Vulkan Bugfixes and Improvements (llama/7084)

* Modify mat mat mul shader for mul_mat_id, modify mat vec mul shaders for single call batch operation

* Further work towards MoE, disabled for now

* Disable MoE code (not ready yet), fix a number of bugs in shaders and Vulkan code

* Add softmax with f16 mask and pos buffer support

* Disable mul_mat_id shaders for now

* Fix flake8

* Fix validation errors caused by empty buffers on larger batch sizes

diff --git a/ggml-vulkan.cpp b/ggml-vulkan.cpp
index f712cdd5a900eb454625d4c1a3a6a1c27cdb89e1..95f71897405397e0f4757c88325993d9913bf33f 100644 (file)
--- a/ggml-vulkan.cpp
+++ b/ggml-vulkan.cpp
@@ -120,8 +120,16 @@ struct vk_device {
 
     vk_matmul_pipeline pipeline_dequant_mul_mat_mat[VK_NUM_TYPES];
 
+    vk_matmul_pipeline pipeline_matmul_id_f32;
+    vk_matmul_pipeline pipeline_matmul_id_f16;
+    vk_matmul_pipeline pipeline_matmul_id_f16_f32;
+
+    vk_matmul_pipeline pipeline_dequant_mul_mat_mat_id[VK_NUM_TYPES];
+
     vk_pipeline pipeline_dequant[VK_NUM_TYPES];
-    vk_pipeline pipeline_dequant_mul_mat_vec_f32[VK_NUM_TYPES];
+    vk_pipeline pipeline_dequant_mul_mat_vec_f32_f32[VK_NUM_TYPES];
+    vk_pipeline pipeline_dequant_mul_mat_vec_f16_f32[VK_NUM_TYPES];
+    vk_pipeline pipeline_dequant_mul_mat_vec_id_f32[VK_NUM_TYPES];
 
     vk_pipeline pipeline_mul_mat_vec_p021_f16_f32;
     vk_pipeline pipeline_mul_mat_vec_nc_f16_f32;
@@ -139,7 +147,7 @@ struct vk_device {
     vk_pipeline pipeline_silu_f32;
     vk_pipeline pipeline_relu_f32;
     vk_pipeline pipeline_diag_mask_inf_f32;
-    vk_pipeline pipeline_soft_max_f32;
+    vk_pipeline pipeline_soft_max_f32, pipeline_soft_max_f32_f16;
     vk_pipeline pipeline_rope_f32, pipeline_rope_f16;
     vk_pipeline pipeline_rope_neox_f32, pipeline_rope_neox_f16;
     vk_pipeline pipeline_argsort_f32;
@@ -215,6 +223,21 @@ struct vk_submission {
 
