_data, "main", 4, sizeof(vk_mat_mat_push_constants), l_wg_denoms, warptile_l, 1);
-        ggml_vk_create_pipeline(ctx, ctx->device->pipeline_matmul_id_f16_f32->m, "matmul_id_f16_f32_m", matmul_id_f16_f32_fp32_len, matmul_id_f16_f32_fp32_data, "main", 4, sizeof(vk_mat_mat_push_constants), m_wg_denoms, warptile_m, 1);
-        ggml_vk_create_pipeline(ctx, ctx->device->pipeline_matmul_id_f16_f32->s, "matmul_id_f16_f32_s", matmul_id_f16_f32_fp32_len, matmul_id_f16_f32_fp32_data, "main", 4, sizeof(vk_mat_mat_push_constants), s_wg_denoms, warptile_s, 1);
-        ggml_vk_create_pipeline(ctx, ctx->device->pipeline_matmul_id_f16_f32->a_l, "matmul_id_f16_f32_aligned_l", matmul_id_f16_f32_aligned_fp32_len, matmul_id_f16_f32_aligned_fp32_data, "main", 4, sizeof(vk_mat_mat_push_constants), l_wg_denoms, warptile_l, l_align);
-        ggml_vk_create_pipeline(ctx, ctx->device->pipeline_matmul_id_f16_f32->a_m, "matmul_id_f16_f32_aligned_m", matmul_id_f16_f32_aligned_fp32_len, matmul_id_f16_f32_aligned_fp32_data, "main", 4, sizeof(vk_mat_mat_push_constants), m_wg_denoms, warptile_m, m_align);
-        ggml_vk_create_pipeline(ctx, ctx->device->pipeline_matmul_id_f16_f32->a_s, "matmul_id_f16_f32_aligned_s", matmul_id_f16_f32_aligned_fp32_len, matmul_id_f16_f32_aligned_fp32_data, "main", 4, sizeof(vk_mat_mat_push_constants), s_wg_denoms, warptile_s, s_align);
-
-        ggml_vk_create_pipeline(ctx, ctx->device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_0]->l, "matmul_id_q4_0_f32_l", matmul_id_q4_0_f32_fp32_len, matmul_id_q4_0_f32_fp32_data, "main", 4, sizeof(vk_mat_mat_push_constants), l_wg_denoms, warptile_mmq_l, l_align);
-        ggml_vk_create_pipeline(ctx, ctx->device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_0]->m, "matmul_id_q4_0_f32_m", matmul_id_q4_0_f32_fp32_len, matmul_id_q4_0_f32_fp32_data, "main", 4, sizeof(vk_mat_mat_push_constants), m_wg_denoms, warptile_mmq_m, m_align);
-        ggml_vk_create_pipeline(ctx, ctx->device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_0]->s, "matmul_id_q4_0_f32_s", matmul_id_q4_0_f32_fp32_len, matmul_id_q4_0_f32_fp32_data, "main", 4, sizeof(vk_mat_mat_push_constants), s_wg_denoms, warptile_mmq_s, s_align);
-        ggml_vk_create_pipeline(ctx, ctx->device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_0]->a_l, "matmul_id_q4_0_f32_aligned_l", matmul_id_q4_0_f32_aligned_fp32_len, matmul_id_q4_0_f32_aligned_fp32_data, "main", 4, sizeof(vk_mat_mat_push_constants), l_wg_denoms, warptile_mmq_l, l_align);
-        ggml_vk_create_pipeline(ctx, ctx->device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_0]->a_m, "matmul_id_q4_0_f32_aligned_m", matmul_id_q4_0_f32_aligned_fp32_len, matmul_id_q4_0_f32_aligned_fp32_data, "main", 4, sizeof(vk_mat_mat_push_constants), m_wg_denoms, warptile_mmq_m, m_align);
-        ggml_vk_create_pipeline(ctx, ctx->device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_0]->a_s, "matmul_id_q4_0_f32_aligned_s", matmul_id_q4_0_f32_aligned_fp32_len, matmul_id_q4_0_f32_aligned_fp32_data, "main", 4, sizeof(vk_mat_mat_push_constants), s_wg_denoms, warptile_mmq_s, s_align);
-
-        ggml_vk_create_pipeline(ctx, ctx->device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_1]->l, "matmul_id_q4_0_f32_l", matmul_id_q4_1_f32_fp32_len, matmul_id_q4_1_f32_fp32_data, "main", 4, sizeof(vk_mat_mat_push_constants), l_wg_denoms, warptile_mmq_l, l_align);
-        ggml_vk_create_pipeline(ctx, ctx->device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_1]->m, "matmul_id_q4_0_f32_m", matmul_id_q4_1_f32_fp32_len, matmul_id_q4_1_f32_fp32_data, "main", 4, sizeof(vk_mat_mat_push_constants), m_wg_denoms, warptile_mmq_m, m_align);
-        ggml_vk_create_pipeline(ctx, ctx->device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_1]->s, "matmul_id_q4_0_f32_s", matmul_id_q4_1_f32_fp32_len, matmul_id_q4_1_f32_fp32_data, "main", 4, sizeof(vk_mat_mat_push_constants), s_wg_denoms, warptile_mmq_s, s_align);
-        ggml_vk_create_pipeline(ctx, ctx->device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_1]->a_l, "matmul_id_q4_0_f32_aligned_l", matmul_id_q4_1_f32_aligned_fp32_len, matmul_id_q4_1_f32_aligned_fp32_data, "main", 4, sizeof(vk_mat_mat_push_constants), l_wg_denoms, warptile_mmq_l, l_align);
-        ggml_vk_create_pipeline(ctx, ctx->device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_1]->a_m, "matmul_id_q4_0_f32_aligned_m", matmul_id_q4_1_f32_aligned_fp32_len, matmul_id_q4_1_f32_aligned_fp32_data, "main", 4, sizeof(vk_mat_mat_push_constants), m_wg_denoms, warptile_mmq_m, m_align);
-        ggml_vk_create_pipeline(ctx, ctx->device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_1]->a_s, "matmul_id_q4_0_f32_aligned_s", matmul_id_q4_1_f32_aligned_fp32_len, matmul_id_q4_1_f32_aligned_fp32_data, "main", 4, sizeof(vk_mat_mat_push_constants), s_wg_denoms, warptile_mmq_s, s_align);
-
-        ggml_vk_create_pipeline(ctx, ctx->device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q5_0]->l, "matmul_id_q5_0_f32_l", matmul_id_q5_0_f32_fp32_len, matmul_id_q5_0_f32_fp32_data, "main", 4, sizeof(vk_mat_mat_push_constants), l_wg_denoms, warptile_mmq_l, l_align);
-        ggml_vk_create_pipeline(ctx, ctx->device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q5_0]->m, "matmul_id_q5_0_f32_m", matmul_id_q5_0_f32_fp32_len, matmul_id_q5_0_f32_fp32_data, "main", 4, sizeof(vk_mat_mat_push_constants), m_wg_denoms, warptile_mmq_m, m_align);
-        ggml_vk_create_pipeline(ctx, ctx->device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q5_0]->s, "matmul_id_q5_0_f32_s", matmul_id_q5_0_f32_fp32_len, matmul_id_q5_0_f32_fp32_data, "main", 4, sizeof(vk_mat_mat_push_constants), s_wg_denoms, warptile_mmq_s, s_align);
-        ggml_vk_create_pipeline(ctx, ctx->device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q5_0]->a_l, "matmul_id_q5_0_f32_aligned_l", matmul_id_q5_0_f32_aligned_fp32_len, matmul_id_q5_0_f32_aligned_fp32_data, "main", 4, sizeof(vk_mat_mat_push_constants), l_wg_denoms, warptile_mmq_l, l_align);
-        ggml_vk_create_pipeline(ctx, ctx->device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q5_0]->a_m, "matmul_id_q5_0_f32_aligned_m", matmul_id_q5_0_f32_aligned_fp32_len, matmul_id_q5_0_f32_aligned_fp32_data, "main", 4, sizeof(vk_mat_mat_push_constants), m_wg_denoms, warptile_mmq_m, m_align);
-        ggml_vk_create_pipeline(ctx, ctx->device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q5_0]->a_s, "matmul_id_q5_0_f32_aligned_s", matmul_id_q5_0_f32_aligned_fp32_len, matmul_id_q5_0_f32_aligned_fp32_data, "main", 4, sizeof(vk_mat_mat_push_constants), s_wg_denoms, warptile_mmq_s, s_align);
-
-        ggml_vk_create_pipeline(ctx, ctx->device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q5_1]->l, "matmul_id_q5_1_f32_l", matmul_id_q5_1_f32_fp32_len, matmul_id_q5_1_f32_fp32_data, "main", 4, sizeof(vk_mat_mat_push_constants), l_wg_denoms, warptile_mmq_l, l_align);
-        ggml_vk_create_pipeline(ctx, ctx->device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q5_1]->m, "matmul_id_q5_1_f32_m", matmul_id_q5_1_f32_fp32_len, matmul_id_q5_1_f32_fp32_data, "main", 4, sizeof(vk_mat_mat_push_constants), m_wg_denoms, warptile_mmq_m, m_align);
-        ggml_vk_create_pipeline(ctx, ctx->device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q5_1]->s, "matmul_id_q5_1_f32_s", matmul_id_q5_1_f32_fp32_len, matmul_id_q5_1_f32_fp32_data, "main", 4, sizeof(vk_mat_mat_push_constants), s_wg_denoms, warptile_mmq_s, s_align);
-        ggml_vk_create_pipeline(ctx, ctx->device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q5_1]->a_l, "matmul_id_q5_1_f32_aligned_l", matmul_id_q5_1_f32_aligned_fp32_len, matmul_id_q5_1_f32_aligned_fp32_data, "main", 4, sizeof(vk_mat_mat_push_constants), l_wg_denoms, warptile_mmq_l, l_align);
-        ggml_vk_create_pipeline(ctx, ctx->device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q5_1]->a_m, "matmul_id_q5_1_f32_aligned_m", matmul_id_q5_1_f32_aligned_fp32_len, matmul_id_q5_1_f32_aligned_fp32_data, "main", 4, sizeof(vk_mat_mat_push_constants), m_wg_denoms, warptile_mmq_m, m_align);
-        ggml_vk_create_pipeline(ctx, ctx->device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q5_1]->a_s, "matmul_id_q5_1_f32_aligned_s", matmul_id_q5_1_f32_aligned_fp32_len, matmul_id_q5_1_f32_aligned_fp32_data, "main", 4, sizeof(vk_mat_mat_push_constants), s_wg_denoms, warptile_mmq_s, s_align);
-
-        ggml_vk_create_pipeline(ctx, ctx->device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q8_0]->l, "matmul_id_q8_0_f32_l", matmul_id_q8_0_f32_fp32_len, matmul_id_q8_0_f32_fp32_data, "main", 4, sizeof(vk_mat_mat_push_constants), l_wg_denoms, warptile_mmq_l, l_align);
-        ggml_vk_create_pipeline(ctx, ctx->device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q8_0]->m, "matmul_id_q8_0_f32_m", matmul_id_q8_0_f32_fp32_len, matmul_id_q8_0_f32_fp32_data, "main", 4, sizeof(vk_mat_mat_push_constants), m_wg_denoms, warptile_mmq_m, m_align);
-        ggml_vk_create_pipeline(ctx, ctx->device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q8_0]->s, "matmul_id_q8_0_f32_s", matmul_id_q8_0_f32_fp32_len, matmul_id_q8_0_f32_fp32_data, "main", 4, sizeof(vk_mat_mat_push_constants), s_wg_denoms, warptile_mmq_s, s_align);
-        ggml_vk_create_pipeline(ctx, ctx->device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q8_0]->a_l, "matmul_id_q8_0_f32_aligned_l", matmul_id_q8_0_f32_aligned_fp32_len, matmul_id_q8_0_f32_aligned_fp32_data, "main", 4, sizeof(vk_mat_mat_push_constants), l_wg_denoms, warptile_mmq_l, l_align);
-        ggml_vk_create_pipeline(ctx, ctx->device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q8_0]->a_m, "matmul_id_q8_0_f32_aligned_m", matmul_id_q8_0_f32_aligned_fp32_len, matmul_id_q8_0_f32_aligned_fp32_data, "main", 4, sizeof(vk_mat_mat_push_constants), m_wg_denoms, warptile_mmq_m, m_align);
-        ggml_vk_create_pipeline(ctx, ctx->device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q8_0]->a_s, "matmul_id_q8_0_f32_aligned_s", matmul_id_q8_0_f32_aligned_fp32_len, matmul_id_q8_0_f32_aligned_fp32_data, "main", 4, sizeof(vk_mat_mat_push_constants), s_wg_denoms, warptile_mmq_s, s_align);
-
-        ggml_vk_create_pipeline(ctx, ctx->device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q2_K]->l, "matmul_id_q2_k_f32_l", matmul_id_q2_k_f32_fp32_len, matmul_id_q2_k_f32_fp32_data, "main", 4, sizeof(vk_mat_mat_push_constants), l_wg_denoms, warptile_mmq_l, l_align);
-        ggml_vk_create_pipeline(ctx, ctx->device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q2_K]->m, "matmul_id_q2_k_f32_m", matmul_id_q2_k_f32_fp32_len, matmul_id_q2_k_f32_fp32_data, "main", 4, sizeof(vk_mat_mat_push_constants), m_wg_denoms, warptile_mmq_m, m_align);
-        ggml_vk_create_pipeline(ctx, ctx->device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q2_K]->s, "matmul_id_q2_k_f32_s", matmul_id_q2_k_f32_fp32_len, matmul_id_q2_k_f32_fp32_data, "main", 4, sizeof(vk_mat_mat_push_constants), s_wg_denoms, warptile_mmq_s, s_align);
-        ggml_vk_create_pipeline(ctx, ctx->device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q2_K]->a_l, "matmul_id_q2_k_f32_aligned_l", matmul_id_q2_k_f32_aligned_fp32_len, matmul_id_q2_k_f32_aligned_fp32_data, "main", 4, sizeof(vk_mat_mat_push_constants), l_wg_denoms, warptile_mmq_l, l_align);
-        ggml_vk_create_pipeline(ctx, ctx->device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q2_K]->a_m, "matmul_id_q2_k_f32_aligned_m", matmul_id_q2_k_f32_aligned_fp32_len, matmul_id_q2_k_f32_aligned_fp32_data, "main", 4, sizeof(vk_mat_mat_push_constants), m_wg_denoms, warptile_mmq_m, m_align);
-        ggml_vk_create_pipeline(ctx, ctx->device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q2_K]->a_s, "matmul_id_q2_k_f32_aligned_s", matmul_id_q2_k_f32_aligned_fp32_len, matmul_id_q2_k_f32_aligned_fp32_data, "main", 4, sizeof(vk_mat_mat_push_constants), s_wg_denoms, warptile_mmq_s, s_align);
-
-        ggml_vk_create_pipeline(ctx, ctx->device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q3_K]->l, "matmul_id_q3_k_f32_l", matmul_id_q3_k_f32_fp32_len, matmul_id_q3_k_f32_fp32_data, "main", 4, sizeof(vk_mat_mat_push_constants), l_wg_denoms, warptile_mmq_l, l_align);
-        ggml_vk_create_pipeline(ctx, ctx->device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q3_K]->m, "matmul_id_q3_k_f32_m", matmul_id_q3_k_f32_fp32_len, matmul_id_q3_k_f32_fp32_data, "main", 4, sizeof(vk_mat_mat_push_constants), m_wg_denoms, warptile_mmq_m, m_align);
-        ggml_vk_create_pipeline(ctx, ctx->device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q3_K]->s, "matmul_id_q3_k_f32_s", matmul_id_q3_k_f32_fp32_len, matmul_id_q3_k_f32_fp32_data, "main", 4, sizeof(vk_mat_mat_push_constants), s_wg_denoms, warptile_mmq_s, s_align);
-        ggml_vk_create_pipeline(ctx, ctx->device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q3_K]->a_l, "matmul_id_q3_k_f32_aligned_l", matmul_id_q3_k_f32_aligned_fp32_len, matmul_id_q3_k_f32_aligned_fp32_data, "main", 4, sizeof(vk_mat_mat_push_constants), l_wg_denoms, warptile_mmq_l, l_align);
-        ggml_vk_create_pipeline(ctx, ctx->device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q3_K]->a_m, "matmul_id_q3_k_f32_aligned_m", matmul_id_q3_k_f32_aligned_fp32_len, matmul_id_q3_k_f32_aligned_fp32_data, "main", 4, sizeof(vk_mat_mat_push_constants), m_wg_denoms, warptile_mmq_m, m_align);
-        ggml_vk_create_pipeline(ctx, ctx->device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q3_K]->a_s, "matmul_id_q3_k_f32_aligned_s", matmul_id_q3_k_f32_aligned_fp32_len, matmul_id_q3_k_f32_aligned_fp32_data, "main", 4, sizeof(vk_mat_mat_push_constants), s_wg_denoms, warptile_mmq_s, s_align);
-
-        ggml_vk_create_pipeline(ctx, ctx->device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_K]->l, "matmul_id_q4_k_f32_l", matmul_id_q4_k_f32_fp32_len, matmul_id_q4_k_f32_fp32_data, "main", 4, sizeof(vk_mat_mat_push_constants), l_wg_denoms, warptile_mmq_l, l_align);
-        ggml_vk_create_pipeline(ctx, ctx->device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_K]->m, "matmul_id_q4_k_f32_m", matmul_id_q4_k_f32_fp32_len, matmul_id_q4_k_f32_fp32_data, "main", 4, sizeof(vk_mat_mat_push_constants), m_wg_denoms, warptile_mmq_m, m_align);
-        ggml_vk_create_pipeline(ctx, ctx->device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_K]->s, "matmul_id_q4_k_f32_s", matmul_id_q4_k_f32_fp32_len, matmul_id_q4_k_f32_fp32_data, "main", 4, sizeof(vk_mat_mat_push_constants), s_wg_denoms, warptile_mmq_s, s_align);
-        ggml_vk_create_pipeline(ctx, ctx->device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_K]->a_l, "matmul_id_q4_k_f32_aligned_l", matmul_id_q4_k_f32_aligned_fp32_len, matmul_id_q4_k_f32_aligned_fp32_data, "main", 4, sizeof(vk_mat_mat_push_constants), l_wg_denoms, warptile_mmq_l, l_align);
-        ggml_vk_create_pipeline(ctx, ctx->device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_K]->a_m, "matmul_id_q4_k_f32_aligned_m", matmul_id_q4_k_f32_aligned_fp32_len, matmul_id_q4_k_f32_aligned_fp32_data, "main", 4, sizeof(vk_mat_mat_push_constants), m_wg_denoms, warptile_mmq_m, m_align);
-        ggml_vk_create_pipeline(ctx, ctx->device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_K]->a_s, "matmul_id_q4_k_f32_aligned_s", matmul_id_q4_k_f32_aligned_fp32_len, matmul_id_q4_k_f32_aligned_fp32_data, "main", 4, sizeof(vk_mat_mat_push_constants), s_wg_denoms, warptile_mmq_s, s_align);
-
-        ggml_vk_create_pipeline(ctx, ctx->device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q5_K]->l, "matmul_id_q5_k_f32_l", matmul_id_q5_k_f32_fp32_len, matmul_id_q5_k_f32_fp32_data, "main", 4, sizeof(vk_mat_mat_push_constants), l_wg_denoms, warptile_mmq_l, l_align);
-        ggml_vk_create_pipeline(ctx, ctx->device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q5_K]->m, "matmul_id_q5_k_f32_m", matmul_id_q5_k_f32_fp32_len, matmul_id_q5_k_f32_fp32_data, "main", 4, sizeof(vk_mat_mat_push_constants), m_wg_denoms, warptile_mmq_m, m_align);
-        ggml_vk_create_pipeline(ctx, ctx->device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q5_K]->s, "matmul_id_q5_k_f32_s", matmul_id_q5_k_f32_fp32_len, matmul_id_q5_k_f32_fp32_data, "main", 4, sizeof(vk_mat_mat_push_constants), s_wg_denoms, warptile_mmq_s, s_align);
-        ggml_vk_create_pipeline(ctx, ctx->device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q5_K]->a_l, "matmul_id_q5_k_f32_aligned_l", matmul_id_q5_k_f32_aligned_fp32_len, matmul_id_q5_k_f32_aligned_fp32_data, "main", 4, sizeof(vk_mat_mat_push_constants), l_wg_denoms, warptile_mmq_l, l_align);
-        ggml_vk_create_pipeline(ctx, ctx->device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q5_K]->a_m, "matmul_id_q5_k_f32_aligned_m", matmul_id_q5_k_f32_aligned_fp32_len, matmul_id_q5_k_f32_aligned_fp32_data, "main", 4, sizeof(vk_mat_mat_push_constants), m_wg_denoms, warptile_mmq_m, m_align);
-        ggml_vk_create_pipeline(ctx, ctx->device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q5_K]->a_s, "matmul_id_q5_k_f32_aligned_s", matmul_id_q5_k_f32_aligned_fp32_len, matmul_id_q5_k_f32_aligned_fp32_data, "main", 4, sizeof(vk_mat_mat_push_constants), s_wg_denoms, warptile_mmq_s, s_align);
-
-        ggml_vk_create_pipeline(ctx, ctx->device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q6_K]->l, "matmul_id_q6_k_f32_l", matmul_id_q6_k_f32_fp32_len, matmul_id_q6_k_f32_fp32_data, "main", 4, sizeof(vk_mat_mat_push_constants), l_wg_denoms, warptile_mmq_l, l_align);
-        ggml_vk_create_pipeline(ctx, ctx->device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q6_K]->m, "matmul_id_q6_k_f32_m", matmul_id_q6_k_f32_fp32_len, matmul_id_q6_k_f32_fp32_data, "main", 4, sizeof(vk_mat_mat_push_constants), m_wg_denoms, warptile_mmq_m, m_align);
-        ggml_vk_create_pipeline(ctx, ctx->device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q6_K]->s, "matmul_id_q6_k_f32_s", matmul_id_q6_k_f32_fp32_len, matmul_id_q6_k_f32_fp32_data, "main", 4, sizeof(vk_mat_mat_push_constants), s_wg_denoms, warptile_mmq_s, s_align);
-        ggml_vk_create_pipeline(ctx, ctx->device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q6_K]->a_l, "matmul_id_q6_k_f32_aligned_l", matmul_id_q6_k_f32_aligned_fp32_len, matmul_id_q6_k_f32_aligned_fp32_data, "main", 4, sizeof(vk_mat_mat_push_constants), l_wg_denoms, warptile_mmq_l, l_align);
-        ggml_vk_create_pipeline(ctx, ctx->device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q6_K]->a_m, "matmul_id_q6_k_f32_aligned_m", matmul_id_q6_k_f32_aligned_fp32_len, matmul_id_q6_k_f32_aligned_fp32_data, "main", 4, sizeof(vk_mat_mat_push_constants), m_wg_denoms, warptile_mmq_m, m_align);
-        ggml_vk_create_pipeline(ctx, ctx->device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q6_K]->a_s, "matmul_id_q6_k_f32_aligned_s", matmul_id_q6_k_f32_aligned_fp32_len, matmul_id_q6_k_f32_aligned_fp32_data, "main", 4, sizeof(vk_mat_mat_push_constants), s_wg_denoms, warptile_mmq_s, s_align);*/
+        ggml_vk_create_pipeline(ctx, ctx->device->pipeline_matmul_id_f32->l, "matmul_id_f32_l", matmul_id_f32_fp32_len, matmul_id_f32_fp32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), l_wg_denoms, warptile_l, 1);
+        ggml_vk_create_pipeline(ctx, ctx->device->pipeline_matmul_id_f32->m, "matmul_id_f32_m", matmul_id_f32_fp32_len, matmul_id_f32_fp32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), m_wg_denoms, warptile_m, 1);
+        ggml_vk_create_pipeline(ctx, ctx->device->pipeline_matmul_id_f32->s, "matmul_id_f32_s", matmul_id_f32_fp32_len, matmul_id_f32_fp32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), s_wg_denoms, warptile_s, 1);
+        ggml_vk_create_pipeline(ctx, ctx->device->pipeline_matmul_id_f32->a_l, "matmul_id_f32_aligned_l", matmul_id_f32_aligned_fp32_len, matmul_id_f32_aligned_fp32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), l_wg_denoms, warptile_l, l_align);
+        ggml_vk_create_pipeline(ctx, ctx->device->pipeline_matmul_id_f32->a_m, "matmul_id_f32_aligned_m", matmul_id_f32_aligned_fp32_len, matmul_id_f32_aligned_fp32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), m_wg_denoms, warptile_m, m_align);
+        ggml_vk_create_pipeline(ctx, ctx->device->pipeline_matmul_id_f32->a_s, "matmul_id_f32_aligned_s", matmul_id_f32_aligned_fp32_len, matmul_id_f32_aligned_fp32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), s_wg_denoms, warptile_s, s_align);
+
+        ggml_vk_create_pipeline(ctx, ctx->device->pipeline_matmul_id_f16->l, "matmul_id_f16_l", matmul_id_f16_fp32_len, matmul_id_f16_fp32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), l_wg_denoms, warptile_l, 1);
+        ggml_vk_create_pipeline(ctx, ctx->device->pipeline_matmul_id_f16->m, "matmul_id_f16_m", matmul_id_f16_fp32_len, matmul_id_f16_fp32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), m_wg_denoms, warptile_m, 1);
+        ggml_vk_create_pipeline(ctx, ctx->device->pipeline_matmul_id_f16->s, "matmul_id_f16_s", matmul_id_f16_fp32_len, matmul_id_f16_fp32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), s_wg_denoms, warptile_s, 1);
+        ggml_vk_create_pipeline(ctx, ctx->device->pipeline_matmul_id_f16->a_l, "matmul_id_f16_aligned_l", matmul_id_f16_aligned_fp32_len, matmul_id_f16_aligned_fp32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), l_wg_denoms, warptile_l, l_align);
+        ggml_vk_create_pipeline(ctx, ctx->device->pipeline_matmul_id_f16->a_m, "matmul_id_f16_aligned_m", matmul_id_f16_aligned_fp32_len, matmul_id_f16_aligned_fp32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), m_wg_denoms, warptile_m, m_align);
+        ggml_vk_create_pipeline(ctx, ctx->device->pipeline_matmul_id_f16->a_s, "matmul_id_f16_aligned_s", matmul_id_f16_aligned_fp32_len, matmul_id_f16_aligned_fp32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), s_wg_denoms, warptile_s, s_align);
+
+        ggml_vk_create_pipeline(ctx, ctx->device->pipeline_matmul_id_f16_f32->l, "matmul_id_f16_f32_l", matmul_id_f16_f32_fp32_len, matmul_id_f16_f32_fp32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), l_wg_denoms, warptile_l, 1);
+        ggml_vk_create_pipeline(ctx, ctx->device->pipeline_matmul_id_f16_f32->m, "matmul_id_f16_f32_m", matmul_id_f16_f32_fp32_len, matmul_id_f16_f32_fp32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), m_wg_denoms, warptile_m, 1);
+        ggml_vk_create_pipeline(ctx, ctx->device->pipeline_matmul_id_f16_f32->s, "matmul_id_f16_f32_s", matmul_id_f16_f32_fp32_len, matmul_id_f16_f32_fp32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), s_wg_denoms, warptile_s, 1);
+        ggml_vk_create_pipeline(ctx, ctx->device->pipeline_matmul_id_f16_f32->a_l, "matmul_id_f16_f32_aligned_l", matmul_id_f16_f32_aligned_fp32_len, matmul_id_f16_f32_aligned_fp32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), l_wg_denoms, warptile_l, l_align);
+        ggml_vk_create_pipeline(ctx, ctx->device->pipeline_matmul_id_f16_f32->a_m, "matmul_id_f16_f32_aligned_m", matmul_id_f16_f32_aligned_fp32_len, matmul_id_f16_f32_aligned_fp32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), m_wg_denoms, warptile_m, m_align);
+        ggml_vk_create_pipeline(ctx, ctx->device->pipeline_matmul_id_f16_f32->a_s, "matmul_id_f16_f32_aligned_s", matmul_id_f16_f32_aligned_fp32_len, matmul_id_f16_f32_aligned_fp32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), s_wg_denoms, warptile_s, s_align);
+
+        ggml_vk_create_pipeline(ctx, ctx->device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_0]->l, "matmul_id_q4_0_f32_l", matmul_id_q4_0_f32_fp32_len, matmul_id_q4_0_f32_fp32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), l_wg_denoms, warptile_mmq_l, l_align);
+        ggml_vk_create_pipeline(ctx, ctx->device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_0]->m, "matmul_id_q4_0_f32_m", matmul_id_q4_0_f32_fp32_len, matmul_id_q4_0_f32_fp32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), m_wg_denoms, warptile_mmq_m, m_align);
+        ggml_vk_create_pipeline(ctx, ctx->device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_0]->s, "matmul_id_q4_0_f32_s", matmul_id_q4_0_f32_fp32_len, matmul_id_q4_0_f32_fp32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), s_wg_denoms, warptile_mmq_s, s_align);
+        ggml_vk_create_pipeline(ctx, ctx->device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_0]->a_l, "matmul_id_q4_0_f32_aligned_l", matmul_id_q4_0_f32_aligned_fp32_len, matmul_id_q4_0_f32_aligned_fp32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), l_wg_denoms, warptile_mmq_l, l_align);
+        ggml_vk_create_pipeline(ctx, ctx->device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_0]->a_m, "matmul_id_q4_0_f32_aligned_m", matmul_id_q4_0_f32_aligned_fp32_len, matmul_id_q4_0_f32_aligned_fp32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), m_wg_denoms, warptile_mmq_m, m_align);
+        ggml_vk_create_pipeline(ctx, ctx->device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_0]->a_s, "matmul_id_q4_0_f32_aligned_s", matmul_id_q4_0_f32_aligned_fp32_len, matmul_id_q4_0_f32_aligned_fp32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), s_wg_denoms, warptile_mmq_s, s_align);
+
+        ggml_vk_create_pipeline(ctx, ctx->device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_1]->l, "matmul_id_q4_1_f32_l", matmul_id_q4_1_f32_fp32_len, matmul_id_q4_1_f32_fp32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), l_wg_denoms, warptile_mmq_l, l_align);
+        ggml_vk_create_pipeline(ctx, ctx->device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_1]->m, "matmul_id_q4_1_f32_m", matmul_id_q4_1_f32_fp32_len, matmul_id_q4_1_f32_fp32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), m_wg_denoms, warptile_mmq_m, m_align);
+        ggml_vk_create_pipeline(ctx, ctx->device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_1]->s, "matmul_id_q4_1_f32_s", matmul_id_q4_1_f32_fp32_len, matmul_id_q4_1_f32_fp32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), s_wg_denoms, warptile_mmq_s, s_align);
+        ggml_vk_create_pipeline(ctx, ctx->device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_1]->a_l, "matmul_id_q4_1_f32_aligned_l", matmul_id_q4_1_f32_aligned_fp32_len, matmul_id_q4_1_f32_aligned_fp32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), l_wg_denoms, warptile_mmq_l, l_align);
+        ggml_vk_create_pipeline(ctx, ctx->device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_1]->a_m, "matmul_id_q4_1_f32_aligned_m", matmul_id_q4_1_f32_aligned_fp32_len, matmul_id_q4_1_f32_aligned_fp32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), m_wg_denoms, warptile_mmq_m, m_align);
+        ggml_vk_create_pipeline(ctx, ctx->device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_1]->a_s, "matmul_id_q4_1_f32_aligned_s", matmul_id_q4_1_f32_aligned_fp32_len, matmul_id_q4_1_f32_aligned_fp32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), s_wg_denoms, warptile_mmq_s, s_align);
+
+        ggml_vk_create_pipeline(ctx, ctx->device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q5_0]->l, "matmul_id_q5_0_f32_l", matmul_id_q5_0_f32_fp32_len, matmul_id_q5_0_f32_fp32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), l_wg_denoms, warptile_mmq_l, l_align);
+        ggml_vk_create_pipeline(ctx, ctx->device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q5_0]->m, "matmul_id_q5_0_f32_m", matmul_id_q5_0_f32_fp32_len, matmul_id_q5_0_f32_fp32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), m_wg_denoms, warptile_mmq_m, m_align);
+        ggml_vk_create_pipeline(ctx, ctx->device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q5_0]->s, "matmul_id_q5_0_f32_s", matmul_id_q5_0_f32_fp32_len, matmul_id_q5_0_f32_fp32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), s_wg_denoms, warptile_mmq_s, s_align);
+        ggml_vk_create_pipeline(ctx, ctx->device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q5_0]->a_l, "matmul_id_q5_0_f32_aligned_l", matmul_id_q5_0_f32_aligned_fp32_len, matmul_id_q5_0_f32_aligned_fp32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), l_wg_denoms, warptile_mmq_l, l_align);
+        ggml_vk_create_pipeline(ctx, ctx->device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q5_0]->a_m, "matmul_id_q5_0_f32_aligned_m", matmul_id_q5_0_f32_aligned_fp32_len, matmul_id_q5_0_f32_aligned_fp32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), m_wg_denoms, warptile_mmq_m, m_align);
+        ggml_vk_create_pipeline(ctx, ctx->device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q5_0]->a_s, "matmul_id_q5_0_f32_aligned_s", matmul_id_q5_0_f32_aligned_fp32_len, matmul_id_q5_0_f32_aligned_fp32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), s_wg_denoms, warptile_mmq_s, s_align);
+
+        ggml_vk_create_pipeline(ctx, ctx->device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q5_1]->l, "matmul_id_q5_1_f32_l", matmul_id_q5_1_f32_fp32_len, matmul_id_q5_1_f32_fp32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), l_wg_denoms, warptile_mmq_l, l_align);
+        ggml_vk_create_pipeline(ctx, ctx->device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q5_1]->m, "matmul_id_q5_1_f32_m", matmul_id_q5_1_f32_fp32_len, matmul_id_q5_1_f32_fp32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), m_wg_denoms, warptile_mmq_m, m_align);
+        ggml_vk_create_pipeline(ctx, ctx->device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q5_1]->s, "matmul_id_q5_1_f32_s", matmul_id_q5_1_f32_fp32_len, matmul_id_q5_1_f32_fp32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), s_wg_denoms, warptile_mmq_s, s_align);
+        ggml_vk_create_pipeline(ctx, ctx->device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q5_1]->a_l, "matmul_id_q5_1_f32_aligned_l", matmul_id_q5_1_f32_aligned_fp32_len, matmul_id_q5_1_f32_aligned_fp32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), l_wg_denoms, warptile_mmq_l, l_align);
+        ggml_vk_create_pipeline(ctx, ctx->device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q5_1]->a_m, "matmul_id_q5_1_f32_aligned_m", matmul_id_q5_1_f32_aligned_fp32_len, matmul_id_q5_1_f32_aligned_fp32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), m_wg_denoms, warptile_mmq_m, m_align);
+        ggml_vk_create_pipeline(ctx, ctx->device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q5_1]->a_s, "matmul_id_q5_1_f32_aligned_s", matmul_id_q5_1_f32_aligned_fp32_len, matmul_id_q5_1_f32_aligned_fp32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), s_wg_denoms, warptile_mmq_s, s_align);
+
+        ggml_vk_create_pipeline(ctx, ctx->device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q8_0]->l, "matmul_id_q8_0_f32_l", matmul_id_q8_0_f32_fp32_len, matmul_id_q8_0_f32_fp32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), l_wg_denoms, warptile_mmq_l, l_align);
+        ggml_vk_create_pipeline(ctx, ctx->device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q8_0]->m, "matmul_id_q8_0_f32_m", matmul_id_q8_0_f32_fp32_len, matmul_id_q8_0_f32_fp32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), m_wg_denoms, warptile_mmq_m, m_align);
+        ggml_vk_create_pipeline(ctx, ctx->device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q8_0]->s, "matmul_id_q8_0_f32_s", matmul_id_q8_0_f32_fp32_len, matmul_id_q8_0_f32_fp32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), s_wg_denoms, warptile_mmq_s, s_align);
+        ggml_vk_create_pipeline(ctx, ctx->device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q8_0]->a_l, "matmul_id_q8_0_f32_aligned_l", matmul_id_q8_0_f32_aligned_fp32_len, matmul_id_q8_0_f32_aligned_fp32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), l_wg_denoms, warptile_mmq_l, l_align);
+        ggml_vk_create_pipeline(ctx, ctx->device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q8_0]->a_m, "matmul_id_q8_0_f32_aligned_m", matmul_id_q8_0_f32_aligned_fp32_len, matmul_id_q8_0_f32_aligned_fp32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), m_wg_denoms, warptile_mmq_m, m_align);
+        ggml_vk_create_pipeline(ctx, ctx->device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q8_0]->a_s, "matmul_id_q8_0_f32_aligned_s", matmul_id_q8_0_f32_aligned_fp32_len, matmul_id_q8_0_f32_aligned_fp32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), s_wg_denoms, warptile_mmq_s, s_align);
+
+        ggml_vk_create_pipeline(ctx, ctx->device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q2_K]->l, "matmul_id_q2_k_f32_l", matmul_id_q2_k_f32_fp32_len, matmul_id_q2_k_f32_fp32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), l_wg_denoms, warptile_mmq_l, l_align);
+        ggml_vk_create_pipeline(ctx, ctx->device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q2_K]->m, "matmul_id_q2_k_f32_m", matmul_id_q2_k_f32_fp32_len, matmul_id_q2_k_f32_fp32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), m_wg_denoms, warptile_mmq_m, m_align);
+        ggml_vk_create_pipeline(ctx, ctx->device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q2_K]->s, "matmul_id_q2_k_f32_s", matmul_id_q2_k_f32_fp32_len, matmul_id_q2_k_f32_fp32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), s_wg_denoms, warptile_mmq_s, s_align);
+        ggml_vk_create_pipeline(ctx, ctx->device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q2_K]->a_l, "matmul_id_q2_k_f32_aligned_l", matmul_id_q2_k_f32_aligned_fp32_len, matmul_id_q2_k_f32_aligned_fp32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), l_wg_denoms, warptile_mmq_l, l_align);
+        ggml_vk_create_pipeline(ctx, ctx->device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q2_K]->a_m, "matmul_id_q2_k_f32_aligned_m", matmul_id_q2_k_f32_aligned_fp32_len, matmul_id_q2_k_f32_aligned_fp32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), m_wg_denoms, warptile_mmq_m, m_align);
+        ggml_vk_create_pipeline(ctx, ctx->device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q2_K]->a_s, "matmul_id_q2_k_f32_aligned_s", matmul_id_q2_k_f32_aligned_fp32_len, matmul_id_q2_k_f32_aligned_fp32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), s_wg_denoms, warptile_mmq_s, s_align);
+
+        ggml_vk_create_pipeline(ctx, ctx->device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q3_K]->l, "matmul_id_q3_k_f32_l", matmul_id_q3_k_f32_fp32_len, matmul_id_q3_k_f32_fp32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), l_wg_denoms, warptile_mmq_l, l_align);
+        ggml_vk_create_pipeline(ctx, ctx->device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q3_K]->m, "matmul_id_q3_k_f32_m", matmul_id_q3_k_f32_fp32_len, matmul_id_q3_k_f32_fp32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), m_wg_denoms, warptile_mmq_m, m_align);
+        ggml_vk_create_pipeline(ctx, ctx->device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q3_K]->s, "matmul_id_q3_k_f32_s", matmul_id_q3_k_f32_fp32_len, matmul_id_q3_k_f32_fp32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), s_wg_denoms, warptile_mmq_s, s_align);
+        ggml_vk_create_pipeline(ctx, ctx->device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q3_K]->a_l, "matmul_id_q3_k_f32_aligned_l", matmul_id_q3_k_f32_aligned_fp32_len, matmul_id_q3_k_f32_aligned_fp32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), l_wg_denoms, warptile_mmq_l, l_align);
+        ggml_vk_create_pipeline(ctx, ctx->device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q3_K]->a_m, "matmul_id_q3_k_f32_aligned_m", matmul_id_q3_k_f32_aligned_fp32_len, matmul_id_q3_k_f32_aligned_fp32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), m_wg_denoms, warptile_mmq_m, m_align);
+        ggml_vk_create_pipeline(ctx, ctx->device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q3_K]->a_s, "matmul_id_q3_k_f32_aligned_s", matmul_id_q3_k_f32_aligned_fp32_len, matmul_id_q3_k_f32_aligned_fp32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), s_wg_denoms, warptile_mmq_s, s_align);
+
+        ggml_vk_create_pipeline(ctx, ctx->device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_K]->l, "matmul_id_q4_k_f32_l", matmul_id_q4_k_f32_fp32_len, matmul_id_q4_k_f32_fp32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), l_wg_denoms, warptile_mmq_l, l_align);
+        ggml_vk_create_pipeline(ctx, ctx->device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_K]->m, "matmul_id_q4_k_f32_m", matmul_id_q4_k_f32_fp32_len, matmul_id_q4_k_f32_fp32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), m_wg_denoms, warptile_mmq_m, m_align);
+        ggml_vk_create_pipeline(ctx, ctx->device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_K]->s, "matmul_id_q4_k_f32_s", matmul_id_q4_k_f32_fp32_len, matmul_id_q4_k_f32_fp32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), s_wg_denoms, warptile_mmq_s, s_align);
+        ggml_vk_create_pipeline(ctx, ctx->device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_K]->a_l, "matmul_id_q4_k_f32_aligned_l", matmul_id_q4_k_f32_aligned_fp32_len, matmul_id_q4_k_f32_aligned_fp32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), l_wg_denoms, warptile_mmq_l, l_align);
+        ggml_vk_create_pipeline(ctx, ctx->device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_K]->a_m, "matmul_id_q4_k_f32_aligned_m", matmul_id_q4_k_f32_aligned_fp32_len, matmul_id_q4_k_f32_aligned_fp32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), m_wg_denoms, warptile_mmq_m, m_align);
+        ggml_vk_create_pipeline(ctx, ctx->device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_K]->a_s, "matmul_id_q4_k_f32_aligned_s", matmul_id_q4_k_f32_aligned_fp32_len, matmul_id_q4_k_f32_aligned_fp32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), s_wg_denoms, warptile_mmq_s, s_align);
+
+        ggml_vk_create_pipeline(ctx, ctx->device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q5_K]->l, "matmul_id_q5_k_f32_l", matmul_id_q5_k_f32_fp32_len, matmul_id_q5_k_f32_fp32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), l_wg_denoms, warptile_mmq_l, l_align);
+        ggml_vk_create_pipeline(ctx, ctx->device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q5_K]->m, "matmul_id_q5_k_f32_m", matmul_id_q5_k_f32_fp32_len, matmul_id_q5_k_f32_fp32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), m_wg_denoms, warptile_mmq_m, m_align);
+        ggml_vk_create_pipeline(ctx, ctx->device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q5_K]->s, "matmul_id_q5_k_f32_s", matmul_id_q5_k_f32_fp32_len, matmul_id_q5_k_f32_fp32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), s_wg_denoms, warptile_mmq_s, s_align);
+        ggml_vk_create_pipeline(ctx, ctx->device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q5_K]->a_l, "matmul_id_q5_k_f32_aligned_l", matmul_id_q5_k_f32_aligned_fp32_len, matmul_id_q5_k_f32_aligned_fp32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), l_wg_denoms, warptile_mmq_l, l_align);
+        ggml_vk_create_pipeline(ctx, ctx->device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q5_K]->a_m, "matmul_id_q5_k_f32_aligned_m", matmul_id_q5_k_f32_aligned_fp32_len, matmul_id_q5_k_f32_aligned_fp32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), m_wg_denoms, warptile_mmq_m, m_align);
+        ggml_vk_create_pipeline(ctx, ctx->device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q5_K]->a_s, "matmul_id_q5_k_f32_aligned_s", matmul_id_q5_k_f32_aligned_fp32_len, matmul_id_q5_k_f32_aligned_fp32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), s_wg_denoms, warptile_mmq_s, s_align);
+
+        ggml_vk_create_pipeline(ctx, ctx->device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q6_K]->l, "matmul_id_q6_k_f32_l", matmul_id_q6_k_f32_fp32_len, matmul_id_q6_k_f32_fp32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), l_wg_denoms, warptile_mmq_l, l_align);
+        ggml_vk_create_pipeline(ctx, ctx->device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q6_K]->m, "matmul_id_q6_k_f32_m", matmul_id_q6_k_f32_fp32_len, matmul_id_q6_k_f32_fp32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), m_wg_denoms, warptile_mmq_m, m_align);
+        ggml_vk_create_pipeline(ctx, ctx->device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q6_K]->s, "matmul_id_q6_k_f32_s", matmul_id_q6_k_f32_fp32_len, matmul_id_q6_k_f32_fp32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), s_wg_denoms, warptile_mmq_s, s_align);
+        ggml_vk_create_pipeline(ctx, ctx->device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q6_K]->a_l, "matmul_id_q6_k_f32_aligned_l", matmul_id_q6_k_f32_aligned_fp32_len, matmul_id_q6_k_f32_aligned_fp32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), l_wg_denoms, warptile_mmq_l, l_align);
+        ggml_vk_create_pipeline(ctx, ctx->device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q6_K]->a_m, "matmul_id_q6_k_f32_aligned_m", matmul_id_q6_k_f32_aligned_fp32_len, matmul_id_q6_k_f32_aligned_fp32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), m_wg_denoms, warptile_mmq_m, m_align);
+        ggml_vk_create_pipeline(ctx, ctx->device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q6_K]->a_s, "matmul_id_q6_k_f32_aligned_s", matmul_id_q6_k_f32_aligned_fp32_len, matmul_id_q6_k_f32_aligned_fp32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), s_wg_denoms, warptile_mmq_s, s_align);
     }
 