 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 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;
+};
+
+struct vk_mat_vec_push_constants {
+    uint32_t ncols; uint32_t stride_a; uint32_t stride_b; uint32_t stride_d;
+    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;
+};
+
 struct vk_op_push_constants {
     uint32_t KX;
     uint32_t KY;
@@ -1003,201 +1026,422 @@ 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_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>();
+    ctx->device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_1] = std::make_shared<vk_matmul_pipeline_struct>();
+    ctx->device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q5_0] = std::make_shared<vk_matmul_pipeline_struct>();
+    ctx->device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q5_1] = std::make_shared<vk_matmul_pipeline_struct>();
+    ctx->device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q8_0] = std::make_shared<vk_matmul_pipeline_struct>();
+    ctx->device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q2_K] = std::make_shared<vk_matmul_pipeline_struct>();
+    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>();*/
+
     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, 14 * sizeof(uint32_t), l_wg_denoms, warptile_l, 1);
-        ggml_vk_create_pipeline(ctx, ctx->device->pipeline_matmul_f32->m, "matmul_f32_m", matmul_f32_len, matmul_f32_data, "main", 3, 14 * sizeof(uint32_t), m_wg_denoms, warptile_m, 1);
-        ggml_vk_create_pipeline(ctx, ctx->device->pipeline_matmul_f32->s, "matmul_f32_s", matmul_f32_len, matmul_f32_data, "main", 3, 14 * sizeof(uint32_t), s_wg_denoms, warptile_s, 1);
-        ggml_vk_create_pipeline(ctx, ctx->device->pipeline_matmul_f32->a_l, "matmul_f32_aligned_l", matmul_f32_aligned_len, matmul_f32_aligned_data, "main", 3, 14 * sizeof(uint32_t), l_wg_denoms, warptile_l, l_align);
-        ggml_vk_create_pipeline(ctx, ctx->device->pipeline_matmul_f32->a_m, "matmul_f32_aligned_m", matmul_f32_aligned_len, matmul_f32_aligned_data, "main", 3, 14 * sizeof(uint32_t), m_wg_denoms, warptile_m, m_align);
-        ggml_vk_create_pipeline(ctx, ctx->device->pipeline_matmul_f32->a_s, "matmul_f32_aligned_s", matmul_f32_aligned_len, matmul_f32_aligned_data, "main", 3, 14 * sizeof(uint32_t), s_wg_denoms, warptile_s, s_align);
-
-        ggml_vk_create_pipeline(ctx, ctx->device->pipeline_matmul_f16->l, "matmul_f16_l", matmul_f16_len, matmul_f16_data, "main", 3, 14 * sizeof(uint32_t), l_wg_denoms, warptile_l, 1);
-        ggml_vk_create_pipeline(ctx, ctx->device->pipeline_matmul_f16->m, "matmul_f16_m", matmul_f16_len, matmul_f16_data, "main", 3, 14 * sizeof(uint32_t), m_wg_denoms, warptile_m, 1);
-        ggml_vk_create_pipeline(ctx, ctx->device->pipeline_matmul_f16->s, "matmul_f16_s", matmul_f16_len, matmul_f16_data, "main", 3, 14 * sizeof(uint32_t), s_wg_denoms, warptile_s, 1);
-        ggml_vk_create_pipeline(ctx, ctx->device->pipeline_matmul_f16->a_l, "matmul_f16_aligned_l", matmul_f16_aligned_len, matmul_f16_aligned_data, "main", 3, 14 * sizeof(uint32_t), l_wg_denoms, warptile_l, l_align);
-        ggml_vk_create_pipeline(ctx, ctx->device->pipeline_matmul_f16->a_m, "matmul_f16_aligned_m", matmul_f16_aligned_len, matmul_f16_aligned_data, "main", 3, 14 * sizeof(uint32_t), m_wg_denoms, warptile_m, m_align);
-        ggml_vk_create_pipeline(ctx, ctx->device->pipeline_matmul_f16->a_s, "matmul_f16_aligned_s", matmul_f16_aligned_len, matmul_f16_aligned_data, "main", 3, 14 * sizeof(uint32_t), s_wg_denoms, warptile_s, s_align);
-
-        ggml_vk_create_pipeline(ctx, ctx->device->pipeline_matmul_f16_f32->l, "matmul_f16_f32_l", matmul_f16_f32_len, matmul_f16_f32_data, "main", 3, 14 * sizeof(uint32_t), l_wg_denoms, warptile_l, 1);
-        ggml_vk_create_pipeline(ctx, ctx->device->pipeline_matmul_f16_f32->m, "matmul_f16_f32_m", matmul_f16_f32_len, matmul_f16_f32_data, "main", 3, 14 * sizeof(uint32_t), m_wg_denoms, warptile_m, 1);
-        ggml_vk_create_pipeline(ctx, ctx->device->pipeline_matmul_f16_f32->s, "matmul_f16_f32_s", matmul_f16_f32_len, matmul_f16_f32_data, "main", 3, 14 * sizeof(uint32_t), s_wg_denoms, warptile_s, 1);
-        ggml_vk_create_pipeline(ctx, ctx->device->pipeline_matmul_f16_f32->a_l, "matmul_f16_f32_aligned_l", matmul_f16_f32_aligned_len, matmul_f16_f32_aligned_data, "main", 3, 14 * sizeof(uint32_t), l_wg_denoms, warptile_l, l_align);
-        ggml_vk_create_pipeline(ctx, ctx->device->pipeline_matmul_f16_f32->a_m, "matmul_f16_f32_aligned_m", matmul_f16_f32_aligned_len, matmul_f16_f32_aligned_data, "main", 3, 14 * sizeof(uint32_t), m_wg_denoms, warptile_m, m_align);
-        ggml_vk_create_pipeline(ctx, ctx->device->pipeline_matmul_f16_f32->a_s, "matmul_f16_f32_aligned_s", matmul_f16_f32_aligned_len, matmul_f16_f32_aligned_data, "main", 3, 14 * sizeof(uint32_t), s_wg_denoms, warptile_s, s_align);
-
-        ggml_vk_create_pipeline(ctx, ctx->device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q4_0]->l, "matmul_q4_0_f32_l", matmul_q4_0_f32_len, matmul_q4_0_f32_data, "main", 3, 14 * sizeof(uint32_t), l_wg_denoms, warptile_mmq_l, l_align);
-        ggml_vk_create_pipeline(ctx, ctx->device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q4_0]->m, "matmul_q4_0_f32_m", matmul_q4_0_f32_len, matmul_q4_0_f32_data, "main", 3, 14 * sizeof(uint32_t), m_wg_denoms, warptile_mmq_m, m_align);
-        ggml_vk_create_pipeline(ctx, ctx->device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q4_0]->s, "matmul_q4_0_f32_s", matmul_q4_0_f32_len, matmul_q4_0_f32_data, "main", 3, 14 * sizeof(uint32_t), s_wg_denoms, warptile_mmq_s, s_align);
-        ggml_vk_create_pipeline(ctx, ctx->device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q4_0]->a_l, "matmul_q4_0_f32_aligned_l", matmul_q4_0_f32_aligned_len, matmul_q4_0_f32_aligned_data, "main", 3, 14 * sizeof(uint32_t), l_wg_denoms, warptile_mmq_l, l_align);
-        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, 14 * sizeof(uint32_t), 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, 14 * sizeof(uint32_t), 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, 14 * sizeof(uint32_t), 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, 14 * sizeof(uint32_t), 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, 14 * sizeof(uint32_t), 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, 14 * sizeof(uint32_t), 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, 14 * sizeof(uint32_t), 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, 14 * sizeof(uint32_t), 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, 14 * sizeof(uint32_t), 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, 14 * sizeof(uint32_t), m_wg_denoms, warptile_mmq_m, m_align);
-        ggml_vk_create_pipeline(ctx, ctx->device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q5_0]->s, "matmul_q5_0_f32_s", matmul_q5_0_f32_len, matmul_q5_0_f32_data, "main", 3, 14 * sizeof(uint32_t), s_wg_denoms, warptile_mmq_s, s_align);
-        ggml_vk_create_pipeline(ctx, ctx->device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q5_0]->a_l, "matmul_q5_0_f32_aligned_l", matmul_q5_0_f32_aligned_len, matmul_q5_0_f32_aligned_data, "main", 3, 14 * sizeof(uint32_t), l_wg_denoms, warptile_mmq_l, l_align);
-        ggml_vk_create_pipeline(ctx, ctx->device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q5_0]->a_m, "matmul_q5_0_f32_aligned_m", matmul_q5_0_f32_aligned_len, matmul_q5_0_f32_aligned_data, "main", 3, 14 * sizeof(uint32_t), m_wg_denoms, warptile_mmq_m, m_align);
-        ggml_vk_create_pipeline(ctx, ctx->device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q5_0]->a_s, "matmul_q5_0_f32_aligned_s", matmul_q5_0_f32_aligned_len, matmul_q5_0_f32_aligned_data, "main", 3, 14 * sizeof(uint32_t), s_wg_denoms, warptile_mmq_s, s_align);
-
-        ggml_vk_create_pipeline(ctx, ctx->device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q5_1]->l, "matmul_q5_1_f32_l", matmul_q5_1_f32_len, matmul_q5_1_f32_data, "main", 3, 14 * sizeof(uint32_t), l_wg_denoms, warptile_mmq_l, l_align);
-        ggml_vk_create_pipeline(ctx, ctx->device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q5_1]->m, "matmul_q5_1_f32_m", matmul_q5_1_f32_len, matmul_q5_1_f32_data, "main", 3, 14 * sizeof(uint32_t), m_wg_denoms, warptile_mmq_m, m_align);
-        ggml_vk_create_pipeline(ctx, ctx->device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q5_1]->s, "matmul_q5_1_f32_s", matmul_q5_1_f32_len, matmul_q5_1_f32_data, "main", 3, 14 * sizeof(uint32_t), s_wg_denoms, warptile_mmq_s, s_align);
-        ggml_vk_create_pipeline(ctx, ctx->device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q5_1]->a_l, "matmul_q5_1_f32_aligned_l", matmul_q5_1_f32_aligned_len, matmul_q5_1_f32_aligned_data, "main", 3, 14 * sizeof(uint32_t), l_wg_denoms, warptile_mmq_l, l_align);
-        ggml_vk_create_pipeline(ctx, ctx->device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q5_1]->a_m, "matmul_q5_1_f32_aligned_m", matmul_q5_1_f32_aligned_len, matmul_q5_1_f32_aligned_data, "main", 3, 14 * sizeof(uint32_t), m_wg_denoms, warptile_mmq_m, m_align);
-        ggml_vk_create_pipeline(ctx, ctx->device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q5_1]->a_s, "matmul_q5_1_f32_aligned_s", matmul_q5_1_f32_aligned_len, matmul_q5_1_f32_aligned_data, "main", 3, 14 * sizeof(uint32_t), s_wg_denoms, warptile_mmq_s, s_align);
-
-        ggml_vk_create_pipeline(ctx, ctx->device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q8_0]->l, "matmul_q8_0_f32_l", matmul_q8_0_f32_len, matmul_q8_0_f32_data, "main", 3, 14 * sizeof(uint32_t), l_wg_denoms, warptile_mmq_l, l_align);
-        ggml_vk_create_pipeline(ctx, ctx->device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q8_0]->m, "matmul_q8_0_f32_m", matmul_q8_0_f32_len, matmul_q8_0_f32_data, "main", 3, 14 * sizeof(uint32_t), m_wg_denoms, warptile_mmq_m, m_align);
-        ggml_vk_create_pipeline(ctx, ctx->device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q8_0]->s, "matmul_q8_0_f32_s", matmul_q8_0_f32_len, matmul_q8_0_f32_data, "main", 3, 14 * sizeof(uint32_t), s_wg_denoms, warptile_mmq_s, s_align);
-        ggml_vk_create_pipeline(ctx, ctx->device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q8_0]->a_l, "matmul_q8_0_f32_aligned_l", matmul_q8_0_f32_aligned_len, matmul_q8_0_f32_aligned_data, "main", 3, 14 * sizeof(uint32_t), l_wg_denoms, warptile_mmq_l, l_align);
-        ggml_vk_create_pipeline(ctx, ctx->device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q8_0]->a_m, "matmul_q8_0_f32_aligned_m", matmul_q8_0_f32_aligned_len, matmul_q8_0_f32_aligned_data, "main", 3, 14 * sizeof(uint32_t), m_wg_denoms, warptile_mmq_m, m_align);
-        ggml_vk_create_pipeline(ctx, ctx->device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q8_0]->a_s, "matmul_q8_0_f32_aligned_s", matmul_q8_0_f32_aligned_len, matmul_q8_0_f32_aligned_data, "main", 3, 14 * sizeof(uint32_t), s_wg_denoms, warptile_mmq_s, s_align);
-
-        ggml_vk_create_pipeline(ctx, ctx->device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q2_K]->l, "matmul_q2_k_f32_l", matmul_q2_k_f32_len, matmul_q2_k_f32_data, "main", 3, 14 * sizeof(uint32_t), l_wg_denoms, warptile_mmq_l, l_align);
-        ggml_vk_create_pipeline(ctx, ctx->device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q2_K]->m, "matmul_q2_k_f32_m", matmul_q2_k_f32_len, matmul_q2_k_f32_data, "main", 3, 14 * sizeof(uint32_t), m_wg_denoms, warptile_mmq_m, m_align);
-        ggml_vk_create_pipeline(ctx, ctx->device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q2_K]->s, "matmul_q2_k_f32_s", matmul_q2_k_f32_len, matmul_q2_k_f32_data, "main", 3, 14 * sizeof(uint32_t), s_wg_denoms, warptile_mmq_s, s_align);
-        ggml_vk_create_pipeline(ctx, ctx->device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q2_K]->a_l, "matmul_q2_k_f32_aligned_l", matmul_q2_k_f32_aligned_len, matmul_q2_k_f32_aligned_data, "main", 3, 14 * sizeof(uint32_t), l_wg_denoms, warptile_mmq_l, l_align);
-        ggml_vk_create_pipeline(ctx, ctx->device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q2_K]->a_m, "matmul_q2_k_f32_aligned_m", matmul_q2_k_f32_aligned_len, matmul_q2_k_f32_aligned_data, "main", 3, 14 * sizeof(uint32_t), m_wg_denoms, warptile_mmq_m, m_align);
-        ggml_vk_create_pipeline(ctx, ctx->device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q2_K]->a_s, "matmul_q2_k_f32_aligned_s", matmul_q2_k_f32_aligned_len, matmul_q2_k_f32_aligned_data, "main", 3, 14 * sizeof(uint32_t), s_wg_denoms, warptile_mmq_s, s_align);
-
-        ggml_vk_create_pipeline(ctx, ctx->device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q3_K]->l, "matmul_q3_k_f32_l", matmul_q3_k_f32_len, matmul_q3_k_f32_data, "main", 3, 14 * sizeof(uint32_t), l_wg_denoms, warptile_mmq_l, l_align);
-        ggml_vk_create_pipeline(ctx, ctx->device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q3_K]->m, "matmul_q3_k_f32_m", matmul_q3_k_f32_len, matmul_q3_k_f32_data, "main", 3, 14 * sizeof(uint32_t), m_wg_denoms, warptile_mmq_m, m_align);
-        ggml_vk_create_pipeline(ctx, ctx->device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q3_K]->s, "matmul_q3_k_f32_s", matmul_q3_k_f32_len, matmul_q3_k_f32_data, "main", 3, 14 * sizeof(uint32_t), s_wg_denoms, warptile_mmq_s, s_align);
-        ggml_vk_create_pipeline(ctx, ctx->device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q3_K]->a_l, "matmul_q3_k_f32_aligned_l", matmul_q3_k_f32_aligned_len, matmul_q3_k_f32_aligned_data, "main", 3, 14 * sizeof(uint32_t), l_wg_denoms, warptile_mmq_l, l_align);
-        ggml_vk_create_pipeline(ctx, ctx->device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q3_K]->a_m, "matmul_q3_k_f32_aligned_m", matmul_q3_k_f32_aligned_len, matmul_q3_k_f32_aligned_data, "main", 3, 14 * sizeof(uint32_t), m_wg_denoms, warptile_mmq_m, m_align);
-        ggml_vk_create_pipeline(ctx, ctx->device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q3_K]->a_s, "matmul_q3_k_f32_aligned_s", matmul_q3_k_f32_aligned_len, matmul_q3_k_f32_aligned_data, "main", 3, 14 * sizeof(uint32_t), s_wg_denoms, warptile_mmq_s, s_align);
-
-        ggml_vk_create_pipeline(ctx, ctx->device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q4_K]->l, "matmul_q4_k_f32_l", matmul_q4_k_f32_len, matmul_q4_k_f32_data, "main", 3, 14 * sizeof(uint32_t), l_wg_denoms, warptile_mmq_l, l_align);
-        ggml_vk_create_pipeline(ctx, ctx->device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q4_K]->m, "matmul_q4_k_f32_m", matmul_q4_k_f32_len, matmul_q4_k_f32_data, "main", 3, 14 * sizeof(uint32_t), m_wg_denoms, warptile_mmq_m, m_align);
-        ggml_vk_create_pipeline(ctx, ctx->device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q4_K]->s, "matmul_q4_k_f32_s", matmul_q4_k_f32_len, matmul_q4_k_f32_data, "main", 3, 14 * sizeof(uint32_t), s_wg_denoms, warptile_mmq_s, s_align);
-        ggml_vk_create_pipeline(ctx, ctx->device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q4_K]->a_l, "matmul_q4_k_f32_aligned_l", matmul_q4_k_f32_aligned_len, matmul_q4_k_f32_aligned_data, "main", 3, 14 * sizeof(uint32_t), l_wg_denoms, warptile_mmq_l, l_align);
-        ggml_vk_create_pipeline(ctx, ctx->device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q4_K]->a_m, "matmul_q4_k_f32_aligned_m", matmul_q4_k_f32_aligned_len, matmul_q4_k_f32_aligned_data, "main", 3, 14 * sizeof(uint32_t), m_wg_denoms, warptile_mmq_m, m_align);
-        ggml_vk_create_pipeline(ctx, ctx->device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q4_K]->a_s, "matmul_q4_k_f32_aligned_s", matmul_q4_k_f32_aligned_len, matmul_q4_k_f32_aligned_data, "main", 3, 14 * sizeof(uint32_t), s_wg_denoms, warptile_mmq_s, s_align);
-
-        ggml_vk_create_pipeline(ctx, ctx->device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q5_K]->l, "matmul_q5_k_f32_l", matmul_q5_k_f32_len, matmul_q5_k_f32_data, "main", 3, 14 * sizeof(uint32_t), l_wg_denoms, warptile_mmq_l, l_align);
-        ggml_vk_create_pipeline(ctx, ctx->device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q5_K]->m, "matmul_q5_k_f32_m", matmul_q5_k_f32_len, matmul_q5_k_f32_data, "main", 3, 14 * sizeof(uint32_t), m_wg_denoms, warptile_mmq_m, m_align);
-        ggml_vk_create_pipeline(ctx, ctx->device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q5_K]->s, "matmul_q5_k_f32_s", matmul_q5_k_f32_len, matmul_q5_k_f32_data, "main", 3, 14 * sizeof(uint32_t), s_wg_denoms, warptile_mmq_s, s_align);
-        ggml_vk_create_pipeline(ctx, ctx->device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q5_K]->a_l, "matmul_q5_k_f32_aligned_l", matmul_q5_k_f32_aligned_len, matmul_q5_k_f32_aligned_data, "main", 3, 14 * sizeof(uint32_t), l_wg_denoms, warptile_mmq_l, l_align);
-        ggml_vk_create_pipeline(ctx, ctx->device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q5_K]->a_m, "matmul_q5_k_f32_aligned_m", matmul_q5_k_f32_aligned_len, matmul_q5_k_f32_aligned_data, "main", 3, 14 * sizeof(uint32_t), m_wg_denoms, warptile_mmq_m, m_align);
-        ggml_vk_create_pipeline(ctx, ctx->device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q5_K]->a_s, "matmul_q5_k_f32_aligned_s", matmul_q5_k_f32_aligned_len, matmul_q5_k_f32_aligned_data, "main", 3, 14 * sizeof(uint32_t), s_wg_denoms, warptile_mmq_s, s_align);
-
-        ggml_vk_create_pipeline(ctx, ctx->device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q6_K]->l, "matmul_q6_k_f32_l", matmul_q6_k_f32_len, matmul_q6_k_f32_data, "main", 3, 14 * sizeof(uint32_t), l_wg_denoms, warptile_mmq_l, l_align);
-        ggml_vk_create_pipeline(ctx, ctx->device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q6_K]->m, "matmul_q6_k_f32_m", matmul_q6_k_f32_len, matmul_q6_k_f32_data, "main", 3, 14 * sizeof(uint32_t), m_wg_denoms, warptile_mmq_m, m_align);
-        ggml_vk_create_pipeline(ctx, ctx->device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q6_K]->s, "matmul_q6_k_f32_s", matmul_q6_k_f32_len, matmul_q6_k_f32_data, "main", 3, 14 * sizeof(uint32_t), s_wg_denoms, warptile_mmq_s, s_align);
-        ggml_vk_create_pipeline(ctx, ctx->device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q6_K]->a_l, "matmul_q6_k_f32_aligned_l", matmul_q6_k_f32_aligned_len, matmul_q6_k_f32_aligned_data, "main", 3, 14 * sizeof(uint32_t), l_wg_denoms, warptile_mmq_l, l_align);
-        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, 14 * sizeof(uint32_t), 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, 14 * sizeof(uint32_t), s_wg_denoms, warptile_mmq_s, s_align);
+        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);
+        ggml_vk_create_pipeline(ctx, ctx->device->pipeline_matmul_f32->m, "matmul_f32_m", matmul_f32_len, matmul_f32_data, "main", 3, sizeof(vk_mat_mat_push_constants), m_wg_denoms, warptile_m, 1);
+        ggml_vk_create_pipeline(ctx, ctx->device->pipeline_matmul_f32->s, "matmul_f32_s", matmul_f32_len, matmul_f32_data, "main", 3, sizeof(vk_mat_mat_push_constants), s_wg_denoms, warptile_s, 1);
+        ggml_vk_create_pipeline(ctx, ctx->device->pipeline_matmul_f32->a_l, "matmul_f32_aligned_l", matmul_f32_aligned_len, matmul_f32_aligned_data, "main", 3, sizeof(vk_mat_mat_push_constants), l_wg_denoms, warptile_l, l_align);
+        ggml_vk_create_pipeline(ctx, ctx->device->pipeline_matmul_f32->a_m, "matmul_f32_aligned_m", matmul_f32_aligned_len, matmul_f32_aligned_data, "main", 3, sizeof(vk_mat_mat_push_constants), m_wg_denoms, warptile_m, m_align);
+        ggml_vk_create_pipeline(ctx, ctx->device->pipeline_matmul_f32->a_s, "matmul_f32_aligned_s", matmul_f32_aligned_len, matmul_f32_aligned_data, "main", 3, sizeof(vk_mat_mat_push_constants), s_wg_denoms, warptile_s, s_align);
+
+        ggml_vk_create_pipeline(ctx, ctx->device->pipeline_matmul_f16->l, "matmul_f16_l", matmul_f16_len, matmul_f16_data, "main", 3, sizeof(vk_mat_mat_push_constants), l_wg_denoms, warptile_l, 1);
+        ggml_vk_create_pipeline(ctx, ctx->device->pipeline_matmul_f16->m, "matmul_f16_m", matmul_f16_len, matmul_f16_data, "main", 3, sizeof(vk_mat_mat_push_constants), m_wg_denoms, warptile_m, 1);
+        ggml_vk_create_pipeline(ctx, ctx->device->pipeline_matmul_f16->s, "matmul_f16_s", matmul_f16_len, matmul_f16_data, "main", 3, sizeof(vk_mat_mat_push_constants), s_wg_denoms, warptile_s, 1);
+        ggml_vk_create_pipeline(ctx, ctx->device->pipeline_matmul_f16->a_l, "matmul_f16_aligned_l", matmul_f16_aligned_len, matmul_f16_aligned_data, "main", 3, sizeof(vk_mat_mat_push_constants), l_wg_denoms, warptile_l, l_align);
+        ggml_vk_create_pipeline(ctx, ctx->device->pipeline_matmul_f16->a_m, "matmul_f16_aligned_m", matmul_f16_aligned_len, matmul_f16_aligned_data, "main", 3, sizeof(vk_mat_mat_push_constants), m_wg_denoms, warptile_m, m_align);
+        ggml_vk_create_pipeline(ctx, ctx->device->pipeline_matmul_f16->a_s, "matmul_f16_aligned_s", matmul_f16_aligned_len, matmul_f16_aligned_data, "main", 3, sizeof(vk_mat_mat_push_constants), s_wg_denoms, warptile_s, s_align);
+
+        ggml_vk_create_pipeline(ctx, ctx->device->pipeline_matmul_f16_f32->l, "matmul_f16_f32_l", matmul_f16_f32_len, matmul_f16_f32_data, "main", 3, sizeof(vk_mat_mat_push_constants), l_wg_denoms, warptile_l, 1);
+        ggml_vk_create_pipeline(ctx, ctx->device->pipeline_matmul_f16_f32->m, "matmul_f16_f32_m", matmul_f16_f32_len, matmul_f16_f32_data, "main", 3, sizeof(vk_mat_mat_push_constants), m_wg_denoms, warptile_m, 1);
+        ggml_vk_create_pipeline(ctx, ctx->device->pipeline_matmul_f16_f32->s, "matmul_f16_f32_s", matmul_f16_f32_len, matmul_f16_f32_data, "main", 3, sizeof(vk_mat_mat_push_constants), s_wg_denoms, warptile_s, 1);
+        ggml_vk_create_pipeline(ctx, ctx->device->pipeline_matmul_f16_f32->a_l, "matmul_f16_f32_aligned_l", matmul_f16_f32_aligned_len, matmul_f16_f32_aligned_data, "main", 3, sizeof(vk_mat_mat_push_constants), l_wg_denoms, warptile_l, l_align);
+        ggml_vk_create_pipeline(ctx, ctx->device->pipeline_matmul_f16_f32->a_m, "matmul_f16_f32_aligned_m", matmul_f16_f32_aligned_len, matmul_f16_f32_aligned_data, "main", 3, sizeof(vk_mat_mat_push_constants), m_wg_denoms, warptile_m, m_align);
+        ggml_vk_create_pipeline(ctx, ctx->device->pipeline_matmul_f16_f32->a_s, "matmul_f16_f32_aligned_s", matmul_f16_f32_aligned_len, matmul_f16_f32_aligned_data, "main", 3, 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[GGML_TYPE_Q4_0]->l, "matmul_q4_0_f32_l", matmul_q4_0_f32_len, matmul_q4_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_Q4_0]->m, "matmul_q4_0_f32_m", matmul_q4_0_f32_len, matmul_q4_0_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_0]->s, "matmul_q4_0_f32_s", matmul_q4_0_f32_len, matmul_q4_0_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_0]->a_l, "matmul_q4_0_f32_aligned_l", matmul_q4_0_f32_aligned_len, matmul_q4_0_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_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_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);
+        ggml_vk_create_pipeline(ctx, ctx->device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q5_0]->s, "matmul_q5_0_f32_s", matmul_q5_0_f32_len, matmul_q5_0_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_Q5_0]->a_l, "matmul_q5_0_f32_aligned_l", matmul_q5_0_f32_aligned_len, matmul_q5_0_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_Q5_0]->a_m, "matmul_q5_0_f32_aligned_m", matmul_q5_0_f32_aligned_len, matmul_q5_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_Q5_0]->a_s, "matmul_q5_0_f32_aligned_s", matmul_q5_0_f32_aligned_len, matmul_q5_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_Q5_1]->l, "matmul_q5_1_f32_l", matmul_q5_1_f32_len, matmul_q5_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_Q5_1]->m, "matmul_q5_1_f32_m", matmul_q5_1_f32_len, matmul_q5_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_Q5_1]->s, "matmul_q5_1_f32_s", matmul_q5_1_f32_len, matmul_q5_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_Q5_1]->a_l, "matmul_q5_1_f32_aligned_l", matmul_q5_1_f32_aligned_len, matmul_q5_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_Q5_1]->a_m, "matmul_q5_1_f32_aligned_m", matmul_q5_1_f32_aligned_len, matmul_q5_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_Q5_1]->a_s, "matmul_q5_1_f32_aligned_s", matmul_q5_1_f32_aligned_len, matmul_q5_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_Q8_0]->l, "matmul_q8_0_f32_l", matmul_q8_0_f32_len, matmul_q8_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_Q8_0]->m, "matmul_q8_0_f32_m", matmul_q8_0_f32_len, matmul_q8_0_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_Q8_0]->s, "matmul_q8_0_f32_s", matmul_q8_0_f32_len, matmul_q8_0_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_Q8_0]->a_l, "matmul_q8_0_f32_aligned_l", matmul_q8_0_f32_aligned_len, matmul_q8_0_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_Q8_0]->a_m, "matmul_q8_0_f32_aligned_m", matmul_q8_0_f32_aligned_len, matmul_q8_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_Q8_0]->a_s, "matmul_q8_0_f32_aligned_s", matmul_q8_0_f32_aligned_len, matmul_q8_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_Q2_K]->l, "matmul_q2_k_f32_l", matmul_q2_k_f32_len, matmul_q2_k_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_Q2_K]->m, "matmul_q2_k_f32_m", matmul_q2_k_f32_len, matmul_q2_k_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_Q2_K]->s, "matmul_q2_k_f32_s", matmul_q2_k_f32_len, matmul_q2_k_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_Q2_K]->a_l, "matmul_q2_k_f32_aligned_l", matmul_q2_k_f32_aligned_len, matmul_q2_k_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_Q2_K]->a_m, "matmul_q