     // mul mat vec
+    ggml_vk_create_pipeline(ctx, ctx->device->pipeline_dequant_mul_mat_vec_f32_f32[GGML_TYPE_F32 ], "mul_mat_vec_f32_f32_f32",  mul_mat_vec_f32_f32_f32_len,  mul_mat_vec_f32_f32_f32_data,  "main", 3, sizeof(vk_mat_vec_push_constants), {1, 1, 1}, { device->subgroup_size }, 1);
     ggml_vk_create_pipeline(ctx, ctx->device->pipeline_dequant_mul_mat_vec_f32_f32[GGML_TYPE_F16 ], "mul_mat_vec_f16_f32_f32",  mul_mat_vec_f16_f32_f32_len,  mul_mat_vec_f16_f32_f32_data,  "main", 3, sizeof(vk_mat_vec_push_constants), {1, 1, 1}, { device->subgroup_size }, 1);
     ggml_vk_create_pipeline(ctx, ctx->device->pipeline_dequant_mul_mat_vec_f32_f32[GGML_TYPE_Q4_0], "mul_mat_vec_q4_0_f32_f32", mul_mat_vec_q4_0_f32_f32_len, mul_mat_vec_q4_0_f32_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {1, 1, 1}, { device->subgroup_size }, 1);
     ggml_vk_create_pipeline(ctx, ctx->device->pipeline_dequant_mul_mat_vec_f32_f32[GGML_TYPE_Q4_1], "mul_mat_vec_q4_1_f32_f32", mul_mat_vec_q4_1_f32_f32_len, mul_mat_vec_q4_1_f32_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {1, 1, 1}, { device->subgroup_size }, 1);
@@ -1435,6 +1448,7 @@ static void ggml_vk_load_shaders(ggml_backend_vk_context * ctx) {
     ggml_vk_create_pipeline(ctx, ctx->device->pipeline_dequant_mul_mat_vec_f32_f32[GGML_TYPE_Q5_K], "mul_mat_vec_q5_K_f32_f32", mul_mat_vec_q5_K_f32_f32_len, mul_mat_vec_q5_K_f32_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {1, 1, 1}, { device->subgroup_size }, 1);
     ggml_vk_create_pipeline(ctx, ctx->device->pipeline_dequant_mul_mat_vec_f32_f32[GGML_TYPE_Q6_K], "mul_mat_vec_q6_K_f32_f32", mul_mat_vec_q6_K_f32_f32_len, mul_mat_vec_q6_K_f32_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {1, 1, 1}, { device->subgroup_size }, 1);
 
+    ggml_vk_create_pipeline(ctx, ctx->device->pipeline_dequant_mul_mat_vec_f16_f32[GGML_TYPE_F32 ], "mul_mat_vec_f32_f16_f32",  mul_mat_vec_f32_f16_f32_len,  mul_mat_vec_f32_f16_f32_data,  "main", 3, sizeof(vk_mat_vec_push_constants), {1, 1, 1}, { device->subgroup_size }, 1);
     ggml_vk_create_pipeline(ctx, ctx->device->pipeline_dequant_mul_mat_vec_f16_f32[GGML_TYPE_F16 ], "mul_mat_vec_f16_f16_f32",  mul_mat_vec_f16_f16_f32_len,  mul_mat_vec_f16_f16_f32_data,  "main", 3, sizeof(vk_mat_vec_push_constants), {1, 1, 1}, { device->subgroup_size }, 1);
     ggml_vk_create_pipeline(ctx, ctx->device->pipeline_dequant_mul_mat_vec_f16_f32[GGML_TYPE_Q4_0], "mul_mat_vec_q4_0_f16_f32", mul_mat_vec_q4_0_f16_f32_len, mul_mat_vec_q4_0_f16_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {1, 1, 1}, { device->subgroup_size }, 1);
     ggml_vk_create_pipeline(ctx, ctx->device->pipeline_dequant_mul_mat_vec_f16_f32[GGML_TYPE_Q4_1], "mul_mat_vec_q4_1_f16_f32", mul_mat_vec_q4_1_f16_f32_len, mul_mat_vec_q4_1_f16_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {1, 1, 1}, { device->subgroup_size }, 1);
@@ -1447,17 +1461,18 @@ static void ggml_vk_load_shaders(ggml_backend_vk_context * ctx) {
     ggml_vk_create_pipeline(ctx, ctx->device->pipeline_dequant_mul_mat_vec_f16_f32[GGML_TYPE_Q5_K], "mul_mat_vec_q5_K_f16_f32", mul_mat_vec_q5_K_f16_f32_len, mul_mat_vec_q5_K_f16_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {1, 1, 1}, { device->subgroup_size }, 1);
     ggml_vk_create_pipeline(ctx, ctx->device->pipeline_dequant_mul_mat_vec_f16_f32[GGML_TYPE_Q6_K], "mul_mat_vec_q6_K_f16_f32", mul_mat_vec_q6_K_f16_f32_len, mul_mat_vec_q6_K_f16_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {1, 1, 1}, { device->subgroup_size }, 1);
 