2_k_f32_aligned_m", matmul_q2_k_f32_aligned_len, matmul_q2_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_Q2_K]->a_s, "matmul_q2_k_f32_aligned_s", matmul_q2_k_f32_aligned_len, matmul_q2_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_dequant_mul_mat_mat[GGML_TYPE_Q3_K]->l, "matmul_q3_k_f32_l", matmul_q3_k_f32_len, matmul_q3_k_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_Q3_K]->m, "matmul_q3_k_f32_m", matmul_q3_k_f32_len, matmul_q3_k_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_Q3_K]->s, "matmul_q3_k_f32_s", matmul_q3_k_f32_len, matmul_q3_k_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_Q3_K]->a_l, "matmul_q3_k_f32_aligned_l", matmul_q3_k_f32_aligned_len, matmul_q3_k_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_Q3_K]->a_m, "matmul_q3_k_f32_aligned_m", matmul_q3_k_f32_aligned_len, matmul_q3_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_Q3_K]->a_s, "matmul_q3_k_f32_aligned_s", matmul_q3_k_f32_aligned_len, matmul_q3_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_dequant_mul_mat_mat[GGML_TYPE_Q4_K]->l, "matmul_q4_k_f32_l", matmul_q4_k_f32_len, matmul_q4_k_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_K]->m, "matmul_q4_k_f32_m", matmul_q4_k_f32_len, matmul_q4_k_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_K]->s, "matmul_q4_k_f32_s", matmul_q4_k_f32_len, matmul_q4_k_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_K]->a_l, "matmul_q4_k_f32_aligned_l", matmul_q4_k_f32_aligned_len, matmul_q4_k_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_K]->a_m, "matmul_q4_k_f32_aligned_m", matmul_q4_k_f32_aligned_len, matmul_q4_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_Q4_K]->a_s, "matmul_q4_k_f32_aligned_s", matmul_q4_k_f32_aligned_len, matmul_q4_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_dequant_mul_mat_mat[GGML_TYPE_Q5_K]->l, "matmul_q5_k_f32_l", matmul_q5_k_f32_len, matmul_q5_k_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_K]->m, "matmul_q5_k_f32_m", matmul_q5_k_f32_len, matmul_q5_k_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_Q5_K]->s, "matmul_q5_k_f32_s", matmul_q5_k_f32_len, matmul_q5_k_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_Q5_K]->a_l, "matmul_q5_k_f32_aligned_l", matmul_q5_k_f32_aligned_len, matmul_q5_k_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_Q5_K]->a_m, "matmul_q5_k_f32_aligned_m", matmul_q5_k_f32_aligned_len, matmul_q5_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_Q5_K]->a_s, "matmul_q5_k_f32_aligned_s", matmul_q5_k_f32_aligned_len, matmul_q5_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_dequant_mul_mat_mat[GGML_TYPE_Q6_K]->l, "matmul_q6_k_f32_l", matmul_q6_k_f32_len, matmul_q6_k_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_Q6_K]->m, "matmul_q6_k_f32_m", matmul_q6_k_f32_len, matmul_q6_k_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_Q6_K]->s, "matmul_q6_k_f32_s", matmul_q6_k_f32_len, matmul_q6_k_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_Q6_K]->a_l, "matmul_q6_k_f32_aligned_l", matmul_q6_k_f32_aligned_len, matmul_q6_k_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_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);*/
     } 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, 14 * sizeof(uint32_t), 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, 14 * sizeof(uint32_t), m_wg_denoms, warptile_m, 1);
-        ggml_vk_create_pipeline(ctx, ctx->device->pipeline_matmul_f32->s, "matmul_f32_s", matmul_f32_fp32_len, matmul_f32_fp32_data, "main", 3, 14 * sizeof(uint32_t), s_wg_denoms, warptile_s, 1);
-        ggml_vk_create_pipeline(ctx, ctx->device->pipeline_matmul_f32->a_l, "matmul_f32_aligned_l", matmul_f32_aligned_fp32_len, matmul_f32_aligned_fp32_data, "main", 3, 14 * sizeof(uint32_t), l_wg_denoms, warptile_l, l_align);
-        ggml_vk_create_pipeline(ctx, ctx->device->pipeline_matmul_f32->a_m, "matmul_f32_aligned_m", matmul_f32_aligned_fp32_len, matmul_f32_aligned_fp32_data, "main", 3, 14 * sizeof(uint32_t), m_wg_denoms, warptile_m, m_align);
-        ggml_vk_create_pipeline(ctx, ctx->device->pipeline_matmul_f32->a_s, "matmul_f32_aligned_s", matmul_f32_aligned_fp32_len, matmul_f32_aligned_fp32_data, "main", 3, 14 * sizeof(uint32_t), s_wg_denoms, warptile_s, s_align);
-
-        ggml_vk_create_pipeline(ctx, ctx->device->pipeline_matmul_f16->l, "matmul_f16_l", matmul_f16_fp32_len, matmul_f16_fp32_data, "main", 3, 14 * sizeof(uint32_t), l_wg_denoms, warptile_l, 1);
-        ggml_vk_create_pipeline(ctx, ctx->device->pipeline_matmul_f16->m, "matmul_f16_m", matmul_f16_fp32_len, matmul_f16_fp32_data, "main", 3, 14 * sizeof(uint32_t), m_wg_denoms, warptile_m, 1);
-        ggml_vk_create_pipeline(ctx, ctx->device->pipeline_matmul_f16->s, "matmul_f16_s", matmul_f16_fp32_len, matmul_f16_fp32_data, "main", 3, 14 * sizeof(uint32_t), s_wg_denoms, warptile_s, 1);
-        ggml_vk_create_pipeline(ctx, ctx->device->pipeline_matmul_f16->a_l, "matmul_f16_aligned_l", matmul_f16_aligned_fp32_len, matmul_f16_aligned_fp32_data, "main", 3, 14 * sizeof(uint32_t), l_wg_denoms, warptile_l, l_align);
-        ggml_vk_create_pipeline(ctx, ctx->device->pipeline_matmul_f16->a_m, "matmul_f16_aligned_m", matmul_f16_aligned_fp32_len, matmul_f16_aligned_fp32_data, "main", 3, 14 * sizeof(uint32_t), m_wg_denoms, warptile_m, m_align);
-        ggml_vk_create_pipeline(ctx, ctx->device->pipeline_matmul_f16->a_s, "matmul_f16_aligned_s", matmul_f16_aligned_fp32_len, matmul_f16_aligned_fp32_data, "main", 3, 14 * sizeof(uint32_t), s_wg_denoms, warptile_s, s_align);
-
-        ggml_vk_create_pipeline(ctx, ctx->device->pipeline_matmul_f16_f32->l, "matmul_f16_f32_l", matmul_f16_f32_fp32_len, matmul_f16_f32_fp32_data, "main", 3, 14 * sizeof(uint32_t), l_wg_denoms, warptile_l, 1);
-        ggml_vk_create_pipeline(ctx, ctx->device->pipeline_matmul_f16_f32->m, "matmul_f16_f32_m", matmul_f16_f32_fp32_len, matmul_f16_f32_fp32_data, "main", 3, 14 * sizeof(uint32_t), m_wg_denoms, warptile_m, 1);
-        ggml_vk_create_pipeline(ctx, ctx->device->pipeline_matmul_f16_f32->s, "matmul_f16_f32_s", matmul_f16_f32_fp32_len, matmul_f16_f32_fp32_data, "main", 3, 14 * sizeof(uint32_t), s_wg_denoms, warptile_s, 1);
-        ggml_vk_create_pipeline(ctx, ctx->device->pipeline_matmul_f16_f32->a_l, "matmul_f16_f32_aligned_l", matmul_f16_f32_aligned_fp32_len, matmul_f16_f32_aligned_fp32_data, "main", 3, 14 * sizeof(uint32_t), l_wg_denoms, warptile_l, l_align);
-        ggml_vk_create_pipeline(ctx, ctx->device->pipeline_matmul_f16_f32->a_m, "matmul_f16_f32_aligned_m", matmul_f16_f32_aligned_fp32_len, matmul_f16_f32_aligned_fp32_data, "main", 3, 14 * sizeof(uint32_t), m_wg_denoms, warptile_m, m_align);
-        ggml_vk_create_pipeline(ctx, ctx->device->pipeline_matmul_f16_f32->a_s, "matmul_f16_f32_aligned_s", matmul_f16_f32_aligned_fp32_len, matmul_f16_f32_aligned_fp32_data, "main", 3, 14 * sizeof(uint32_t), s_wg_denoms, warptile_s, s_align);
-
-        ggml_vk_create_pipeline(ctx, ctx->device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q4_0]->l, "matmul_q4_0_f32_l", matmul_q4_0_f32_fp32_len, matmul_q4_0_f32_fp32_data, "main", 3, 14 * sizeof(uint32_t), l_wg_denoms, warptile_mmq_l, l_align);
-        ggml_vk_create_pipeline(ctx, ctx->device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q4_0]->m, "matmul_q4_0_f32_m", matmul_q4_0_f32_fp32_len, matmul_q4_0_f32_fp32_data, "main", 3, 14 * sizeof(uint32_t), m_wg_denoms, warptile_mmq_m, m_align);
-        ggml_vk_create_pipeline(ctx, ctx->device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q4_0]->s, "matmul_q4_0_f32_s", matmul_q4_0_f32_fp32_len, matmul_q4_0_f32_fp32_data, "main", 3, 14 * sizeof(uint32_t), s_wg_denoms, warptile_mmq_s, s_align);
-        ggml_vk_create_pipeline(ctx, ctx->device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q4_0]->a_l, "matmul_q4_0_f32_aligned_l", matmul_q4_0_f32_aligned_fp32_len, matmul_q4_0_f32_aligned_fp32_data, "main", 3, 14 * sizeof(uint32_t), l_wg_denoms, warptile_mmq_l, l_align);
-        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_fp32_len, matmul_q4_0_f32_aligned_fp32_data, "main", 3, 14 * sizeof(uint32_t), 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_fp32_len, matmul_q4_0_f32_aligned_fp32_data, "main", 3, 14 * sizeof(uint32_t), 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_fp32_len, matmul_q4_1_f32_fp32_data, "main", 3, 14 * sizeof(uint32_t), 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_fp32_len, matmul_q4_1_f32_fp32_data, "main", 3, 14 * sizeof(uint32_t), 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_fp32_len, matmul_q4_1_f32_fp32_data, "main", 3, 14 * sizeof(uint32_t), 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_fp32_len, matmul_q4_1_f32_aligned_fp32_data, "main", 3, 14 * sizeof(uint32_t), 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_fp32_len, matmul_q4_1_f32_aligned_fp32_data, "main", 3, 14 * sizeof(uint32_t), 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_fp32_len, matmul_q4_1_f32_aligned_fp32_data, "main", 3, 14 * sizeof(uint32_t), 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_fp32_len, matmul_q5_0_f32_fp32_data, "main", 3, 14 * sizeof(uint32_t), 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_fp32_len, matmul_q5_0_f32_fp32_data, "main", 3, 14 * sizeof(uint32_t), m_wg_denoms, warptile_mmq_m, m_align);
-        ggml_vk_create_pipeline(ctx, ctx->device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q5_0]->s, "matmul_q5_0_f32_s", matmul_q5_0_f32_fp32_len, matmul_q5_0_f32_fp32_data, "main", 3, 14 * sizeof(uint32_t), s_wg_denoms, warptile_mmq_s, s_align);
-        ggml_vk_create_pipeline(ctx, ctx->device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q5_0]->a_l, "matmul_q5_0_f32_aligned_l", matmul_q5_0_f32_aligned_fp32_len, matmul_q5_0_f32_aligned_fp32_data, "main", 3, 14 * sizeof(uint32_t), l_wg_denoms, warptile_mmq_l, l_align);
-        ggml_vk_create_pipeline(ctx, ctx->device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q5_0]->a_m, "matmul_q5_0_f32_aligned_m", matmul_q5_0_f32_aligned_fp32_len, matmul_q5_0_f32_aligned_fp32_data, "main", 3, 14 * sizeof(uint32_t), m_wg_denoms, warptile_mmq_m, m_align);
-        ggml_vk_create_pipeline(ctx, ctx->device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q5_0]->a_s, "matmul_q5_0_f32_aligned_s", matmul_q5_0_f32_aligned_fp32_len, matmul_q5_0_f32_aligned_fp32_data, "main", 3, 14 * sizeof(uint32_t), s_wg_denoms, warptile_mmq_s, s_align);
-
-        ggml_vk_create_pipeline(ctx, ctx->device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q5_1]->l, "matmul_q5_1_f32_l", matmul_q5_1_f32_fp32_len, matmul_q5_1_f32_fp32_data, "main", 3, 14 * sizeof(uint32_t), l_wg_denoms, warptile_mmq_l, l_align);
-        ggml_vk_create_pipeline(ctx, ctx->device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q5_1]->m, "matmul_q5_1_f32_m", matmul_q5_1_f32_fp32_len, matmul_q5_1_f32_fp32_data, "main", 3, 14 * sizeof(uint32_t), m_wg_denoms, warptile_mmq_m, m_align);
-        ggml_vk_create_pipeline(ctx, ctx->device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q5_1]->s, "matmul_q5_1_f32_s", matmul_q5_1_f32_fp32_len, matmul_q5_1_f32_fp32_data, "main", 3, 14 * sizeof(uint32_t), s_wg_denoms, warptile_mmq_s, s_align);
-        ggml_vk_create_pipeline(ctx, ctx->device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q5_1]->a_l, "matmul_q5_1_f32_aligned_l", matmul_q5_1_f32_aligned_fp32_len, matmul_q5_1_f32_aligned_fp32_data, "main", 3, 14 * sizeof(uint32_t), l_wg_denoms, warptile_mmq_l, l_align);
-        ggml_vk_create_pipeline(ctx, ctx->device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q5_1]->a_m, "matmul_q5_1_f32_aligned_m", matmul_q5_1_f32_aligned_fp32_len, matmul_q5_1_f32_aligned_fp32_data, "main", 3, 14 * sizeof(uint32_t), m_wg_denoms, warptile_mmq_m, m_align);
-        ggml_vk_create_pipeline(ctx, ctx->device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q5_1]->a_s, "matmul_q5_1_f32_aligned_s", matmul_q5_1_f32_aligned_fp32_len, matmul_q5_1_f32_aligned_fp32_data, "main", 3, 14 * sizeof(uint32_t), s_wg_denoms, warptile_mmq_s, s_align);
-
-        ggml_vk_create_pipeline(ctx, ctx->device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q8_0]->l, "matmul_q8_0_f32_l", matmul_q8_0_f32_fp32_len, matmul_q8_0_f32_fp32_data, "main", 3, 14 * sizeof(uint32_t), l_wg_denoms, warptile_mmq_l, l_align);
-        ggml_vk_create_pipeline(ctx, ctx->device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q8_0]->m, "matmul_q8_0_f32_m", matmul_q8_0_f32_fp32_len, matmul_q8_0_f32_fp32_data, "main", 3, 14 * sizeof(uint32_t), m_wg_denoms, warptile_mmq_m, m_align);
-        ggml_vk_create_pipeline(ctx, ctx->device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q8_0]->s, "matmul_q8_0_f32_s", matmul_q8_0_f32_fp32_len, matmul_q8_0_f32_fp32_data, "main", 3, 14 * sizeof(uint32_t), s_wg_denoms, warptile_mmq_s, s_align);