-    /*ggml_vk_create_pipeline(ctx, ctx->device->pipeline_dequant_mul_mat_vec_id_f32[GGML_TYPE_F16 ], "mul_mat_vec_id_f16_f32",  mul_mat_vec_id_f16_f32_len,  mul_mat_vec_id_f16_f32_data,  "main", 4, sizeof(vk_mat_vec_push_constants), {1, 1, 1}, { device->subgroup_size }, 1);
-    ggml_vk_create_pipeline(ctx, ctx->device->pipeline_dequant_mul_mat_vec_id_f32[GGML_TYPE_Q4_0], "mul_mat_vec_id_q4_0_f32", mul_mat_vec_id_q4_0_f32_len, mul_mat_vec_id_q4_0_f32_data, "main", 4, sizeof(vk_mat_vec_push_constants), {1, 1, 1}, { device->subgroup_size }, 1);
-    ggml_vk_create_pipeline(ctx, ctx->device->pipeline_dequant_mul_mat_vec_id_f32[GGML_TYPE_Q4_1], "mul_mat_vec_id_q4_1_f32", mul_mat_vec_id_q4_1_f32_len, mul_mat_vec_id_q4_1_f32_data, "main", 4, sizeof(vk_mat_vec_push_constants), {1, 1, 1}, { device->subgroup_size }, 1);
-    ggml_vk_create_pipeline(ctx, ctx->device->pipeline_dequant_mul_mat_vec_id_f32[GGML_TYPE_Q5_0], "mul_mat_vec_id_q5_0_f32", mul_mat_vec_id_q5_0_f32_len, mul_mat_vec_id_q5_0_f32_data, "main", 4, sizeof(vk_mat_vec_push_constants), {1, 1, 1}, { device->subgroup_size }, 1);
-    ggml_vk_create_pipeline(ctx, ctx->device->pipeline_dequant_mul_mat_vec_id_f32[GGML_TYPE_Q5_1], "mul_mat_vec_id_q5_1_f32", mul_mat_vec_id_q5_1_f32_len, mul_mat_vec_id_q5_1_f32_data, "main", 4, sizeof(vk_mat_vec_push_constants), {1, 1, 1}, { device->subgroup_size }, 1);
-    ggml_vk_create_pipeline(ctx, ctx->device->pipeline_dequant_mul_mat_vec_id_f32[GGML_TYPE_Q8_0], "mul_mat_vec_id_q8_0_f32", mul_mat_vec_id_q8_0_f32_len, mul_mat_vec_id_q8_0_f32_data, "main", 4, sizeof(vk_mat_vec_push_constants), {1, 1, 1}, { device->subgroup_size }, 1);
-    ggml_vk_create_pipeline(ctx, ctx->device->pipeline_dequant_mul_mat_vec_id_f32[GGML_TYPE_Q2_K], "mul_mat_vec_id_q2_K_f32", mul_mat_vec_id_q2_K_f32_len, mul_mat_vec_id_q2_K_f32_data, "main", 4, sizeof(vk_mat_vec_push_constants), {1, 1, 1}, { device->subgroup_size }, 1);
-    ggml_vk_create_pipeline(ctx, ctx->device->pipeline_dequant_mul_mat_vec_id_f32[GGML_TYPE_Q3_K], "mul_mat_vec_id_q3_K_f32", mul_mat_vec_id_q3_K_f32_len, mul_mat_vec_id_q3_K_f32_data, "main", 4, sizeof(vk_mat_vec_push_constants), {1, 1, 1}, { device->subgroup_size }, 1);
-    ggml_vk_create_pipeline(ctx, ctx->device->pipeline_dequant_mul_mat_vec_id_f32[GGML_TYPE_Q4_K], "mul_mat_vec_id_q4_K_f32", mul_mat_vec_id_q4_K_f32_len, mul_mat_vec_id_q4_K_f32_data, "main", 4, sizeof(vk_mat_vec_push_constants), {1, 1, 1}, { device->subgroup_size }, 1);
-    ggml_vk_create_pipeline(ctx, ctx->device->pipeline_dequant_mul_mat_vec_id_f32[GGML_TYPE_Q5_K], "mul_mat_vec_id_q5_K_f32", mul_mat_vec_id_q5_K_f32_len, mul_mat_vec_id_q5_K_f32_data, "main", 4, sizeof(vk_mat_vec_push_constants), {1, 1, 1}, { device->subgroup_size }, 1);
-    ggml_vk_create_pipeline(ctx, ctx->device->pipeline_dequant_mul_mat_vec_id_f32[GGML_TYPE_Q6_K], "mul_mat_vec_id_q6_K_f32", mul_mat_vec_id_q6_K_f32_len, mul_mat_vec_id_q6_K_f32_data, "main", 4, sizeof(vk_mat_vec_push_constants), {1, 1, 1}, { device->subgroup_size }, 1);*/
+    ggml_vk_create_pipeline(ctx, ctx->device->pipeline_dequant_mul_mat_vec_id_f32[GGML_TYPE_F32 ], "mul_mat_vec_id_f32_f32",  mul_mat_vec_id_f32_f32_len,  mul_mat_vec_id_f32_f32_data,  "main", 4, sizeof(vk_mat_vec_id_push_constants), {1, 1, 1}, { device->subgroup_size }, 1);
+    ggml_vk_create_pipeline(ctx, ctx->device->pipeline_dequant_mul_mat_vec_id_f32[GGML_TYPE_F16 ], "mul_mat_vec_id_f16_f32",  mul_mat_vec_id_f16_f32_len,  mul_mat_vec_id_f16_f32_data,  "main", 4, sizeof(vk_mat_vec_id_push_constants), {1, 1, 1}, { device->subgroup_size }, 1);
+    ggml_vk_create_pipeline(ctx, ctx->device->pipeline_dequant_mul_mat_vec_id_f32[GGML_TYPE_Q4_0], "mul_mat_vec_id_q4_0_f32", mul_mat_vec_id_q4_0_f32_len, mul_mat_vec_id_q4_0_f32_data, "main", 4, sizeof(vk_mat_vec_id_push_constants), {1, 1, 1}, { device->subgroup_size }, 1);
+    ggml_vk_create_pipeline(ctx, ctx->device->pipeline_dequant_mul_mat_vec_id_f32[GGML_TYPE_Q4_1], "mul_mat_vec_id_q4_1_f32", mul_mat_vec_id_q4_1_f32_len, mul_mat_vec_id_q4_1_f32_data, "main", 4, sizeof(vk_mat_vec_id_push_constants), {1, 1, 1}, { device->subgroup_size }, 1);
+    ggml_vk_create_pipeline(ctx, ctx->device->pipeline_dequant_mul_mat_vec_id_f32[GGML_TYPE_Q5_0], "mul_mat_vec_id_q5_0_f32", mul_mat_vec_id_q5_0_f32_len, mul_mat_vec_id_q5_0_f32_data, "main", 4, sizeof(vk_mat_vec_id_push_constants), {1, 1, 1}, { device->subgroup_size }, 1);
+    ggml_vk_create_pipeline(ctx, ctx->device->pipeline_dequant_mul_mat_vec_id_f32[GGML_TYPE_Q5_1], "mul_mat_vec_id_q5_1_f32", mul_mat_vec_id_q5_1_f32_len, mul_mat_vec_id_q5_1_f32_data, "main", 4, sizeof(vk_mat_vec_id_push_constants), {1, 1, 1}, { device->subgroup_size }, 1);
+    ggml_vk_create_pipeline(ctx, ctx->device->pipeline_dequant_mul_mat_vec_id_f32[GGML_TYPE_Q8_0], "mul_mat_vec_id_q8_0_f32", mul_mat_vec_id_q8_0_f32_len, mul_mat_vec_id_q8_0_f32_data, "main", 4, sizeof(vk_mat_vec_id_push_constants), {1, 1, 1}, { device->subgroup_size }, 1);
+    ggml_vk_create_pipeline(ctx, ctx->device->pipeline_dequant_mul_mat_vec_id_f32[GGML_TYPE_Q2_K], "mul_mat_vec_id_q2_K_f32", mul_mat_vec_id_q2_K_f32_len, mul_mat_vec_id_q2_K_f32_data, "main", 4, sizeof(vk_mat_vec_id_push_constants), {1, 1, 1}, { device->subgroup_size }, 1);
+    ggml_vk_create_pipeline(ctx, ctx->device->pipeline_dequant_mul_mat_vec_id_f32[GGML_TYPE_Q3_K], "mul_mat_vec_id_q3_K_f32", mul_mat_vec_id_q3_K_f32_len, mul_mat_vec_id_q3_K_f32_data, "main", 4, sizeof(vk_mat_vec_id_push_constants), {1, 1, 1}, { device->subgroup_size }, 1);
+    ggml_vk_create_pipeline(ctx, ctx->device->pipeline_dequant_mul_mat_vec_id_f32[GGML_TYPE_Q4_K], "mul_mat_vec_id_q4_K_f32", mul_mat_vec_id_q4_K_f32_len, mul_mat_vec_id_q4_K_f32_data, "main", 4, sizeof(vk_mat_vec_id_push_constants), {1, 1, 1}, { device->subgroup_size }, 1);
+    ggml_vk_create_pipeline(ctx, ctx->device->pipeline_dequant_mul_mat_vec_id_f32[GGML_TYPE_Q5_K], "mul_mat_vec_id_q5_K_f32", mul_mat_vec_id_q5_K_f32_len, mul_mat_vec_id_q5_K_f32_data, "main", 4, sizeof(vk_mat_vec_id_push_constants), {1, 1, 1}, { device->subgroup_size }, 1);
+    ggml_vk_create_pipeline(ctx, ctx->device->pipeline_dequant_mul_mat_vec_id_f32[GGML_TYPE_Q6_K], "mul_mat_vec_id_q6_K_f32", mul_mat_vec_id_q6_K_f32_len, mul_mat_vec_id_q6_K_f32_data, "main", 4, sizeof(vk_mat_vec_id_push_constants), {1, 1, 1}, { device->subgroup_size }, 1);
 
     // dequant shaders
     ggml_vk_create_pipeline(ctx, ctx->device->pipeline_dequant[GGML_TYPE_F32 ], "f32_to_f16",   dequant_f32_len,  dequant_f32_data,  "main", 2, 5 * sizeof(uint32_t), {256 * 16, 1, 1}, {}, 1);
@@ -1505,6 +1520,8 @@ static void ggml_vk_load_shaders(ggml_backend_vk_context * ctx) {
 
     ggml_vk_create_pipeline(ctx, ctx->device->pipeline_mul_f32, "mul_f32", mul_f32_len, mul_f32_data, "main", 3, sizeof(vk_op_binary_push_constants), {512, 1, 1}, {}, 1);
 
+    ggml_vk_create_pipeline(ctx, ctx->device->pipeline_div_f32, "div_f32", div_f32_len, div_f32_data, "main", 3, sizeof(vk_op_binary_push_constants), {512, 1, 1}, {}, 1);
+
     ggml_vk_create_pipeline(ctx, ctx->device->pipeline_scale_f32, "scale_f32", scale_f32_len, scale_f32_data, "main", 2, sizeof(vk_op_unary_push_constants), {512, 1, 1}, {}, 1);
 
     ggml_vk_create_pipeline(ctx, ctx->device->pipeline_sqr_f32, "sqr_f32", sqr_f32_len, sqr_f32_data, "main", 2, sizeof(vk_op_unary_push_constants), {512, 1, 1}, {}, 1);
@@ -1527,6 +1544,8 @@ static void ggml_vk_load_shaders(ggml_backend_vk_context * ctx) {
     ggml_vk_create_pipeline(ctx, ctx->device->pipeline_rope_neox_f16, "rope_neox_f16", rope_neox_f16_len, rope_neox_f16_data, "main", 4, sizeof(vk_op_rope_neox_push_constants), {1, 512, 1}, {}, 1);
 
     ggml_vk_create_pipeline(ctx, ctx->device->pipeline_argsort_f32, "argsort_f32", argsort_f32_len, argsort_f32_data, "main", 2, sizeof(vk_op_argsort_push_constants), {1024, 1, 1}, {}, 1);
+
+    ggml_vk_create_pipeline(ctx, ctx->device->pipeline_sum_rows_f32, "sum_rows_f32", sum_rows_f32_len, sum_rows_f32_data, "main", 2, sizeof(vk_op_push_constants), {1, 1, 1}, { device->subgroup_size }, 1);
 }
 
 static void ggml_vk_print_gpu_info(size_t idx) {
@@ -1949,6 +1968,33 @@ static vk_matmul_pipeline ggml_vk_get_mul_mat_mat_pipeline(ggml_backend_vk_conte
     return ctx->device->pipeline_dequant_mul_mat_mat[src0_type];
 }
 
+static vk_pipeline ggml_vk_get_dequantize_mul_mat_vec(ggml_backend_vk_context * ctx, ggml_type a_type, ggml_type b_type) {
+#ifdef GGML_VULKAN_DEBUG
+    std::cerr << "ggml_vk_get_dequantize_mul_mat_vec()" << std::endl;
+#endif
+    GGML_ASSERT(b_type == GGML_TYPE_F32 || b_type == GGML_TYPE_F16);
+
+    switch (a_type) {
+        case GGML_TYPE_F32:
+        case GGML_TYPE_F16:
+        case GGML_TYPE_Q4_0:
+        case GGML_TYPE_Q4_1:
+        case GGML_TYPE_Q5_0:
+        case GGML_TYPE_Q5_1:
+        case GGML_TYPE_Q8_0:
+        case GGML_TYPE_Q2_K:
+        case GGML_TYPE_Q3_K:
+        case GGML_TYPE_Q4_K:
+        case GGML_TYPE_Q5_K:
+        case GGML_TYPE_Q6_K:
+            break;
+        default:
+            return nullptr;
+    }
+
+    return b_type == GGML_TYPE_F32 ? ctx->device->pipeline_dequant_mul_mat_vec_f32_f32[a_type] : ctx->device->pipeline_dequant_mul_mat_vec_f16_f32[a_type];
+}
+
 static vk_matmul_pipeline ggml_vk_get_mul_mat_mat_id_pipeline(ggml_backend_vk_context * ctx, ggml_type src0_type, ggml_type src1_type) {
 #ifdef GGML_VULKAN_DEBUG
     std::cerr << "ggml_vk_get_mul_mat_mat_id_pipeline()" << std::endl;
@@ -1984,13 +2030,14 @@ static vk_matmul_pipeline ggml_vk_get_mul_mat_mat_id_pipeline(ggml_backend_vk_co
     return ctx->device->pipeline_dequant_mul_mat_mat_id[src0_type];
 }
 
-static vk_pipeline ggml_vk_get_dequantize_mul_mat_vec(ggml_backend_vk_context * ctx, ggml_type a_type, ggml_type b_type) {
+static vk_pipeline ggml_vk_get_dequantize_mul_mat_vec_id(ggml_backend_vk_context * ctx, ggml_type a_type, ggml_type b_type) {
 #ifdef GGML_VULKAN_DEBUG
     std::cerr << "ggml_vk_get_dequantize_mul_mat_vec()" << std::endl;
 #endif
-    GGML_ASSERT(b_type == GGML_TYPE_F32 || b_type == GGML_TYPE_F16);
+    GGML_ASSERT(b_type == GGML_TYPE_F32);
 
     switch (a_type) {
+        case GGML_TYPE_F32:
         case GGML_TYPE_F16:
         case GGML_TYPE_Q4_0:
         case GGML_TYPE_Q4_1:
@@ -2007,7 +2054,7 @@ static vk_pipeline ggml_vk_get_dequantize_mul_mat_vec(ggml_backend_vk_context *
             return nullptr;
     }
 
-    return b_type == GGML_TYPE_F32 ? ctx->device->pipeline_dequant_mul_mat_vec_f32_f32[a_type] : ctx->device->pipeline_dequant_mul_mat_vec_f16_f32[a_type];
+    return ctx->device->pipeline_dequant_mul_mat_vec_id_f32[a_type];
 }
 
 static vk_buffer ggml_vk_pool_malloc(ggml_backend_vk_context * ctx, size_t size) {
@@ -2155,7 +2202,11 @@ static void ggml_vk_dispatch_pipeline(ggml_backend_vk_context * ctx, vk_context
     const uint32_t wg1 = CEIL_DIV(elements[1], pipeline->wg_denoms[1]);
     const uint32_t wg2 = CEIL_DIV(elements[2], pipeline->wg_denoms[2]);
 #ifdef GGML_VULKAN_DEBUG
-    std::cerr << "ggml_vk_dispatch_pipeline(" << pipeline->name << ", (" << wg0 << "," << wg1 << "," << wg2 << "))" << std::endl;
+    std::cerr << "ggml_vk_dispatch_pipeline(" << pipeline->name << ", {";
+    for (auto& buffer : buffers) {
+        std::cerr << "(" << buffer.buffer << ", " << buffer.offset << ", " << buffer.size << "), ";
+    }
+    std::cerr << "}, (" << wg0 << "," << wg1 << "," << wg2 << "))" << std::endl;
 #endif
     std::vector<vk::DescriptorBufferInfo> descriptor_buffer_infos;
     std::vector<vk::WriteDescriptorSet> write_descriptor_sets;
@@ -2736,22 +2787,21 @@ static void ggml_vk_matmul(
         ggml_backend_vk_context * ctx, vk_context * subctx, vk_pipeline& pipeline,
         vk_subbuffer&& a, vk_subbuffer&& b, vk_subbuffer&& d, vk_subbuffer&& split_k_buffer,
         uint32_t m, uint32_t n, uint32_t k, uint32_t stride_a, uint32_t stride_b, uint32_t stride_d,
-        uint32_t split_k, uint32_t batch, uint32_t ne02, uint32_t ne12, uint32_t broadcast2, uint32_t broadcast3,
         uint32_t batch_stride_a, uint32_t batch_stride_b, uint32_t batch_stride_d,
-        uint32_t expert_stride_b, uint32_t expert_stride_d, uint32_t idx, uint32_t nbi1, uint32_t n_as) {
+        uint32_t split_k, uint32_t batch, uint32_t ne02, uint32_t ne12, uint32_t broadcast2, uint32_t broadcast3) {
 #ifdef GGML_VULKAN_DEBUG
-    std::cerr << "ggml_vk_matmul(a: (" << a.buffer->buffer << ", " << a.offset << ", " << a.size << "), b: (" << b.buffer->buffer << ", " << b.offset << ", " << b.size << "), c: (" << d.buffer->buffer << ", " << d.offset << ", " << d.size << "), split_k: (" << (split_k_buffer.buffer != nullptr ? split_k_buffer.buffer->buffer : VK_NULL_HANDLE) << ", " << split_k_buffer.offset << ", " << split_k_buffer.size << "), m: " << m << ", n: " << n << ", k: " << k << ", stride_a: " << stride_a << ", stride_b: " << stride_b << ", stride_d: " << stride_d << ", split_k: " << split_k << ", batch: " << batch << ", ne02: " << ne02 << ", ne12: " << ne12 << ", broadcast2: " << broadcast2 << ", broadcast3: " << broadcast3 << ", batch_stride_a: " << batch_stride_a << ", batch_stride_b: " << batch_stride_b << ", batch_stride_d: " << batch_stride_d << ")" << std::endl;
+    std::cerr << "ggml_vk_matmul(a: (" << a.buffer->buffer << ", " << a.offset << ", " << a.size << "), b: (" << b.buffer->buffer << ", " << b.offset << ", " << b.size << "), d: (" << d.buffer->buffer << ", " << d.offset << ", " << d.size << "), split_k: (" << (split_k_buffer.buffer != nullptr ? split_k_buffer.buffer->buffer : VK_NULL_HANDLE) << ", " << split_k_buffer.offset << ", " << split_k_buffer.size << "), m: " << m << ", n: " << n << ", k: " << k << ", stride_a: " << stride_a << ", stride_b: " << stride_b << ", stride_d: " << stride_d << ", batch_stride_a: " << batch_stride_a << ", batch_stride_b: " << batch_stride_b << ", batch_stride_d: " << batch_stride_d << ", split_k: " << split_k << ", batch: " << batch << ", ne02: " << ne02 << ", ne12: " << ne12 << ", broadcast2: " << broadcast2 << ", broadcast3: " << broadcast3 << ")" << std::endl;
 #endif
     ggml_vk_sync_buffers(subctx);
     if (split_k == 1) {
-        const vk_mat_mat_push_constants pc = { m, n, k, stride_a, stride_b, stride_d, k, ne02, ne12, broadcast2, broadcast3, batch_stride_a, batch_stride_b, batch_stride_d, expert_stride_b, expert_stride_d, idx, nbi1, n_as };
+        const vk_mat_mat_push_constants pc = { m, n, k, stride_a, stride_b, stride_d, batch_stride_a, batch_stride_b, batch_stride_d, k, ne02, ne12, broadcast2, broadcast3 };
         ggml_vk_dispatch_pipeline(ctx, subctx, pipeline, { a, b, d }, sizeof(vk_mat_mat_push_constants), &pc, { m, n, batch });
         return;
     }
 
     GGML_ASSERT(batch_stride_d == m * n);
 
-    const vk_mat_mat_push_constants pc1 = { m, n, k, stride_a, stride_b, stride_d, CEIL_DIV(k, split_k), ne02, ne12, broadcast2, broadcast3, batch_stride_a, batch_stride_b, batch_stride_d, expert_stride_b, expert_stride_d, idx, nbi1, n_as };
+    const vk_mat_mat_push_constants pc1 = { m, n, k, stride_a, stride_b, stride_d, batch_stride_a, batch_stride_b, batch_stride_d, CEIL_DIV(k, split_k), ne02, ne12, broadcast2, broadcast3 };
     // Make sure enough workgroups get assigned for split k to work
     ggml_vk_dispatch_pipeline(ctx, subctx, pipeline, { a, b, split_k_buffer }, sizeof(vk_mat_mat_push_constants), &pc1, { (CEIL_DIV(m, pipeline->wg_denoms[0]) * pipeline->wg_denoms[0]) * split_k, n, batch });
     ggml_vk_sync_buffers(subctx);
@@ -2761,29 +2811,20 @@ static void ggml_vk_matmul(
 
 static void ggml_vk_matmul_id(
         ggml_backend_vk_context * ctx, vk_context * subctx, vk_pipeline& pipeline,
-        vk_subbuffer&& ids, vk_subbuffer&& b, vk_subbuffer&& d, vk_subbuffer&& a, vk_subbuffer&& split_k_buffer,
+        vk_subbuffer&& a, vk_subbuffer&& b, vk_subbuffer&& d, vk_subbuffer&& ids,
         uint32_t m, uint32_t n, uint32_t k, uint32_t stride_a, uint32_t stride_b, uint32_t stride_d,
-        uint32_t split_k, uint32_t batch, uint32_t ne02, uint32_t ne12, uint32_t broadcast2, uint32_t broadcast3,
         uint32_t batch_stride_a, uint32_t batch_stride_b, uint32_t batch_stride_d,
-        uint32_t expert_stride_b, uint32_t expert_stride_d, uint32_t idx, uint32_t nbi1, uint32_t n_as) {
+        uint32_t n_as, uint32_t nei0, uint32_t nei1, uint32_t nbi1, uint32_t ne11) {
 #ifdef GGML_VULKAN_DEBUG
-    std::cerr << "ggml_vk_matmul(a: (" << a.buffer->buffer << ", " << a.offset << ", " << a.size << "), b: (" << b.buffer->buffer << ", " << b.offset << ", " << b.size << "), c: (" << d.buffer->buffer << ", " << d.offset << ", " << d.size << "), split_k: (" << split_k_buffer.buffer->buffer << ", " << split_k_buffer.offset << ", " << split_k_buffer.size << "), m: " << m << ", n: " << n << ", k: " << k << ", stride_a: " << stride_a << ", stride_b: " << stride_b << ", stride_d: " << stride_d << ", split_k: " << split_k << ", batch: " << batch << ", ne02: " << ne02 << ", ne12: " << ne12 << ", broadcast2: " << broadcast2 << ", broadcast3: " << broadcast3 << ", batch_stride_a: " << batch_stride_a << ", batch_stride_b: " << batch_stride_b << ", batch_stride_d: " << batch_stride_d << ")" << std::endl;
+    std::cerr << "ggml_vk_matmul_id(a: (" << a.buffer->buffer << ", " << a.offset << ", " << a.size << "), b: (" << b.buffer->buffer << ", " << b.offset << ", " << b.size << "), d: (" << d.buffer->buffer << ", " << d.offset << ", " << d.size << "), ids: (" << ids.buffer->buffer << ", " << ids.offset << ", " << ids.size << "), " <<
+        "m: " << m << ", n: " << n << ", k: " << k << ", stride_a: " << stride_a << ", stride_b: " << stride_b << ", stride_d: " << stride_d << ", " <<
+        "batch_stride_a: " << batch_stride_a << ", batch_stride_b: " << batch_stride_b << ", batch_stride_d: " << batch_stride_d << ", " <<
+        "n_as: " << n_as << ", nei0: " << nei0 << ", nei1: " << nei1 << ", nbi1: " << nbi1 << ", ne11: " << ne11 << ")" << std::endl;
 #endif
     ggml_vk_sync_buffers(subctx);
-    if (split_k == 1) {
-        const vk_mat_mat_push_constants pc = { m, n, k, stride_a, stride_b, stride_d, k, ne02, ne12, broadcast2, broadcast3, batch_stride_a, batch_stride_b, batch_stride_d, expert_stride_b, expert_stride_d, idx, nbi1, n_as };
-        ggml_vk_dispatch_pipeline(ctx, subctx, pipeline, { ids, b, d, a }, sizeof(vk_mat_mat_push_constants), &pc, { m, n, batch });
-        return;
-    }
-
-    GGML_ASSERT(batch_stride_d == m * n);
-
-    const vk_mat_mat_push_constants pc1 = { m, n, k, stride_a, stride_b, stride_d, CEIL_DIV(k, split_k), ne02, ne12, broadcast2, broadcast3, batch_stride_a, batch_stride_b, batch_stride_d, expert_stride_b, expert_stride_d, idx, nbi1, n_as };
-    // Make sure enough workgroups get assigned for split k to work
-    ggml_vk_dispatch_pipeline(ctx, subctx, pipeline, { ids, b, split_k_buffer, a }, sizeof(vk_mat_mat_push_constants), &pc1, { (CEIL_DIV(m, pipeline->wg_denoms[0]) * pipeline->wg_denoms[0]) * split_k, n, batch });
-    ggml_vk_sync_buffers(subctx);
-    const std::array<uint32_t, 2> pc2 = { (uint32_t)(m * n * batch), split_k };
-    ggml_vk_dispatch_pipeline(ctx, subctx, ctx->device->pipeline_matmul_split_k_reduce, { split_k_buffer, d }, pc2.size() * sizeof(uint32_t), pc2.data(), { m * n * batch, 1, 1 });
+    const vk_mat_mat_id_push_constants pc = { m, n, k, stride_a, stride_b, stride_d, batch_stride_a, batch_stride_b, batch_stride_d,
+                                              nei0, nei1, nbi1, ne11 };
+    ggml_vk_dispatch_pipeline(ctx, subctx, pipeline, { a, b, d, ids }, sizeof(vk_mat_mat_id_push_constants), &pc, { m, nei1, n_as });
 }
 
 static bool ggml_vk_dim01_contiguous(const ggml_tensor * tensor) {
@@ -2997,8 +3038,9 @@ static void ggml_vk_mul_mat_q_f16(ggml_backend_vk_context * ctx, vk_context * su
         ctx, subctx, pipeline,
         { d_X, x_buf_offset, x_sz * ne02 * ne03 }, { d_Y, y_buf_offset, y_sz * ne12 * ne13 },
         { d_D, d_buf_offset, d_sz * ne12 * ne13 }, { ctx->prealloc_split_k, 0, d_sz * ne12 * ne13 * split_k },
-        ne01, ne11, ne10, ne10, ne10, ne01, split_k, ne12*ne13, ne02, ne12, r2, r3, stride_batch_x, stride_batch_y, ne20*ne21,
-        0, 0, 0, 0, 1
+        ne01, ne11, ne10,
+        ne10, ne10, ne01, stride_batch_x, stride_batch_y, ne20*ne21,
+        split_k, ne12*ne13, ne02, ne12, r2, r3
     );  // NOLINT
 }
 
@@ -3150,8 +3192,8 @@ static void ggml_vk_mul_mat_vec_q_f16(ggml_backend_vk_context * ctx, vk_context
     // compute
     const vk_mat_vec_push_constants pc = {
         (uint32_t)ne00, (uint32_t)ne10, (uint32_t)ne10, (uint32_t)ne01,
-        (uint32_t)ne02, (uint32_t)ne12, (uint32_t)r2, (uint32_t)r3,
         stride_batch_x, stride_batch_y, (uint32_t)(ne20*ne21),
+        (uint32_t)ne02, (uint32_t)ne12, (uint32_t)r2, (uint32_t)r3,
     };
     ggml_vk_sync_buffers(subctx);
     ggml_vk_dispatch_pipeline(ctx, subctx, dmmv, { { d_X, x_buf_offset, x_sz * ne02 * ne03 }, { d_Y, y_buf_offset, y_sz * ne12 * ne13 }, { d_D, d_buf_offset, d_sz * ne22 * ne23} }, sizeof(vk_mat_vec_push_constants), &pc, { (uint32_t)ne01, (uint32_t)(ne12 * ne13), 1});
@@ -3311,26 +3353,26 @@ static void ggml_vk_mul_mat(ggml_backend_vk_context * ctx, vk_context * subctx,
 #ifdef GGML_VULKAN_DEBUG
     std::cerr << "ggml_vk_mul_mat(" << src0 << ", " << src1 << ", " << dst << ")" << std::endl;
 #endif
-    if (src0->type == GGML_TYPE_F16 && ggml_is_permuted(src0) && ggml_is_permuted(src1) && src1->ne[1] == 1) {
+    if (src0->type == GGML_TYPE_F16 && ggml_is_permuted(src0) && ggml_is_permuted(src1) && dst->ne[1] == 1) {
         ggml_vk_mul_mat_vec_p021_f16_f32(ctx, subctx, src0, src1, dst);
-    } else if (src0->type == GGML_TYPE_F16 && !ggml_is_contiguous(src0) && !ggml_is_transposed(src1) && src1->ne[1] == 1) {
+    } else if (src0->type == GGML_TYPE_F16 && !ggml_is_contiguous(src0) && !ggml_is_transposed(src1) && dst->ne[1] == 1) {
         ggml_vk_mul_mat_vec_nc_f16_f32(ctx, subctx, src0, src1, dst);
-    } else if (src1->ne[1] == 1 && (src0->type == GGML_TYPE_F16 || ggml_is_quantized(src0->type))) {
+    } else if (dst->ne[1] == 1 && (src0->type == GGML_TYPE_F32 || src0->type == GGML_TYPE_F16 || ggml_is_quantized(src0->type))) {
         ggml_vk_mul_mat_vec_q_f16(ctx, subctx, src0, src1, dst);
     } else {
         ggml_vk_mul_mat_q_f16(ctx, subctx, src0, src1, dst);
     }
 }
 