-        ggml_vk_create_pipeline(ctx, ctx->device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q8_0]->a_l, "matmul_q8_0_f32_aligned_l", matmul_q8_0_f32_aligned_fp32_len, matmul_q8_0_f32_aligned_fp32_data, "main", 3, 14 * sizeof(uint32_t), l_wg_denoms, warptile_mmq_l, l_align);
-        ggml_vk_create_pipeline(ctx, ctx->device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q8_0]->a_m, "matmul_q8_0_f32_aligned_m", matmul_q8_0_f32_aligned_fp32_len, matmul_q8_0_f32_aligned_fp32_data, "main", 3, 14 * sizeof(uint32_t), m_wg_denoms, warptile_mmq_m, m_align);
-        ggml_vk_create_pipeline(ctx, ctx->device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q8_0]->a_s, "matmul_q8_0_f32_aligned_s", matmul_q8_0_f32_aligned_fp32_len, matmul_q8_0_f32_aligned_fp32_data, "main", 3, 14 * sizeof(uint32_t), s_wg_denoms, warptile_mmq_s, s_align);
-
-        ggml_vk_create_pipeline(ctx, ctx->device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q2_K]->l, "matmul_q2_k_f32_l", matmul_q2_k_f32_fp32_len, matmul_q2_k_f32_fp32_data, "main", 3, 14 * sizeof(uint32_t), l_wg_denoms, warptile_mmq_l, l_align);
-        ggml_vk_create_pipeline(ctx, ctx->device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q2_K]->m, "matmul_q2_k_f32_m", matmul_q2_k_f32_fp32_len, matmul_q2_k_f32_fp32_data, "main", 3, 14 * sizeof(uint32_t), m_wg_denoms, warptile_mmq_m, m_align);
-        ggml_vk_create_pipeline(ctx, ctx->device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q2_K]->s, "matmul_q2_k_f32_s", matmul_q2_k_f32_fp32_len, matmul_q2_k_f32_fp32_data, "main", 3, 14 * sizeof(uint32_t), s_wg_denoms, warptile_mmq_s, s_align);
-        ggml_vk_create_pipeline(ctx, ctx->device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q2_K]->a_l, "matmul_q2_k_f32_aligned_l", matmul_q2_k_f32_aligned_fp32_len, matmul_q2_k_f32_aligned_fp32_data, "main", 3, 14 * sizeof(uint32_t), l_wg_denoms, warptile_mmq_l, l_align);
-        ggml_vk_create_pipeline(ctx, ctx->device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q2_K]->a_m, "matmul_q2_k_f32_aligned_m", matmul_q2_k_f32_aligned_fp32_len, matmul_q2_k_f32_aligned_fp32_data, "main", 3, 14 * sizeof(uint32_t), m_wg_denoms, warptile_mmq_m, m_align);
-        ggml_vk_create_pipeline(ctx, ctx->device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q2_K]->a_s, "matmul_q2_k_f32_aligned_s", matmul_q2_k_f32_aligned_fp32_len, matmul_q2_k_f32_aligned_fp32_data, "main", 3, 14 * sizeof(uint32_t), s_wg_denoms, warptile_mmq_s, s_align);
-
-        ggml_vk_create_pipeline(ctx, ctx->device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q3_K]->l, "matmul_q3_k_f32_l", matmul_q3_k_f32_fp32_len, matmul_q3_k_f32_fp32_data, "main", 3, 14 * sizeof(uint32_t), l_wg_denoms, warptile_mmq_l, l_align);
-        ggml_vk_create_pipeline(ctx, ctx->device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q3_K]->m, "matmul_q3_k_f32_m", matmul_q3_k_f32_fp32_len, matmul_q3_k_f32_fp32_data, "main", 3, 14 * sizeof(uint32_t), m_wg_denoms, warptile_mmq_m, m_align);
-        ggml_vk_create_pipeline(ctx, ctx->device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q3_K]->s, "matmul_q3_k_f32_s", matmul_q3_k_f32_fp32_len, matmul_q3_k_f32_fp32_data, "main", 3, 14 * sizeof(uint32_t), s_wg_denoms, warptile_mmq_s, s_align);
-        ggml_vk_create_pipeline(ctx, ctx->device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q3_K]->a_l, "matmul_q3_k_f32_aligned_l", matmul_q3_k_f32_aligned_fp32_len, matmul_q3_k_f32_aligned_fp32_data, "main", 3, 14 * sizeof(uint32_t), l_wg_denoms, warptile_mmq_l, l_align);
-        ggml_vk_create_pipeline(ctx, ctx->device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q3_K]->a_m, "matmul_q3_k_f32_aligned_m", matmul_q3_k_f32_aligned_fp32_len, matmul_q3_k_f32_aligned_fp32_data, "main", 3, 14 * sizeof(uint32_t), m_wg_denoms, warptile_mmq_m, m_align);
-        ggml_vk_create_pipeline(ctx, ctx->device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q3_K]->a_s, "matmul_q3_k_f32_aligned_s", matmul_q3_k_f32_aligned_fp32_len, matmul_q3_k_f32_aligned_fp32_data, "main", 3, 14 * sizeof(uint32_t), s_wg_denoms, warptile_mmq_s, s_align);
-
-        ggml_vk_create_pipeline(ctx, ctx->device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q4_K]->l, "matmul_q4_k_f32_l", matmul_q4_k_f32_fp32_len, matmul_q4_k_f32_fp32_data, "main", 3, 14 * sizeof(uint32_t), l_wg_denoms, warptile_mmq_l, l_align);
-        ggml_vk_create_pipeline(ctx, ctx->device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q4_K]->m, "matmul_q4_k_f32_m", matmul_q4_k_f32_fp32_len, matmul_q4_k_f32_fp32_data, "main", 3, 14 * sizeof(uint32_t), m_wg_denoms, warptile_mmq_m, m_align);
-        ggml_vk_create_pipeline(ctx, ctx->device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q4_K]->s, "matmul_q4_k_f32_s", matmul_q4_k_f32_fp32_len, matmul_q4_k_f32_fp32_data, "main", 3, 14 * sizeof(uint32_t), s_wg_denoms, warptile_mmq_s, s_align);
-        ggml_vk_create_pipeline(ctx, ctx->device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q4_K]->a_l, "matmul_q4_k_f32_aligned_l", matmul_q4_k_f32_aligned_fp32_len, matmul_q4_k_f32_aligned_fp32_data, "main", 3, 14 * sizeof(uint32_t), l_wg_denoms, warptile_mmq_l, l_align);
-        ggml_vk_create_pipeline(ctx, ctx->device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q4_K]->a_m, "matmul_q4_k_f32_aligned_m", matmul_q4_k_f32_aligned_fp32_len, matmul_q4_k_f32_aligned_fp32_data, "main", 3, 14 * sizeof(uint32_t), m_wg_denoms, warptile_mmq_m, m_align);
-        ggml_vk_create_pipeline(ctx, ctx->device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q4_K]->a_s, "matmul_q4_k_f32_aligned_s", matmul_q4_k_f32_aligned_fp32_len, matmul_q4_k_f32_aligned_fp32_data, "main", 3, 14 * sizeof(uint32_t), s_wg_denoms, warptile_mmq_s, s_align);
-
-        ggml_vk_create_pipeline(ctx, ctx->device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q5_K]->l, "matmul_q5_k_f32_l", matmul_q5_k_f32_fp32_len, matmul_q5_k_f32_fp32_data, "main", 3, 14 * sizeof(uint32_t), l_wg_denoms, warptile_mmq_l, l_align);
-        ggml_vk_create_pipeline(ctx, ctx->device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q5_K]->m, "matmul_q5_k_f32_m", matmul_q5_k_f32_fp32_len, matmul_q5_k_f32_fp32_data, "main", 3, 14 * sizeof(uint32_t), m_wg_denoms, warptile_mmq_m, m_align);
-        ggml_vk_create_pipeline(ctx, ctx->device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q5_K]->s, "matmul_q5_k_f32_s", matmul_q5_k_f32_fp32_len, matmul_q5_k_f32_fp32_data, "main", 3, 14 * sizeof(uint32_t), s_wg_denoms, warptile_mmq_s, s_align);
-        ggml_vk_create_pipeline(ctx, ctx->device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q5_K]->a_l, "matmul_q5_k_f32_aligned_l", matmul_q5_k_f32_aligned_fp32_len, matmul_q5_k_f32_aligned_fp32_data, "main", 3, 14 * sizeof(uint32_t), l_wg_denoms, warptile_mmq_l, l_align);
-        ggml_vk_create_pipeline(ctx, ctx->device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q5_K]->a_m, "matmul_q5_k_f32_aligned_m", matmul_q5_k_f32_aligned_fp32_len, matmul_q5_k_f32_aligned_fp32_data, "main", 3, 14 * sizeof(uint32_t), m_wg_denoms, warptile_mmq_m, m_align);
-        ggml_vk_create_pipeline(ctx, ctx->device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q5_K]->a_s, "matmul_q5_k_f32_aligned_s", matmul_q5_k_f32_aligned_fp32_len, matmul_q5_k_f32_aligned_fp32_data, "main", 3, 14 * sizeof(uint32_t), s_wg_denoms, warptile_mmq_s, s_align);
-
-        ggml_vk_create_pipeline(ctx, ctx->device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q6_K]->l, "matmul_q6_k_f32_l", matmul_q6_k_f32_fp32_len, matmul_q6_k_f32_fp32_data, "main", 3, 14 * sizeof(uint32_t), l_wg_denoms, warptile_mmq_l, l_align);
-        ggml_vk_create_pipeline(ctx, ctx->device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q6_K]->m, "matmul_q6_k_f32_m", matmul_q6_k_f32_fp32_len, matmul_q6_k_f32_fp32_data, "main", 3, 14 * sizeof(uint32_t), m_wg_denoms, warptile_mmq_m, m_align);
-        ggml_vk_create_pipeline(ctx, ctx->device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q6_K]->s, "matmul_q6_k_f32_s", matmul_q6_k_f32_fp32_len, matmul_q6_k_f32_fp32_data, "main", 3, 14 * sizeof(uint32_t), s_wg_denoms, warptile_mmq_s, s_align);
-        ggml_vk_create_pipeline(ctx, ctx->device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q6_K]->a_l, "matmul_q6_k_f32_aligned_l", matmul_q6_k_f32_aligned_fp32_len, matmul_q6_k_f32_aligned_fp32_data, "main", 3, 14 * sizeof(uint32_t), l_wg_denoms, warptile_mmq_l, l_align);
-        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, 14 * sizeof(uint32_t), 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, 14 * sizeof(uint32_t), s_wg_denoms, warptile_mmq_s, s_align);
-    }
-
-    ggml_vk_create_pipeline(ctx, ctx->device->pipeline_dequant_mul_mat_vec_f32[GGML_TYPE_F16 ], "mul_mat_vec_f16_f32",  mul_mat_vec_f16_f32_len,  mul_mat_vec_f16_f32_data,  "main", 3, 3 * sizeof(uint32_t), {1, 1, 1}, { device->subgroup_size }, 1);
-    ggml_vk_create_pipeline(ctx, ctx->device->pipeline_dequant_mul_mat_vec_f32[GGML_TYPE_Q4_0], "mul_mat_vec_q4_0_f32", mul_mat_vec_q4_0_f32_len, mul_mat_vec_q4_0_f32_data, "main", 3, 3 * sizeof(uint32_t), {1, 1, 1}, { device->subgroup_size }, 1);
-    ggml_vk_create_pipeline(ctx, ctx->device->pipeline_dequant_mul_mat_vec_f32[GGML_TYPE_Q4_1], "mul_mat_vec_q4_1_f32", mul_mat_vec_q4_1_f32_len, mul_mat_vec_q4_1_f32_data, "main", 3, 3 * sizeof(uint32_t), {1, 1, 1}, { device->subgroup_size }, 1);
-    ggml_vk_create_pipeline(ctx, ctx->device->pipeline_dequant_mul_mat_vec_f32[GGML_TYPE_Q5_0], "mul_mat_vec_q5_0_f32", mul_mat_vec_q5_0_f32_len, mul_mat_vec_q5_0_f32_data, "main", 3, 3 * sizeof(uint32_t), {1, 1, 1}, { device->subgroup_size }, 1);
-    ggml_vk_create_pipeline(ctx, ctx->device->pipeline_dequant_mul_mat_vec_f32[GGML_TYPE_Q5_1], "mul_mat_vec_q5_1_f32", mul_mat_vec_q5_1_f32_len, mul_mat_vec_q5_1_f32_data, "main", 3, 3 * sizeof(uint32_t), {1, 1, 1}, { device->subgroup_size }, 1);
-    ggml_vk_create_pipeline(ctx, ctx->device->pipeline_dequant_mul_mat_vec_f32[GGML_TYPE_Q8_0], "mul_mat_vec_q8_0_f32", mul_mat_vec_q8_0_f32_len, mul_mat_vec_q8_0_f32_data, "main", 3, 3 * sizeof(uint32_t), {1, 1, 1}, { device->subgroup_size }, 1);
-    ggml_vk_create_pipeline(ctx, ctx->device->pipeline_dequant_mul_mat_vec_f32[GGML_TYPE_Q2_K], "mul_mat_vec_q2_K_f32", mul_mat_vec_q2_K_f32_len, mul_mat_vec_q2_K_f32_data, "main", 3, 3 * sizeof(uint32_t), {1, 1, 1}, { device->subgroup_size }, 1);
-    ggml_vk_create_pipeline(ctx, ctx->device->pipeline_dequant_mul_mat_vec_f32[GGML_TYPE_Q3_K], "mul_mat_vec_q3_K_f32", mul_mat_vec_q3_K_f32_len, mul_mat_vec_q3_K_f32_data, "main", 3, 3 * sizeof(uint32_t), {1, 1, 1}, { device->subgroup_size }, 1);
-    ggml_vk_create_pipeline(ctx, ctx->device->pipeline_dequant_mul_mat_vec_f32[GGML_TYPE_Q4_K], "mul_mat_vec_q4_K_f32", mul_mat_vec_q4_K_f32_len, mul_mat_vec_q4_K_f32_data, "main", 3, 3 * sizeof(uint32_t), {1, 1, 1}, { device->subgroup_size }, 1);
-    ggml_vk_create_pipeline(ctx, ctx->device->pipeline_dequant_mul_mat_vec_f32[GGML_TYPE_Q5_K], "mul_mat_vec_q5_K_f32", mul_mat_vec_q5_K_f32_len, mul_mat_vec_q5_K_f32_data, "main", 3, 3 * sizeof(uint32_t), {1, 1, 1}, { device->subgroup_size }, 1);
-    ggml_vk_create_pipeline(ctx, ctx->device->pipeline_dequant_mul_mat_vec_f32[GGML_TYPE_Q6_K], "mul_mat_vec_q6_K_f32", mul_mat_vec_q6_K_f32_len, mul_mat_vec_q6_K_f32_data, "main", 3, 3 * sizeof(uint32_t), {1, 1, 1}, { device->subgroup_size }, 1);
+        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);
+        ggml_vk_create_pipeline(ctx, ctx->device->pipeline_matmul_f32->s, "matmul_f32_s", matmul_f32_fp32_len, matmul_f32_fp32_data, "main", 3, sizeof(vk_mat_mat_push_constants), s_wg_denoms, warptile_s, 1);
+        ggml_vk_create_pipeline(ctx, ctx->device->pipeline_matmul_f32->a_l, "matmul_f32_aligned_l", matmul_f32_aligned_fp32_len, matmul_f32_aligned_fp32_data, "main", 3, sizeof(vk_mat_mat_push_constants), l_wg_denoms, warptile_l, l_align);
+        ggml_vk_create_pipeline(ctx, ctx->device->pipeline_matmul_f32->a_m, "matmul_f32_aligned_m", matmul_f32_aligned_fp32_len, matmul_f32_aligned_fp32_data, "main", 3, sizeof(vk_mat_mat_push_constants), m_wg_denoms, warptile_m, m_align);
+        ggml_vk_create_pipeline(ctx, ctx->device->pipeline_matmul_f32->a_s, "matmul_f32_aligned_s", matmul_f32_aligned_fp32_len, matmul_f32_aligned_fp32_data, "main", 3, sizeof(vk_mat_mat_push_constants), s_wg_denoms, warptile_s, s_align);
+
+        ggml_vk_create_pipeline(ctx, ctx->device->pipeline_matmul_f16->l, "matmul_f16_l", matmul_f16_fp32_len, matmul_f16_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_f16->m, "matmul_f16_m", matmul_f16_fp32_len, matmul_f16_fp32_data, "main", 3, sizeof(vk_mat_mat_push_constants), m_wg_denoms, warptile_m, 1);