-/*static void ggml_vk_mul_mat_id_q_f16(ggml_backend_vk_context * ctx, vk_context * subctx, const ggml_tensor * src0, const ggml_tensor * src1, const ggml_tensor * ids, ggml_tensor * dst) {
+static void ggml_vk_mul_mat_id_q_f16(ggml_backend_vk_context * ctx, vk_context * subctx, const ggml_tensor * src0, const ggml_tensor * src1, const ggml_tensor * ids, ggml_tensor * dst) {
 #ifdef GGML_VULKAN_DEBUG
-    std::cerr << "ggml_vk_mul_mat_id_q_f16((" << src0 << ", name=" << src0->name << ", type=" << src0->type << ", backend=" << src0->backend << ", ne0=" << src0->ne[0] << ", ne1=" << src0->ne[1] << ", ne2=" << src0->ne[2] << ", ne3=" << src0->ne[3] << ", nb0=" << src0->nb[0] << ", nb1=" << src0->nb[1] << ", nb2=" << src0->nb[2] << ", nb3=" << src0->nb[3];
-    std::cerr << "), (" << src1 << ", name=" << src1->name << ", type=" << src1->type << ", backend=" << src1->backend << ", ne0=" << src1->ne[0] << ", ne1=" << src1->ne[1] << ", ne2=" << src1->ne[2] << ", ne3=" << src1->ne[3] << ", nb0=" << src1->nb[0] << ", nb1=" << src1->nb[1] << ", nb2=" << src1->nb[2] << ", nb3=" << src1->nb[3];
-    std::cerr << "), (" << ids << ", name=" << ids->name << ", type=" << ids->type << ", backend=" << ids->backend << ", ne0=" << ids->ne[0] << ", ne1=" << ids->ne[1] << ", ne2=" << ids->ne[2] << ", ne3=" << ids->ne[3] << ", nb0=" << ids->nb[0] << ", nb1=" << ids->nb[1] << ", nb2=" << ids->nb[2] << ", nb3=" << ids->nb[3];
-    std::cerr << "), (" << dst << ", name=" << dst->name << ", type=" << dst->type << ", backend=" << dst->backend << ", ne0=" << dst->ne[0] << ", ne1=" << dst->ne[1] << ", ne2=" << dst->ne[2] << ", ne3=" << dst->ne[3] << ", nb0=" << dst->nb[0] << ", nb1=" << dst->nb[1] << ", nb2=" << dst->nb[2] << ", nb3=" << dst->nb[3] << "),)" << std::endl;
+    std::cerr << "ggml_vk_mul_mat_id_q_f16((" << src0 << ", name=" << src0->name << ", type=" << src0->type << ", ne0=" << src0->ne[0] << ", ne1=" << src0->ne[1] << ", ne2=" << src0->ne[2] << ", ne3=" << src0->ne[3] << ", nb0=" << src0->nb[0] << ", nb1=" << src0->nb[1] << ", nb2=" << src0->nb[2] << ", nb3=" << src0->nb[3];
+    std::cerr << "), (" << src1 << ", name=" << src1->name << ", type=" << src1->type << ", ne0=" << src1->ne[0] << ", ne1=" << src1->ne[1] << ", ne2=" << src1->ne[2] << ", ne3=" << src1->ne[3] << ", nb0=" << src1->nb[0] << ", nb1=" << src1->nb[1] << ", nb2=" << src1->nb[2] << ", nb3=" << src1->nb[3];
+    std::cerr << "), (" << ids << ", name=" << ids->name << ", type=" << ids->type << ", ne0=" << ids->ne[0] << ", ne1=" << ids->ne[1] << ", ne2=" << ids->ne[2] << ", ne3=" << ids->ne[3] << ", nb0=" << ids->nb[0] << ", nb1=" << ids->nb[1] << ", nb2=" << ids->nb[2] << ", nb3=" << ids->nb[3];
+    std::cerr << "), (" << dst << ", name=" << dst->name << ", type=" << dst->type << ", ne0=" << dst->ne[0] << ", ne1=" << dst->ne[1] << ", ne2=" << dst->ne[2] << ", ne3=" << dst->ne[3] << ", nb0=" << dst->nb[0] << ", nb1=" << dst->nb[1] << ", nb2=" << dst->nb[2] << ", nb3=" << dst->nb[3] << "),)" << std::endl;
 #endif
-    GGML_ASSERT(src0->type == GGML_TYPE_I32);
     GGML_ASSERT(ggml_vk_dim01_contiguous(src1) || src1->type == GGML_TYPE_F32 || src1->type == GGML_TYPE_F16);  // NOLINT
+    GGML_ASSERT(ids->type == GGML_TYPE_I32);
 
     const uint64_t ne00 = src0->ne[0];
     const uint64_t ne01 = src0->ne[1];
@@ -3342,16 +3384,18 @@ static void ggml_vk_mul_mat(ggml_backend_vk_context * ctx, vk_context * subctx,
     const uint64_t ne12 = src1->ne[2];
     const uint64_t ne13 = src1->ne[3];
 
-    const uint32_t nb11 = src1->nb[1];
+    const uint64_t nei0 = ids->ne[0];
+    const uint64_t nei1 = ids->ne[1];
+    GGML_ASSERT(nei0 * nei1 <= 2048);
+
+    const uint32_t nbi1 = ids->nb[1];
+    const uint32_t nbi2 = ids->nb[2];
 
     const uint64_t ne20 = dst->ne[0];
     const uint64_t ne21 = dst->ne[1];
+    const uint64_t ne22 = dst->ne[2];
+    const uint64_t ne23 = dst->ne[3];
 
-    const uint64_t r2 = ne12 / ne02;
-    const uint64_t r3 = ne13 / ne03;
-
-    const uint32_t nbi1 = src0->nb[1];
-    const uint32_t idx = ((uint32_t *) dst->op_params)[0];
     const uint64_t n_as = ne02;
 
     GGML_ASSERT(n_as <= 8);
@@ -3365,15 +3409,20 @@ static void ggml_vk_mul_mat(ggml_backend_vk_context * ctx, vk_context * subctx,
     size_t qx_buf_offset = 0;
     vk_buffer d_Qy;
     size_t qy_buf_offset = 0;
+    vk_buffer d_ids;
+    size_t ids_buf_offset = 0;
 
     bool src0_uma = false;
     bool src1_uma = false;
+    bool ids_uma = false;
 
     if (ctx->device->uma) {
         ggml_vk_host_get(ctx, src0->data, d_Qx, qx_buf_offset);
         ggml_vk_host_get(ctx, src1->data, d_Qy, qy_buf_offset);
+        ggml_vk_host_get(ctx, ids->data, d_ids, ids_buf_offset);
         src0_uma = d_Qx != nullptr;
         src1_uma = d_Qy != nullptr;
+        ids_uma = d_ids != nullptr;
     }
 
     const bool x_non_contig = !ggml_vk_dim01_contiguous(src0);
@@ -3393,27 +3442,25 @@ static void ggml_vk_mul_mat(ggml_backend_vk_context * ctx, vk_context * subctx,
     // Not implemented
     GGML_ASSERT(y_non_contig || !qy_needs_dequant);  // NOLINT
 
-    const int x_ne = ne01 * ne00;
-    const int y_ne = ne11 * ne10;
-    const int d_ne = ne11 * ne01;
-
-    const uint32_t kpad = ggml_vk_align_size(ne10, ggml_vk_guess_matmul_pipeline_align(ctx, mmp, ne01, ne11));
-    const bool aligned = ne10 == kpad && ne01 > 8 && ne11 > 8;
+    const uint64_t x_ne = ne01 * ne00;
+    const uint64_t y_ne = ne11 * ne10;
+    const uint64_t d_ne = ne21 * ne20;
 
-    const uint32_t split_k = ggml_vk_guess_split_k(ne01, ne11, ne10);
+    const uint32_t kpad = ggml_vk_align_size(ne10, ggml_vk_guess_matmul_pipeline_align(ctx, mmp, ne01, nei1));
+    const bool aligned = ne10 == kpad && ne01 > 8 && nei1 > 8;
 
-    vk_pipeline pipeline = ggml_vk_guess_matmul_pipeline(ctx, mmp, ne01, ne11, aligned);
+    vk_pipeline pipeline = ggml_vk_guess_matmul_pipeline(ctx, mmp, ne01, nei1, aligned);
 
     const uint64_t qx_sz = ggml_type_size(src0->type) * x_ne / ggml_blck_size(src0->type);
     const uint64_t qy_sz = ggml_type_size(src1->type) * y_ne / ggml_blck_size(src1->type);
     const uint64_t x_sz = !qx_needs_dequant ? qx_sz : sizeof(ggml_fp16_t) * x_ne;
     const uint64_t y_sz = y_f32_kernel ? sizeof(float) * y_ne : sizeof(ggml_fp16_t) * y_ne;
+    const uint64_t ids_sz = nbi2;
     const uint64_t d_sz = sizeof(float) * d_ne;
 
     vk_buffer d_D = extra->buffer_gpu.lock();
     const uint64_t d_buf_offset = extra->offset;
     GGML_ASSERT(d_D != nullptr);
-    GGML_ASSERT(d_D->size >= d_buf_offset + d_sz * ne02 * ne03);
     vk_buffer d_X;
     uint64_t x_buf_offset = 0;
     vk_buffer d_Y;
@@ -3428,6 +3475,11 @@ static void ggml_vk_mul_mat(ggml_backend_vk_context * ctx, vk_context * subctx,
         qy_buf_offset = extra_src1->offset;
         GGML_ASSERT(d_Qy != nullptr);
     }
+    if (!ids_uma) {
+        d_ids = extra_ids->buffer_gpu.lock();
+        ids_buf_offset = extra_ids->offset;
+        GGML_ASSERT(d_ids != nullptr);
+    }
     if (qx_needs_dequant) {
         d_X = ctx->prealloc_x;
         GGML_ASSERT(d_X->size >= x_sz * ne02 * ne03);
@@ -3469,9 +3521,6 @@ static void ggml_vk_mul_mat(ggml_backend_vk_context * ctx, vk_context * subctx,
     if (qy_needs_dequant) {
         ggml_pipeline_allocate_descriptor_sets(ctx, to_fp16_vk_1, 1);
     }
-    if (split_k > 1) {
-        ggml_pipeline_allocate_descriptor_sets(ctx, ctx->device->pipeline_matmul_split_k_reduce, 1);
-    }
 
     if (x_non_contig) {
         ggml_vk_cpy_to_contiguous(ctx, subctx, to_fp16_vk_0, src0, { d_Qx, qx_buf_offset, VK_WHOLE_SIZE }, { d_X, 0, VK_WHOLE_SIZE });
@@ -3496,23 +3545,26 @@ static void ggml_vk_mul_mat(ggml_backend_vk_context * ctx, vk_context * subctx,
     }
 
     // compute
-    ggml_vk_matmul(
+    ggml_vk_matmul_id(
         ctx, subctx, pipeline,
         { d_X, x_buf_offset, x_sz * ne02 * ne03 }, { d_Y, y_buf_offset, y_sz * ne12 * ne13 },
-        { d_D, d_buf_offset, d_sz * ne12 * ne13 }, { ctx->prealloc_split_k, 0, d_sz * ne12 * ne13 * split_k },
-        ne01, ne11, ne10, ne10, ne10, ne01, split_k, ne12*ne13, ne02, ne12, r2, r3, stride_batch_x, stride_batch_y, ne20*ne21,
-        nb11 / ggml_type_size(src1->type), ne20, idx, nbi1, n_as
+        { d_D, d_buf_offset, d_sz * ne22 * ne23 }, { d_ids, ids_buf_offset, ids_sz },
+        ne01, ne21, ne10, ne10, ne10, ne01,
+        stride_batch_x, stride_batch_y, ne20*ne21,
+        n_as, nei0, nei1, nbi1 / ggml_type_size(ids->type), ne11
     );  // NOLINT
 }
 
-static void ggml_vk_mul_mat_vec_id_q_f16(ggml_backend_vk_context * ctx, vk_context * subctx, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
+static void ggml_vk_mul_mat_vec_id_q_f16(ggml_backend_vk_context * ctx, vk_context * subctx, const ggml_tensor * src0, const ggml_tensor * src1, const ggml_tensor * ids, ggml_tensor * dst) {
 #ifdef GGML_VULKAN_DEBUG
-    std::cerr << "ggml_vk_mul_mat_vec_id_q_f16((" << src0 << ", name=" << src0->name << ", type=" << src0->type << ",  backend=" << src0->backend << ", ne0=" << src0->ne[0] << ", ne1=" << src0->ne[1] << ", ne2=" << src0->ne[2] << ", ne3=" << src0->ne[3] << ", nb0=" << src0->nb[0] << ", nb1=" << src0->nb[1] << ", nb2=" << src0->nb[2] << ", nb3=" << src0->nb[3];
-    std::cerr << "), (" << src1 << ", name=" << src1->name << ", type=" << src1->type << ",  backend=" << src1->backend << ", ne0=" << src1->ne[0] << ", ne1=" << src1->ne[1] << ", ne2=" << src1->ne[2] << ", ne3=" << src1->ne[3] << ", nb0=" << src1->nb[0] << ", nb1=" << src1->nb[1] << ", nb2=" << src1->nb[2] << ", nb3=" << src1->nb[3];
-    std::cerr << "), (" << dst << ", name=" << dst->name << ", type=" << dst->type << ",  backend=" << dst->backend << ", ne0=" << dst->ne[0] << ", ne1=" << dst->ne[1] << ", ne2=" << dst->ne[2] << ", ne3=" << dst->ne[3] << ", nb0=" << dst->nb[0] << ", nb1=" << dst->nb[1] << ", nb2=" << dst->nb[2] << ", nb3=" << dst->nb[3] << "),)" << std::endl;
+    std::cerr << "ggml_vk_mul_mat_vec_id_q_f16((" << src0 << ", name=" << src0->name << ", type=" << src0->type << ", ne0=" << src0->ne[0] << ", ne1=" << src0->ne[1] << ", ne2=" << src0->ne[2] << ", ne3=" << src0->ne[3] << ", nb0=" << src0->nb[0] << ", nb1=" << src0->nb[1] << ", nb2=" << src0->nb[2] << ", nb3=" << src0->nb[3];
+    std::cerr << "), (" << src1 << ", name=" << src1->name << ", type=" << src1->type << ", ne0=" << src1->ne[0] << ", ne1=" << src1->ne[1] << ", ne2=" << src1->ne[2] << ", ne3=" << src1->ne[3] << ", nb0=" << src1->nb[0] << ", nb1=" << src1->nb[1] << ", nb2=" << src1->nb[2] << ", nb3=" << src1->nb[3];
+    std::cerr << "), (" << ids << ", name=" << ids->name << ", type=" << ids->type << ", ne0=" << ids->ne[0] << ", ne1=" << ids->ne[1] << ", ne2=" << ids->ne[2] << ", ne3=" << ids->ne[3] << ", nb0=" << ids->nb[0] << ", nb1=" << ids->nb[1] << ", nb2=" << ids->nb[2] << ", nb3=" << ids->nb[3];
+    std::cerr << "), (" << dst << ", name=" << dst->name << ", type=" << dst->type << ", ne0=" << dst->ne[0] << ", ne1=" << dst->ne[1] << ", ne2=" << dst->ne[2] << ", ne3=" << dst->ne[3] << ", nb0=" << dst->nb[0] << ", nb1=" << dst->nb[1] << ", nb2=" << dst->nb[2] << ", nb3=" << dst->nb[3] << "),)" << std::endl;
 #endif
     GGML_ASSERT(ggml_vk_dim01_contiguous(src0) || src0->type == GGML_TYPE_F32 || src0->type == GGML_TYPE_F16);  // NOLINT
     GGML_ASSERT(ggml_vk_dim01_contiguous(src1) || src1->type == GGML_TYPE_F32 || src1->type == GGML_TYPE_F16);  // NOLINT
+    GGML_ASSERT(ids->type == GGML_TYPE_I32);
 
     const uint64_t ne00 = src0->ne[0];
     const uint64_t ne01 = src0->ne[1];
@@ -3524,36 +3576,41 @@ static void ggml_vk_mul_mat_vec_id_q_f16(ggml_backend_vk_context * ctx, vk_conte
     const uint64_t ne12 = src1->ne[2];
     const uint64_t ne13 = src1->ne[3];
 
-    GGML_ASSERT(ne11 == 1);
+    const uint64_t nei0 = ids->ne[0];
+    const uint64_t nei1 = ids->ne[1];
+
+    const uint64_t nbi2 = ids->nb[2];
+
+    GGML_ASSERT(nei1 == 1);
 
     const uint64_t ne20 = dst->ne[0];
     const uint64_t ne21 = dst->ne[1];
     const uint64_t ne22 = dst->ne[2];
     const uint64_t ne23 = dst->ne[3];
 
-    const uint64_t nb22  = dst->nb[2];
-    const uint64_t nb23  = dst->nb[3];
-
-    const uint64_t r2 = ne12 / ne02;
-    const uint64_t r3 = ne13 / ne03;
-
     ggml_tensor_extra_gpu * extra = (ggml_tensor_extra_gpu *) dst->extra;
     ggml_tensor_extra_gpu * extra_src0 = (ggml_tensor_extra_gpu *) src0->extra;
     ggml_tensor_extra_gpu * extra_src1 = (ggml_tensor_extra_gpu *) src1->extra;
+    ggml_tensor_extra_gpu * extra_ids = (ggml_tensor_extra_gpu *) ids->extra;
 
     vk_buffer d_Qx;
     size_t qx_buf_offset = 0;
     vk_buffer d_Qy;
     size_t qy_buf_offset = 0;
+    vk_buffer d_ids;
+    size_t ids_buf_offset = 0;
 
     bool src0_uma = false;
     bool src1_uma = false;
+    bool ids_uma = false;
 
     if (ctx->device->uma) {
         ggml_vk_host_get(ctx, src0->data, d_Qx, qx_buf_offset);
         ggml_vk_host_get(ctx, src1->data, d_Qy, qy_buf_offset);
+        ggml_vk_host_get(ctx, ids->data, d_ids, ids_buf_offset);
         src0_uma = d_Qx != nullptr;
         src1_uma = d_Qy != nullptr;
+        ids_uma = d_ids != nullptr;
     }
 
     const bool x_non_contig = !ggml_vk_dim01_contiguous(src0);
@@ -3569,12 +3626,13 @@ static void ggml_vk_mul_mat_vec_id_q_f16(ggml_backend_vk_context * ctx, vk_conte
 
     const uint64_t x_ne = ne01 * ne00;
     const uint64_t y_ne = ne11 * ne10;
-    const uint64_t d_ne = ne11 * ne01;
+    const uint64_t d_ne = ne21 * ne20;
 
     const uint64_t qx_sz = ggml_vk_align_size(ggml_type_size(src0->type) * x_ne / ggml_blck_size(src0->type), ctx->device->properties.limits.minStorageBufferOffsetAlignment);
     const uint64_t qy_sz = ggml_type_size(src1->type) * y_ne / ggml_blck_size(src1->type);
     const uint64_t x_sz = x_non_contig ? ggml_vk_align_size(ggml_type_size(src0->type) * x_ne, ctx->device->properties.limits.minStorageBufferOffsetAlignment) : qx_sz;
     const uint64_t y_sz = f16_f32_kernel ? sizeof(float) * y_ne : sizeof(ggml_fp16_t) * y_ne;
+    const uint64_t ids_sz = nbi2;
     const uint64_t d_sz = sizeof(float) * d_ne;
 
     vk_buffer d_D = extra->buffer_gpu.lock();
@@ -3594,6 +3652,11 @@ static void ggml_vk_mul_mat_vec_id_q_f16(ggml_backend_vk_context * ctx, vk_conte
         qy_buf_offset = extra_src1->offset;
         GGML_ASSERT(d_Qy != nullptr);
     }
+    if(!ids_uma) {
+        d_ids = extra_ids->buffer_gpu.lock();
+        ids_buf_offset = extra_ids->offset;
+        GGML_ASSERT(d_ids != nullptr);
+    }
     if (qx_needs_dequant) {
         d_X = ctx->prealloc_x;
     } else {
@@ -3619,7 +3682,7 @@ static void ggml_vk_mul_mat_vec_id_q_f16(ggml_backend_vk_context * ctx, vk_conte
     } else {
         to_fp16_vk_1 = ggml_vk_get_to_fp16(ctx, src1->type);
     }
-    vk_pipeline dmmv = ggml_vk_get_dequantize_mul_mat_vec(ctx, src0->type, src1->type);
+    vk_pipeline dmmv = ggml_vk_get_dequantize_mul_mat_vec_id(ctx, src0->type, src1->type);
     GGML_ASSERT(!qx_needs_dequant || to_fp16_vk_0 != nullptr);  // NOLINT
     GGML_ASSERT(!qy_needs_dequant || to_fp16_vk_1 != nullptr);  // NOLINT
     GGML_ASSERT(dmmv != nullptr);
@@ -3642,27 +3705,34 @@ static void ggml_vk_mul_mat_vec_id_q_f16(ggml_backend_vk_context * ctx, vk_conte
         ggml_vk_cpy_to_contiguous(ctx, subctx, to_fp16_vk_1, src1, { d_Qy, qy_buf_offset, VK_WHOLE_SIZE }, { d_Y, 0, VK_WHOLE_SIZE });
     }
 
-    uint32_t stride_batch_x = ne00*ne01;
     uint32_t stride_batch_y = ne10*ne11;
 
-    if (!ggml_vk_dim01_contiguous(src0) && !qx_needs_dequant) {
-        stride_batch_x = src0->nb[0] / ggml_type_size(src0->type);
-    }
-
     if (!ggml_vk_dim01_contiguous(src1) && !qy_needs_dequant) {
         stride_batch_y = src1->nb[0] / ggml_type_size(src1->type);
     }
 
     // compute
-    const vk_mat_vec_push_constants pc = {
+    const vk_mat_vec_id_push_constants pc = {
         (uint32_t)ne00, (uint32_t)ne10, (uint32_t)ne10, (uint32_t)ne01,
-        (uint32_t)ne02, (uint32_t)ne12, (uint32_t)r2, (uint32_t)r3,
-        stride_batch_x, stride_batch_y, (uint32_t)(ne20*ne21),
-        // 0, 0, 0, 0, 1
+        (uint32_t)x_ne, stride_batch_y, (uint32_t)(ne20*ne21),
+        (uint32_t)nei0, (uint32_t)ne11,
     };
     ggml_vk_sync_buffers(subctx);
-    ggml_vk_dispatch_pipeline(ctx, subctx, dmmv, { { d_X, x_buf_offset, x_sz * ne02 * ne03 }, { d_Y, y_buf_offset, y_sz * ne12 * ne13 }, { d_D, d_buf_offset, d_sz * ne22 * ne23} }, sizeof(vk_mat_vec_push_constants), &pc, { (uint32_t)ne01, (uint32_t)(ne12 * ne13), 1});
-}*/
+    ggml_vk_dispatch_pipeline(ctx, subctx, dmmv,
+        { { d_X, x_buf_offset, x_sz * ne02 * ne03 }, { d_Y, y_buf_offset, y_sz * ne12 * ne13 }, { d_D, d_buf_offset, d_sz * ne22 * ne23}, { d_ids, ids_buf_offset, ids_sz } },
+        sizeof(vk_mat_vec_id_push_constants), &pc, { (uint32_t)ne01, (uint32_t)nei0, 1 });
+}
+
+static void ggml_vk_mul_mat_id(ggml_backend_vk_context * ctx, vk_context * subctx, const ggml_tensor * src0, const ggml_tensor * src1, const ggml_tensor * src2, ggml_tensor * dst) {
+#ifdef GGML_VULKAN_DEBUG
+    std::cerr << "ggml_vk_mul_mat_id(" << src0 << ", " << src1 << ", " << src2 << ", " << dst << ")" << std::endl;
+#endif
+    if (src2->ne[1] == 1 && (src0->type == GGML_TYPE_F32 || src0->type == GGML_TYPE_F16 || ggml_is_quantized(src0->type))) {
+        ggml_vk_mul_mat_vec_id_q_f16(ctx, subctx, src0, src1, src2, dst);
+    } else {
+        ggml_vk_mul_mat_id_q_f16(ctx, subctx, src0, src1, src2, dst);
+    }
+}
 
 static void ggml_vk_op_repeat(ggml_backend_vk_context * ctx, vk_context * subctx, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
     // guaranteed to be an integer due to the check in ggml_can_repeat
@@ -3754,6 +3824,11 @@ static vk_pipeline ggml_vk_op_get_pipeline(ggml_backend_vk_context * ctx, const
             return ctx->device->pipeline_mul_f32;
         }
         return nullptr;
+    case GGML_OP_DIV:
+        if (src0->type == GGML_TYPE_F32 && src1->type == GGML_TYPE_F32 && dst->type == GGML_TYPE_F32) {
+            return ctx->device->pipeline_div_f32;
+        }
+        return nullptr;
     case GGML_OP_SCALE:
         if (src0->type == GGML_TYPE_F32 && dst->type == GGML_TYPE_F32) {
             return ctx->device->pipeline_scale_f32;
@@ -3851,6 +3926,11 @@ static vk_pipeline ggml_vk_op_get_pipeline(ggml_backend_vk_context * ctx, const
             return ctx->device->pipeline_argsort_f32;
         }
         return nullptr;
+    case GGML_OP_SUM_ROWS:
+        if (src0->type == GGML_TYPE_F32 && dst->type == GGML_TYPE_F32) {
+            return ctx->device->pipeline_sum_rows_f32;
+        }
+        return nullptr;
     default:
         return nullptr;
     }
@@ -3873,6 +3953,7 @@ static bool ggml_vk_op_supports_incontiguous(ggml_op op) {
     case GGML_OP_GET_ROWS:
     case GGML_OP_ADD:
     case GGML_OP_MUL:
+    case GGML_OP_DIV:
     case GGML_OP_SCALE:
     case GGML_OP_SQR:
     case GGML_OP_CLAMP:
@@ -3895,7 +3976,7 @@ static void ggml_vk_op_f32(ggml_backend_vk_context * ctx, vk_context * subctx, c
     std::cerr << "), (" << dst << ", name=" << dst->name << ", type=" << dst->type << ", ne0=" << dst->ne[0] << ", ne1=" << dst->ne[1] << ", ne2=" << dst->ne[2] << ", ne3=" << dst->ne[3] << ", nb0=" << dst->nb[0] << ", nb1=" << dst->nb[1] << ", nb2=" << dst->nb[2] << ", nb3=" << dst->nb[3] << "), " << ggml_op_name(op) << ")" << std::endl;
 #endif
     GGML_ASSERT(op == GGML_OP_GET_ROWS || (!ggml_is_quantized(src0->type) && (src1 == nullptr || !ggml_is_quantized(src1->type))));  // NOLINT
-    GGML_ASSERT(op == GGML_OP_CPY || ggml_vk_dim01_contiguous(src0));  // NOLINT
+    GGML_ASSERT(ggml_vk_op_supports_incontiguous(op) || ggml_vk_dim01_contiguous(src0));  // NOLINT
     GGML_ASSERT(dst->extra != nullptr);
     const uint64_t ne00 = src0->ne[0];
     const uint64_t ne01 = src0->ne[1];
@@ -3918,6 +3999,12 @@ static void ggml_vk_op_f32(ggml_backend_vk_context * ctx, vk_context * subctx, c
     const uint64_t ne23 = use_src2 ? src2->ne[3] : 0;
     const uint64_t ne2 = ne20 * ne21;
 
+    const uint64_t ned0 = dst->ne[0];
+    const uint64_t ned1 = dst->ne[1];
+    const uint64_t ned2 = dst->ne[2];
+    const uint64_t ned3 = dst->ne[3];
+    const uint64_t ned = ned0 * ned1;
+
     vk_pipeline pipeline = ggml_vk_op_get_pipeline(ctx, src0, src1, src2, dst, op);
     ggml_vk_func_t op_func;
 
@@ -3967,10 +4054,10 @@ static void ggml_vk_op_f32(ggml_backend_vk_context * ctx, vk_context * subctx, c
         }
     }
 
-    uint64_t x_sz = ggml_vk_align_size(ggml_type_size(src0->type)/ggml_blck_size(src0->type) * ne0, ctx->device->properties.limits.minStorageBufferOffsetAlignment);
-    uint64_t y_sz = use_src1 ? ggml_vk_align_size(ggml_type_size(src1->type) * ne1, ctx->device->properties.limits.minStorageBufferOffsetAlignment) : 0;
-    uint64_t z_sz = use_src2 ? ggml_vk_align_size(ggml_type_size(src2->type) * ne2, ctx->device->properties.limits.minStorageBufferOffsetAlignment) : 0;
-    uint64_t d_sz = ggml_type_size(dst->type) * ne0;
+    uint64_t x_sz = ggml_type_size(src0->type)/ggml_blck_size(src0->type) * ne0;
+    uint64_t y_sz = use_src1 ? ggml_type_size(src1->type) * ne1 : 0;
+    uint64_t z_sz = use_src2 ? ggml_type_size(src2->type) * ne2 : 0;
+    uint64_t d_sz = ggml_type_size(dst->type) * ned;
 
     vk_buffer d_D = extra->buffer_gpu.lock();
 
@@ -4028,6 +4115,7 @@ static void ggml_vk_op_f32(ggml_backend_vk_context * ctx, vk_context * subctx, c
         case GGML_OP_NORM:
         case GGML_OP_RMS_NORM:
         case GGML_OP_SOFT_MAX:
+        case GGML_OP_SUM_ROWS:
             elements = { (uint32_t)ggml_nrows(src0), 1, 1 };
             break;
         case GGML_OP_DIAG_MASK_INF:
@@ -4056,7 +4144,7 @@ static void ggml_vk_op_f32(ggml_backend_vk_context * ctx, vk_context * subctx, c
                 z_sz *= ne22 * ne23;
             }
             if (d_sz != VK_WHOLE_SIZE) {
-                d_sz *= ne02 * ne03;
+                d_sz *= ned2 * ned3;
             }
         }
 
@@ -4193,6 +4281,21 @@ static void ggml_vk_mul(ggml_backend_vk_context * ctx, vk_context * subctx, cons
     });
 }
 
+static void ggml_vk_div(ggml_backend_vk_context * ctx, vk_context * subctx, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
+    const uint32_t src0_type_size = ggml_type_size(src0->type);
+    const uint32_t src1_type_size = ggml_type_size(src1->type);
+    const uint32_t dst_type_size = ggml_type_size(dst->type);
+
+    ggml_vk_op_f32<vk_op_binary_push_constants>(ctx, subctx, src0, src1, nullptr, dst, GGML_OP_DIV, {
+        (uint32_t)ggml_nelements(src0),
+        (uint32_t)src0->ne[0], (uint32_t)src0->ne[1], (uint32_t)src0->ne[2],(uint32_t)src0->ne[3], (uint32_t)src0->nb[0] / src0_type_size, (uint32_t)src0->nb[1] / src0_type_size, (uint32_t)src0->nb[2] / src0_type_size, (uint32_t)src0->nb[3] / src0_type_size,
+        (uint32_t)src1->ne[0], (uint32_t)src1->ne[1], (uint32_t)src1->ne[2],(uint32_t)src1->ne[3], (uint32_t)src1->nb[0] / src1_type_size, (uint32_t)src1->nb[1] / src1_type_size, (uint32_t)src1->nb[2] / src1_type_size, (uint32_t)src1->nb[3] / src1_type_size,
+        (uint32_t) dst->ne[0], (uint32_t) dst->ne[1], (uint32_t) dst->ne[2],(uint32_t) dst->ne[3], (uint32_t) dst->nb[0] /  dst_type_size, (uint32_t) dst->nb[1] /  dst_type_size, (uint32_t) dst->nb[2] /  dst_type_size, (uint32_t) dst->nb[3] /  dst_type_size,
+        0,
+        0.0f, 0.0f,
+    });
+}
+
 static void ggml_vk_scale(ggml_backend_vk_context * ctx, vk_context * subctx, const ggml_tensor * src0, ggml_tensor * dst) {
     float * op_params = (float *)dst->op_params;
     const uint32_t src0_type_size = ggml_type_size(src0->type);
@@ -4342,10 +4445,6 @@ static void ggml_vk_argsort(ggml_backend_vk_context * ctx, vk_context * subctx,
 