+        ggml_vk_create_pipeline(ctx, ctx->device->pipeline_matmul_f16->s, "matmul_f16_s", matmul_f16_fp32_len, matmul_f16_fp32_data, "main", 3, sizeof(vk_mat_mat_push_constants), s_wg_denoms, warptile_s, 1);
+        ggml_vk_create_pipeline(ctx, ctx->device->pipeline_matmul_f16->a_l, "matmul_f16_aligned_l", matmul_f16_aligned_fp32_len, matmul_f16_aligned_fp32_data, "main", 3, sizeof(vk_mat_mat_push_constants), l_wg_denoms, warptile_l, l_align);
+        ggml_vk_create_pipeline(ctx, ctx->device->pipeline_matmul_f16->a_m, "matmul_f16_aligned_m", matmul_f16_aligned_fp32_len, matmul_f16_aligned_fp32_data, "main", 3, sizeof(vk_mat_mat_push_constants), m_wg_denoms, warptile_m, m_align);
+        ggml_vk_create_pipeline(ctx, ctx->device->pipeline_matmul_f16->a_s, "matmul_f16_aligned_s", matmul_f16_aligned_fp32_len, matmul_f16_aligned_fp32_data, "main", 3, sizeof(vk_mat_mat_push_constants), s_wg_denoms, warptile_s, s_align);
+
+        ggml_vk_create_pipeline(ctx, ctx->device->pipeline_matmul_f16_f32->l, "matmul_f16_f32_l", matmul_f16_f32_fp32_len, matmul_f16_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_f16_f32->m, "matmul_f16_f32_m", matmul_f16_f32_fp32_len, matmul_f16_f32_fp32_data, "main", 3, sizeof(vk_mat_mat_push_constants), m_wg_denoms, warptile_m, 1);
+        ggml_vk_create_pipeline(ctx, ctx->device->pipeline_matmul_f16_f32->s, "matmul_f16_f32_s", matmul_f16_f32_fp32_len, matmul_f16_f32_fp32_data, "main", 3, sizeof(vk_mat_mat_push_constants), s_wg_denoms, warptile_s, 1);
+        ggml_vk_create_pipeline(ctx, ctx->device->pipeline_matmul_f16_f32->a_l, "matmul_f16_f32_aligned_l", matmul_f16_f32_aligned_fp32_len, matmul_f16_f32_aligned_fp32_data, "main", 3, sizeof(vk_mat_mat_push_constants), l_wg_denoms, warptile_l, l_align);
+        ggml_vk_create_pipeline(ctx, ctx->device->pipeline_matmul_f16_f32->a_m, "matmul_f16_f32_aligned_m", matmul_f16_f32_aligned_fp32_len, matmul_f16_f32_aligned_fp32_data, "main", 3, sizeof(vk_mat_mat_push_constants), m_wg_denoms, warptile_m, m_align);
+        ggml_vk_create_pipeline(ctx, ctx->device->pipeline_matmul_f16_f32->a_s, "matmul_f16_f32_aligned_s", matmul_f16_f32_aligned_fp32_len, matmul_f16_f32_aligned_fp32_data, "main", 3, 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[GGML_TYPE_Q4_0]->l, "matmul_q4_0_f32_l", matmul_q4_0_f32_fp32_len, matmul_q4_0_f32_fp32_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_0]->m, "matmul_q4_0_f32_m", matmul_q4_0_f32_fp32_len, matmul_q4_0_f32_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_Q4_0]->s, "matmul_q4_0_f32_s", matmul_q4_0_f32_fp32_len, matmul_q4_0_f32_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_dequant_mul_mat_mat[GGML_TYPE_Q4_0]->a_l, "matmul_q4_0_f32_aligned_l", matmul_q4_0_f32_aligned_fp32_len, matmul_q4_0_f32_aligned_fp32_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_0]->a_m, "matmul_q4_0_f32_aligned_m", matmul_q4_0_f32_aligned_fp32_len, matmul_q4_0_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_Q4_0]->a_s, "matmul_q4_0_f32_aligned_s", matmul_q4_0_f32_aligned_fp32_len, matmul_q4_0_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_dequant_mul_mat_mat[GGML_TYPE_Q4_1]->l, "matmul_q4_1_f32_l", matmul_q4_1_f32_fp32_len, matmul_q4_1_f32_fp32_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_fp32_len, matmul_q4_1_f32_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_Q4_1]->s, "matmul_q4_1_f32_s", matmul_q4_1_f32_fp32_len, matmul_q4_1_f32_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_dequant_mul_mat_mat[GGML_TYPE_Q4_1]->a_l, "matmul_q4_1_f32_aligned_l", matmul_q4_1_f32_aligned_fp32_len, matmul_q4_1_f32_aligned_fp32_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_fp32_len, matmul_q4_1_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_Q4_1]->a_s, "matmul_q4_1_f32_aligned_s", matmul_q4_1_f32_aligned_fp32_len, matmul_q4_1_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_dequant_mul_mat_mat[GGML_TYPE_Q5_0]->l, "matmul_q5_0_f32_l", matmul_q5_0_f32_fp32_len, matmul_q5_0_f32_fp32_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_fp32_len, matmul_q5_0_f32_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_Q5_0]->s, "matmul_q5_0_f32_s", matmul_q5_0_f32_fp32_len, matmul_q5_0_f32_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_dequant_mul_mat_mat[GGML_TYPE_Q5_0]->a_l, "matmul_q5_0_f32_aligned_l", matmul_q5_0_f32_aligned_fp32_len, matmul_q5_0_f32_aligned_fp32_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]->a_m, "matmul_q5_0_f32_aligned_m", matmul_q5_0_f32_aligned_fp32_len, matmul_q5_0_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_Q5_0]->a_s, "matmul_q5_0_f32_aligned_s", matmul_q5_0_f32_aligned_fp32_len, matmul_q5_0_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_dequant_mul_mat_mat[GGML_TYPE_Q5_1]->l, "matmul_q5_1_f32_l", matmul_q5_1_f32_fp32_len, matmul_q5_1_f32_fp32_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_1]->m, "matmul_q5_1_f32_m", matmul_q5_1_f32_fp32_len, matmul_q5_1_f32_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_Q5_1]->s, "matmul_q5_1_f32_s", matmul_q5_1_f32_fp32_len, matmul_q5_1_f32_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_dequant_mul_mat_mat[GGML_TYPE_Q5_1]->a_l, "matmul_q5_1_f32_aligned_l", matmul_q5_1_f32_aligned_fp32_len, matmul_q5_1_f32_aligned_fp32_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_1]->a_m, "matmul_q5_1_f32_aligned_m", matmul_q5_1_f32_aligned_fp32_len, matmul_q5_1_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_Q5_1]->a_s, "matmul_q5_1_f32_aligned_s", matmul_q5_1_f32_aligned_fp32_len, matmul_q5_1_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_dequant_mul_mat_mat[GGML_TYPE_Q8_0]->l, "matmul_q8_0_f32_l", matmul_q8_0_f32_fp32_len, matmul_q8_0_f32_fp32_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_Q8_0]->m, "matmul_q8_0_f32_m", matmul_q8_0_f32_fp32_len, matmul_q8_0_f32_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_Q8_0]->s, "matmul_q8_0_f32_s", matmul_q8_0_f32_fp32_len, matmul_q8_0_f32_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_dequant_mul_mat_mat[GGML_TYPE_Q8_0]->a_l, "matmul_q8_0_f32_aligned_l", matmul_q8_0_f32_aligned_fp32_len, matmul_q8_0_f32_aligned_fp32_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_Q8_0]->a_m, "matmul_q8_0_f32_aligned_m", matmul_q8_0_f32_aligned_fp32_len, matmul_q8_0_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_Q8_0]->a_s, "matmul_q8_0_f32_aligned_s", matmul_q8_0_f32_aligned_fp32_len, matmul_q8_0_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_dequant_mul_mat_mat[GGML_TYPE_Q2_K]->l, "matmul_q2_k_f32_l", matmul_q2_k_f32_fp32_len, matmul_q2_k_f32_fp32_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_Q2_K]->m, "matmul_q2_k_f32_m", matmul_q2_k_f32_fp32_len, matmul_q2_k_f32_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_Q2_K]->s, "matmul_q2_k_f32_s", matmul_q2_k_f32_fp32_len, matmul_q2_k_f32_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_dequant_mul_mat_mat[GGML_TYPE_Q2_K]->a_l, "matmul_q2_k_f32_aligned_l", matmul_q2_k_f32_aligned_fp32_len, matmul_q2_k_f32_aligned_fp32_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_Q2_K]->a_m, "matmul_q2_k_f32_aligned_m", matmul_q2_k_f32_aligned_fp32_len, matmul_q2_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_Q2_K]->a_s, "matmul_q2_k_f32_aligned_s", matmul_q2_k_f32_aligned_fp32_len, matmul_q2_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_dequant_mul_mat_mat[GGML_TYPE_Q3_K]->l, "matmul_q3_k_f32_l", matmul_q3_k_f32_fp32_len, matmul_q3_k_f32_fp32_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_Q3_K]->m, "matmul_q3_k_f32_m", matmul_q3_k_f32_fp32_len, matmul_q3_k_f32_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_Q3_K]->s, "matmul_q3_k_f32_s", matmul_q3_k_f32_fp32_len, matmul_q3_k_f32_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_dequant_mul_mat_mat[GGML_TYPE_Q3_K]->a_l, "matmul_q3_k_f32_aligned_l", matmul_q3_k_f32_aligned_fp32_len, matmul_q3_k_f32_aligned_fp32_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_Q3_K]->a_m, "matmul_q3_k_f32_aligned_m", matmul_q3_k_f32_aligned_fp32_len, matmul_q3_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_Q3_K]->a_s, "matmul_q3_k_f32_aligned_s", matmul_q3_k_f32_aligned_fp32_len, matmul_q3_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_dequant_mul_mat_mat[GGML_TYPE_Q4_K]->l, "matmul_q4_k_f32_l", matmul_q4_k_f32_fp32_len, matmul_q4_k_f32_fp32_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_K]->m, "matmul_q4_k_f32_m", matmul_q4_k_f32_fp32_len, matmul_q4_k_f32_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_Q4_K]->s, "matmul_q4_k_f32_s", matmul_q4_k_f32_fp32_len, matmul_q4_k_f32_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_dequant_mul_mat_mat[GGML_TYPE_Q4_K]->a_l, "matmul_q4_k_f32_aligned_l", matmul_q4_k_f32_aligned_fp32_len, matmul_q4_k_f32_aligned_fp32_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_K]->a_m, "matmul_q4_k_f32_aligned_m", matmul_q4_k_f32_aligned_fp32_len, matmul_q4_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_Q4_K]->a_s, "matmul_q4_k_f32_aligned_s", matmul_q4_k_f32_aligned_fp32_len, matmul_q4_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_dequant_mul_mat_mat[GGML_TYPE_Q5_K]->l, "matmul_q5_k_f32_l", matmul_q5_k_f32_fp32_len, matmul_q5_k_f32_fp32_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_K]->m, "matmul_q5_k_f32_m", matmul_q5_k_f32_fp32_len, matmul_q5_k_f32_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_Q5_K]->s, "matmul_q5_k_f32_s", matmul_q5_k_f32_fp32_len, matmul_q5_k_f32_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_dequant_mul_mat_mat[GGML_TYPE_Q5_K]->a_l, "matmul_q5_k_f32_aligned_l", matmul_q5_k_f32_aligned_fp32_len, matmul_q5_k_f32_aligned_fp32_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_K]->a_m, "matmul_q5_k_f32_aligned_m", matmul_q5_k_f32_aligned_fp32_len, matmul_q5_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_Q5_K]->a_s, "matmul_q5_k_f32_aligned_s", matmul_q5_k_f32_aligned_fp32_len, matmul_q5_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_dequant_mul_mat_mat[GGML_TYPE_Q6_K]->l, "matmul_q6_k_f32_l", matmul_q6_k_f32_fp32_len, matmul_q6_k_f32_fp32_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_Q6_K]->m, "matmul_q6_k_f32_m", matmul_q6_k_f32_fp32_len, matmul_q6_k_f32_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]->s, "matmul_q6_k_f32_s", matmul_q6_k_f32_fp32_len, matmul_q6_k_f32_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_dequant_mul_mat_mat[GGML_TYPE_Q6_K]->a_l, "matmul_q6_k_f32_aligned_l", matmul_q6_k_f32_aligned_fp32_len, matmul_q6_k_f32_aligned_fp32_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_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);*/
+    }
+
+    // mul mat vec
+    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);
+    ggml_vk_create_pipeline(ctx, ctx->device->pipeline_dequant_mul_mat_vec_f32_f32[GGML_TYPE_Q5_0], "mul_mat_vec_q5_0_f32_f32", mul_mat_vec_q5_0_f32_f32_len, mul_mat_vec_q5_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_Q5_1], "mul_mat_vec_q5_1_f32_f32", mul_mat_vec_q5_1_f32_f32_len, mul_mat_vec_q5_1_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_Q8_0], "mul_mat_vec_q8_0_f32_f32", mul_mat_vec_q8_0_f32_f32_len, mul_mat_vec_q8_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_Q2_K], "mul_mat_vec_q2_K_f32_f32", mul_mat_vec_q2_K_f32_f32_len, mul_mat_vec_q2_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_Q3_K], "mul_mat_vec_q3_K_f32_f32", mul_mat_vec_q3_K_f32_f32_len, mul_mat_vec_q3_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_Q4_K], "mul_mat_vec_q4_K_f32_f32", mul_mat_vec_q4_K_f32_f32_len, mul_mat_vec_q4_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_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_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);
+    ggml_vk_create_pipeline(ctx, ctx->device->pipeline_dequant_mul_mat_vec_f16_f32[GGML_TYPE_Q5_0], "mul_mat_vec_q5_0_f16_f32", mul_mat_vec_q5_0_f16_f32_len, mul_mat_vec_q5_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_Q5_1], "mul_mat_vec_q5_1_f16_f32", mul_mat_vec_q5_1_f16_f32_len, mul_mat_vec_q5_1_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_Q8_0], "mul_mat_vec_q8_0_f16_f32", mul_mat_vec_q8_0_f16_f32_len, mul_mat_vec_q8_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_Q2_K], "mul_mat_vec_q2_K_f16_f32", mul_mat_vec_q2_K_f16_f32_len, mul_mat_vec_q2_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_Q3_K], "mul_mat_vec_q3_K_f16_f32", mul_mat_vec_q3_K_f16_f32_len, mul_mat_vec_q3_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_Q4_K], "mul_mat_vec_q4_K_f16_f32", mul_mat_vec_q4_K_f16_f32_len, mul_mat_vec_q4_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_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);*/
 
     // 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);
@@ -1258,6 +1502,7 @@ static void ggml_vk_load_shaders(ggml_backend_vk_context * ctx) {
     ggml_vk_create_pipeline(ctx, ctx->device->pipeline_diag_mask_inf_f32, "diag_mask_inf_f32", diag_mask_inf_f32_len, diag_mask_inf_f32_data, "main", 2, sizeof(vk_op_diag_mask_push_constants), {512, 1, 1}, {}, 1);
 
     ggml_vk_create_pipeline(ctx, ctx->device->pipeline_soft_max_f32, "soft_max_f32", soft_max_f32_len, soft_max_f32_data, "main", 4, sizeof(vk_op_soft_max_push_constants), {1, 1, 1}, {}, 1);
+    ggml_vk_create_pipeline(ctx, ctx->device->pipeline_soft_max_f32_f16, "soft_max_f32_f16", soft_max_f32_f16_len, soft_max_f32_f16_data, "main", 4, sizeof(vk_op_soft_max_push_constants), {1, 1, 1}, {}, 1);
 
     ggml_vk_create_pipeline(ctx, ctx->device->pipeline_rope_f32, "rope_f32", rope_f32_len, rope_f32_data, "main", 3, sizeof(vk_op_rope_push_constants), {1, 512, 1}, {}, 1);
     ggml_vk_create_pipeline(ctx, ctx->device->pipeline_rope_f16, "rope_f16", rope_f16_len, rope_f16_data, "main", 3, sizeof(vk_op_rope_push_constants), {1, 512, 1}, {}, 1);
@@ -1686,11 +1931,48 @@ 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 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;
+#endif
+    if (src0_type == GGML_TYPE_F32 && src1_type == GGML_TYPE_F32) {
+        return ctx->device->pipeline_matmul_id_f32;
+    }
+    if (src0_type == GGML_TYPE_F16 && src1_type == GGML_TYPE_F32) {
+        return ctx->device->pipeline_matmul_id_f16_f32;
+    }
+    if (src0_type == GGML_TYPE_F16 && src1_type == GGML_TYPE_F16) {
+        return ctx->device->pipeline_matmul_id_f16;
+    }
+
+    GGML_ASSERT(src1_type == GGML_TYPE_F32);
+
+    switch (src0_type) {
+        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 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) {
 #ifdef GGML_VULKAN_DEBUG
     std::cerr << "ggml_vk_get_dequantize_mul_mat_vec()" << std::endl;
 #endif
-    switch (type) {
+    GGML_ASSERT(b_type == GGML_TYPE_F32 || b_type == GGML_TYPE_F16);
+
+    switch (a_type) {
         case GGML_TYPE_F16:
         case GGML_TYPE_Q4_0:
         case GGML_TYPE_Q4_1:
@@ -1707,7 +1989,7 @@ static vk_pipeline ggml_vk_get_dequantize_mul_mat_vec(ggml_backend_vk_context *
             return nullptr;
     }
 