     GGML_ASSERT(ncols_pad <= 1024);
 
-    std::cerr << "ncols=" << ncols << " ncols_pad=" << ncols_pad << " ascending=" << op_params[0] << std::endl;
-
-    std::cerr << ((ggml_sort_order) op_params[0]) << " " << GGML_SORT_ORDER_ASC << std::endl;
-
     ggml_vk_op_f32<vk_op_argsort_push_constants>(ctx, subctx, src0, nullptr, nullptr, dst, GGML_OP_ARGSORT, {
         ncols,
         ncols_pad,
@@ -4353,6 +4452,10 @@ static void ggml_vk_argsort(ggml_backend_vk_context * ctx, vk_context * subctx,
     });
 }
 
+static void ggml_vk_sum_rows(ggml_backend_vk_context * ctx, vk_context * subctx, const ggml_tensor * src0, ggml_tensor * dst) {
+    ggml_vk_op_f32<vk_op_push_constants>(ctx, subctx, src0, nullptr, nullptr, dst, GGML_OP_SUM_ROWS, { (uint32_t)src0->ne[0], 0, 0.0f, 0.0f });
+}
+
 #ifdef GGML_VULKAN_RUN_TESTS
 static void ggml_vk_print_matrix_area(const void * data, ggml_type type, int ne0, int ne1, int i0, int i1, int i2) {
     if (type != GGML_TYPE_F32 && type != GGML_TYPE_F16) {
@@ -4548,7 +4651,9 @@ static void ggml_vk_test_matmul(ggml_backend_vk_context * ctx, size_t m, size_t
         ggml_vk_ctx_begin(ctx, subctx);
         ggml_vk_matmul(
             ctx, subctx, p, ggml_vk_subbuffer(d_X), ggml_vk_subbuffer(d_Y), ggml_vk_subbuffer(d_D), ggml_vk_subbuffer(ctx->prealloc_split_k),
-            m, n, k, k, k, m, split_k, batch, batch, batch, 1, 1, k*m, k*n, m*n, 0, 0, 0, 0, 1
+            m, n, k,
+            k, k, m, k*m, k*n, m*n,
+            split_k, batch, batch, batch, 1, 1
         );
         ggml_vk_ctx_end(subctx);
     }
@@ -5052,7 +5157,9 @@ static void ggml_vk_test_dequant_matmul(ggml_backend_vk_context * ctx, size_t m,
         ggml_vk_ctx_begin(ctx, subctx);
         ggml_vk_matmul(
             ctx, subctx, p, ggml_vk_subbuffer(qx_buf), ggml_vk_subbuffer(y_buf), ggml_vk_subbuffer(d_buf), ggml_vk_subbuffer(ctx->prealloc_split_k),
-            m, n, k, k, k, m, split_k, batch, batch, batch, 1, 1, k*m, k*n, m*n, 0, 0, 0, 0, 1
+            m, n, k,
+            k, k, m, k*m, k*n, m*n,
+            split_k, batch, batch, batch, 1, 1
         );
         ggml_vk_ctx_end(subctx);
     }
@@ -5237,12 +5344,14 @@ static void ggml_vk_preallocate_buffers_graph(ggml_backend_vk_context * ctx, ggm
     case GGML_OP_CONT:
     case GGML_OP_DUP:
     case GGML_OP_MUL:
+    case GGML_OP_DIV:
     case GGML_OP_NORM:
     case GGML_OP_RMS_NORM:
     case GGML_OP_DIAG_MASK_INF:
     case GGML_OP_SOFT_MAX:
     case GGML_OP_ROPE:
     case GGML_OP_ARGSORT:
+    case GGML_OP_SUM_ROWS:
         break;
     case GGML_OP_UNARY:
         switch (ggml_get_unary_op(node)) {
@@ -5479,6 +5588,7 @@ static void ggml_vk_build_graph(ggml_backend_vk_context * ctx, ggml_tensor * nod
     case GGML_OP_GET_ROWS:
     case GGML_OP_ADD:
     case GGML_OP_MUL:
+    case GGML_OP_DIV:
     case GGML_OP_SCALE:
     case GGML_OP_SQR:
     case GGML_OP_CLAMP:
@@ -5498,6 +5608,7 @@ static void ggml_vk_build_graph(ggml_backend_vk_context * ctx, ggml_tensor * nod
     case GGML_OP_MUL_MAT_ID:
     case GGML_OP_NONE:
     case GGML_OP_ARGSORT:
+    case GGML_OP_SUM_ROWS:
         break;
     default:
         std::cerr << "ggml_vulkan: Error: Missing op: " << ggml_op_name(node->op) << std::endl;
@@ -5526,6 +5637,10 @@ static void ggml_vk_build_graph(ggml_backend_vk_context * ctx, ggml_tensor * nod
     case GGML_OP_MUL:
         ggml_vk_mul(ctx, ctx->compute_ctx, src0, src1, node);
 
+        break;
+    case GGML_OP_DIV:
+        ggml_vk_div(ctx, ctx->compute_ctx, src0, src1, node);
+
         break;
     case GGML_OP_SCALE:
         ggml_vk_scale(ctx, ctx->compute_ctx, src0, node);
@@ -5584,15 +5699,18 @@ static void ggml_vk_build_graph(ggml_backend_vk_context * ctx, ggml_tensor * nod
         break;
     case GGML_OP_ARGSORT:
         ggml_vk_argsort(ctx, ctx->compute_ctx, src0, node);
+
+        break;
+    case GGML_OP_SUM_ROWS:
+        ggml_vk_sum_rows(ctx, ctx->compute_ctx, src0, node);
+
         break;
     case GGML_OP_MUL_MAT:
         ggml_vk_mul_mat(ctx, ctx->compute_ctx, src0, src1, node);
 
         break;
     case GGML_OP_MUL_MAT_ID:
-        //ggml_vk_mul_mat_id(ctx, ctx->compute_ctx, src0, src1, node);
-        std::cerr << "ggml_vulkan: GGML_OP_MUL_MAT_ID not implemented yet." << std::endl;
-        GGML_ASSERT(false);
+        ggml_vk_mul_mat_id(ctx, ctx->compute_ctx, src0, src1, src2, node);
 
         break;
     default:
@@ -5622,6 +5740,7 @@ static bool ggml_vk_compute_forward(ggml_backend_vk_context * ctx, ggml_compute_
     case GGML_OP_ADD:
     case GGML_OP_GET_ROWS:
     case GGML_OP_MUL:
+    case GGML_OP_DIV:
     case GGML_OP_SCALE:
     case GGML_OP_SQR:
     case GGML_OP_CLAMP:
@@ -5639,6 +5758,7 @@ static bool ggml_vk_compute_forward(ggml_backend_vk_context * ctx, ggml_compute_
     case GGML_OP_TRANSPOSE:
     case GGML_OP_NONE:
     case GGML_OP_ARGSORT:
+    case GGML_OP_SUM_ROWS:
         extra = (ggml_tensor_extra_gpu *) tensor->extra;
 
         break;
@@ -6293,7 +6413,7 @@ GGML_CALL static bool ggml_backend_vk_supports_op(ggml_backend_t backend, const
             }
             break;
         case GGML_OP_MUL_MAT:
-        // case GGML_OP_MUL_MAT_ID:
+        case GGML_OP_MUL_MAT_ID:
             {
                 switch (op->src[0]->type) {
                     case GGML_TYPE_F32:
@@ -6377,6 +6497,7 @@ GGML_CALL static bool ggml_backend_vk_supports_op(ggml_backend_t backend, const
         case GGML_OP_NORM:
         case GGML_OP_ADD:
         case GGML_OP_MUL:
+        case GGML_OP_DIV:
         case GGML_OP_RMS_NORM:
         case GGML_OP_SCALE:
         case GGML_OP_SQR:
@@ -6385,6 +6506,7 @@ GGML_CALL static bool ggml_backend_vk_supports_op(ggml_backend_t backend, const
         case GGML_OP_DIAG_MASK_INF:
         case GGML_OP_SOFT_MAX:
         case GGML_OP_ARGSORT:
+        case GGML_OP_SUM_ROWS:
             return true;
         default:
             return false;
@@ -6394,15 +6516,10 @@ GGML_CALL static bool ggml_backend_vk_supports_op(ggml_backend_t backend, const
 }
 
 GGML_CALL static bool ggml_backend_vk_offload_op(ggml_backend_t backend, const ggml_tensor * op) {
-    const ggml_tensor * dst = op;
-
     const int min_batch_size = 32;
 
-    if (dst->ne[1] > min_batch_size && dst->op != GGML_OP_GET_ROWS) {
-        return true;
-    }
-
-    return false;
+    return (op->ne[1] >= min_batch_size && op->op != GGML_OP_GET_ROWS) ||
+           (op->ne[2] >= min_batch_size && op->op == GGML_OP_MUL_MAT_ID);
 
     UNUSED(backend);
 }
@@ -6681,9 +6798,9 @@ static void ggml_vk_check_results_0(ggml_backend_vk_context * ctx, ggml_compute_
     size_t src1_size;
     size_t src2_size;
 
-    void * src0_buffer;
-    void * src1_buffer;
-    void * src2_buffer;
+    void * src0_buffer = nullptr;
+    void * src1_buffer = nullptr;
+    void * src2_buffer = nullptr;
 
     if (src0 != nullptr) {
         src0_clone = ggml_dup_tensor(ggml_ctx, src0);
@@ -6846,8 +6963,12 @@ static void ggml_vk_check_results_0(ggml_backend_vk_context * ctx, ggml_compute_
 
     if (tensor->op == GGML_OP_MUL_MAT) {
         tensor_clone = ggml_mul_mat(ggml_ctx, src0_clone, src1_clone);
+    } else if (tensor->op == GGML_OP_MUL_MAT_ID) {
+        tensor_clone = ggml_mul_mat_id(ggml_ctx, src0_clone, src1_clone, src2_clone);
     } else if (tensor->op == GGML_OP_MUL) {
         tensor_clone = ggml_mul(ggml_ctx, src0_clone, src1_clone);
+    } else if (tensor->op == GGML_OP_DIV) {
+        tensor_clone = ggml_div(ggml_ctx, src0_clone, src1_clone);
     } else if (tensor->op == GGML_OP_SCALE) {
         tensor_clone = ggml_scale(ggml_ctx, src0_clone, ((float *)tensor->op_params)[0]);
     } else if (tensor->op == GGML_OP_SQR) {
@@ -6917,6 +7038,8 @@ static void ggml_vk_check_results_0(ggml_backend_vk_context * ctx, ggml_compute_
         tensor_clone = ggml_get_rows(ggml_ctx, src0_clone, src1_clone);
     } else if (tensor->op == GGML_OP_ARGSORT) {
         tensor_clone = ggml_argsort(ggml_ctx, src0_clone, (ggml_sort_order) *(int *)tensor->op_params);
+    } else if (tensor->op == GGML_OP_SUM_ROWS) {
+        tensor_clone = ggml_sum_rows(ggml_ctx, src0_clone);
     } else {
         std::cerr << "Missing vk_check_results OP: " << ggml_op_name(tensor->op) << std::endl;
         GGML_ASSERT(false);
@@ -6964,6 +7087,7 @@ static void ggml_vk_check_results_1(ggml_backend_vk_context * ctx, ggml_compute_
 
     ggml_tensor * src0 = tensor->src[0];
     ggml_tensor * src1 = tensor->src[1];
+    ggml_tensor * src2 = tensor->src[2];
 
     void * tensor_data = tensor->data;
 
@@ -7022,6 +7146,9 @@ static void ggml_vk_check_results_1(ggml_backend_vk_context * ctx, ggml_compute_
                         if (src1 != nullptr) {
                             std::cerr << "src1=" << src1 << " src1->name=" << src1->name << " op=" << ggml_op_name(src1->op) << " type=" << ggml_type_name(src1->type) << " ne0=" << src1->ne[0] << " nb0=" << src1->nb[0] << " ne1=" << src1->ne[1] << " nb1=" << src1->nb[1] << " ne2=" << src1->ne[2] << " nb2=" << src1->nb[2] << " ne3=" << src1->ne[3] << " nb3=" << src1->nb[3] << " offset=" << src1->view_offs << std::endl;
                         }
+                        if (src2 != nullptr) {
+                            std::cerr << "src2=" << src2 << " src2->name=" << src2->name << " op=" << ggml_op_name(src2->op) << " type=" << ggml_type_name(src2->type) << " ne0=" << src2->ne[0] << " nb0=" << src2->nb[0] << " ne1=" << src2->ne[1] << " nb1=" << src2->nb[1] << " ne2=" << src2->ne[2] << " nb2=" << src2->nb[2] << " ne3=" << src2->ne[3] << " nb3=" << src2->nb[3] << " offset=" << src2->view_offs << std::endl;
+                        }
                         std::cerr << "First error: result=" << first_error_result << " correct=" << first_error_correct  << " i3=" << first_error[3] << " i2=" << first_error[2] << " i1=" << first_error[1] << " i0=" << first_error[0] << std::endl;
                         std::cerr << std::endl << "Result:" << std::endl;
                         ggml_vk_print_tensor_area(tensor, tensor_data, i0, i1, i2, i3);
@@ -7063,6 +7190,9 @@ static void ggml_vk_check_results_1(ggml_backend_vk_context * ctx, ggml_compute_
         if (src1 != nullptr) {
             std::cerr << "src1=" << src1 << " op=" << ggml_op_name(src1->op) << " type=" << ggml_type_name(src1->type) << " ne0=" << src1->ne[0] << " nb0=" << src1->nb[0] << " ne1=" << src1->ne[1] << " nb1=" << src1->nb[1] << " ne2=" << src1->ne[2] << " nb2=" << src1->nb[2] << " ne3=" << src1->ne[3] << " nb3=" << src1->nb[3] << " offset=" << src1->view_offs << std::endl;
         }
+        if (src2 != nullptr) {
+            std::cerr << "src2=" << src2 << " op=" << ggml_op_name(src2->op) << " type=" << ggml_type_name(src2->type) << " ne0=" << src2->ne[0] << " nb0=" << src2->nb[0] << " ne1=" << src2->ne[1] << " nb1=" << src2->nb[1] << " ne2=" << src2->ne[2] << " nb2=" << src2->nb[2] << " ne3=" << src2->ne[3] << " nb3=" << src2->nb[3] << " offset=" << src2->view_offs << std::endl;
+        }
         std::cerr << "First error: result=" << first_error_result << " correct=" << first_error_correct  << " i3=" << first_error[3] << " i2=" << first_error[2] << " i1=" << first_error[1] << " i0=" << first_error[0] << std::endl;
         std::cerr << std::endl << "Result:" << std::endl;
         ggml_vk_print_tensor_area(tensor, tensor_data, 5, 5, 0, 0);
@@ -7087,6 +7217,9 @@ static void ggml_vk_check_results_1(ggml_backend_vk_context * ctx, ggml_compute_
         if (src1 != nullptr) {
             std::cerr << "src1=" << src1 << " op=" << ggml_op_name(src1->op) << " type=" << ggml_type_name(src1->type) << " ne0=" << src1->ne[0] << " nb0=" << src1->nb[0] << " ne1=" << src1->ne[1] << " nb1=" << src1->nb[1] << " ne2=" << src1->ne[2] << " nb2=" << src1->nb[2] << " ne3=" << src1->ne[3] << " nb3=" << src1->nb[3] << " offset=" << src1->view_offs << std::endl;
         }
+        if (src2 != nullptr) {
+            std::cerr << "src2=" << src2 << " op=" << ggml_op_name(src2->op) << " type=" << ggml_type_name(src2->type) << " ne0=" << src2->ne[0] << " nb0=" << src2->nb[0] << " ne1=" << src2->ne[1] << " nb1=" << src2->nb[1] << " ne2=" << src2->ne[2] << " nb2=" << src2->nb[2] << " ne3=" << src2->ne[3] << " nb3=" << src2->nb[3] << " offset=" << src2->view_offs << std::endl;
+        }
         std::cerr << "First error: result=" << first_error_result << " correct=" << first_error_correct  << " i3=" << first_error[3] << " i2=" << first_error[2] << " i1=" << first_error[1] << " i0=" << first_error[0] << std::endl;
         std::cerr << std::endl << "Result:" << std::endl;
         ggml_vk_print_tensor_area(tensor, tensor_data, first_error[0], first_error[1], first_error[2], first_error[3]);