-    return ctx->device->pipeline_dequant_mul_mat_vec_f32[type];
+    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_buffer ggml_vk_pool_malloc(ggml_backend_vk_context * ctx, size_t size) {
@@ -1913,6 +2195,9 @@ static void ggml_vk_ctx_begin(ggml_backend_vk_context * ctx, vk_context * subctx
 }
 
 static size_t ggml_vk_align_size(size_t width, size_t align) {
+#ifdef GGML_VULKAN_DEBUG
+    std::cerr << "ggml_vk_align_size(" << width << ", " << align << ")" << std::endl;
+#endif
     return CEIL_DIV(width, align) * align;
 }
 
@@ -2368,11 +2653,13 @@ static uint32_t ggml_vk_guess_split_k(int m, int n, int k) {
 #ifdef GGML_VULKAN_DEBUG
     std::cerr << "ggml_vk_guess_split_k(" << m << ", " << n << ", " << k << ")" << std::endl;
 #endif
-    if (k > 128 && (m < 128 || n < 128) && m > 2 && n > 2) {
-        return 4;
-    }
+    // if (k > 128 && (m < 128 || n < 128) && m > 2 && n > 2) {
+    //     return 4;
+    // }
 
     return 1;
+
+    GGML_UNUSED(m); GGML_UNUSED(n); GGML_UNUSED(k);
 }
 
 static vk_pipeline ggml_vk_guess_matmul_pipeline_amd(ggml_backend_vk_context * ctx, vk_matmul_pipeline& mmp, int m, int n, bool aligned) {
@@ -2424,25 +2711,58 @@ static uint32_t ggml_vk_guess_matmul_pipeline_align(ggml_backend_vk_context * ct
 #ifdef GGML_VULKAN_DEBUG
     std::cerr << "ggml_vk_guess_matmul_pipeline_align(" << m << ", " << n << ")" << std::endl;
 #endif
-    return ggml_vk_guess_matmul_pipeline(ctx, mmp, m, n, false)->align;
+    return ggml_vk_guess_matmul_pipeline(ctx, mmp, m, n, true)->align;
+}
+
+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) {
+#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;
+#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, { 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 };
+    // 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);
+    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 });
 }
 
-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) {
+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,
+        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) {
 #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;
 #endif
     ggml_vk_sync_buffers(subctx);
     if (split_k == 1) {
-        const std::array<uint32_t, 14> pc = { m, n, k, stride_a, stride_b, stride_d, k, ne02, ne12, broadcast2, broadcast3, batch_stride_a, batch_stride_b, batch_stride_d };
-        ggml_vk_dispatch_pipeline(ctx, subctx, pipeline, { a, b, d }, pc.size() * sizeof(uint32_t), pc.data(), { m, n, batch });
+        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 std::array<uint32_t, 14> 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 };
+    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, { a, b, split_k_buffer }, pc1.size() * sizeof(uint32_t), pc1.data(), { (CEIL_DIV(m, pipeline->wg_denoms[0]) * pipeline->wg_denoms[0]) * split_k, n, batch });
+    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 });
@@ -2557,7 +2877,7 @@ static void ggml_vk_mul_mat_q_f16(ggml_backend_vk_context * ctx, vk_context * su
     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;
+    const bool aligned = ne10 == kpad && ne01 > 8 && ne11 > 8;
 
     const uint32_t split_k = ggml_vk_guess_split_k(ne01, ne11, ne10);
 
@@ -2655,7 +2975,13 @@ static void ggml_vk_mul_mat_q_f16(ggml_backend_vk_context * ctx, vk_context * su
     }
 
     // compute
-    ggml_vk_matmul(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);  // NOLINT
+    ggml_vk_matmul(
+        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
+    );  // NOLINT
 
     if (dst->backend == GGML_BACKEND_TYPE_CPU) {
         // copy dst to host
@@ -2685,8 +3011,10 @@ static void ggml_vk_mul_mat_vec_q_f16(ggml_backend_vk_context * ctx, vk_context
 
     GGML_ASSERT(ne11 == 1);
 
-    const uint64_t nb2  = dst->nb[2];
-    const uint64_t nb3  = dst->nb[3];
+    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;
@@ -2718,6 +3046,9 @@ static void ggml_vk_mul_mat_vec_q_f16(ggml_backend_vk_context * ctx, vk_context
     const bool qx_needs_dequant = x_non_contig;
     const bool qy_needs_dequant = (src1->type != GGML_TYPE_F16 && !f16_f32_kernel) || y_non_contig;
 
+    // Not implemented
+    GGML_ASSERT(y_non_contig || !qy_needs_dequant);  // NOLINT
+
     const uint64_t x_ne = ne01 * ne00;
     const uint64_t y_ne = ne11 * ne10;
     const uint64_t d_ne = ne11 * ne01;
@@ -2770,7 +3101,7 @@ static void ggml_vk_mul_mat_vec_q_f16(ggml_backend_vk_context * ctx, vk_context
     } 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);
+    vk_pipeline dmmv = ggml_vk_get_dequantize_mul_mat_vec(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);
@@ -2793,43 +3124,25 @@ static void ggml_vk_mul_mat_vec_q_f16(ggml_backend_vk_context * ctx, vk_context
         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 });
     }
 
-    for (uint64_t i13 = 0; i13 < ne13; i13++) {
-        const uint64_t i03 = i13 / r3;
-        for (uint64_t i12 = 0; i12 < ne12; i12++) {
-            const uint64_t i02 = i12 / r2;
-
-            const uint64_t it_idx0 = (i03 * ne02 + i02);
-            const uint64_t it_idx1 = (i13 * ne12 + i12);
-            const uint64_t x_offset = x_buf_offset + x_sz * it_idx0;
-            const uint64_t qy_offset = qy_buf_offset + qy_sz * it_idx1;
-            const uint64_t y_offset = y_buf_offset + y_sz * it_idx1;
-            const uint64_t d_offset = d_buf_offset + d_sz * it_idx1;
-
-            const uint64_t y_buffer_offset = (y_offset / ctx->device->properties.limits.minStorageBufferOffsetAlignment) * ctx->device->properties.limits.minStorageBufferOffsetAlignment;
-            const uint64_t y_shader_offset = y_offset - y_buffer_offset;
-
-            const uint64_t d_buffer_offset = (d_offset / ctx->device->properties.limits.minStorageBufferOffsetAlignment) * ctx->device->properties.limits.minStorageBufferOffsetAlignment;
-            const uint64_t d_shader_offset = d_offset - d_buffer_offset;
-
-            if (!y_non_contig && qy_needs_dequant) {
-                const std::vector<uint32_t> pc = { (uint32_t)ne11, (uint32_t)ne10, (uint32_t)ne10, (uint32_t)ne10, (uint32_t)(y_ne / 32) };
-                ggml_vk_sync_buffers(subctx);
-                ggml_vk_dispatch_pipeline(ctx, subctx, to_fp16_vk_1, { { d_Qy, qy_offset, qy_sz }, { d_Y, y_offset, y_sz } }, pc.size() * sizeof(uint32_t), pc.data(), { (uint32_t)y_ne, 1, 1});
-            }
+    uint32_t stride_batch_x = ne00*ne01;
+    uint32_t stride_batch_y = ne10*ne11;
 
-            // compute
-            const std::array<uint32_t, 3> pc = { (uint32_t)ne00, (uint32_t)(y_shader_offset / ggml_type_size(src1->type)), (uint32_t)(d_shader_offset / ggml_type_size(dst->type))};
-            ggml_vk_sync_buffers(subctx);
-            ggml_vk_dispatch_pipeline(ctx, subctx, dmmv, { { d_X, x_offset, x_sz }, { d_Y, y_buffer_offset, y_sz + y_shader_offset }, { d_D, d_buffer_offset, d_sz + d_shader_offset } }, 3 * sizeof(int), &pc, { (uint32_t)ne01, 1, 1});
+    if (!ggml_vk_dim01_contiguous(src0) && !qx_needs_dequant) {
+        stride_batch_x = src0->nb[0] / ggml_type_size(src0->type);
+    }
 
-            if (dst->backend == GGML_BACKEND_TYPE_CPU) {
-                // copy dst to host
-                float * d = (float *) ((char *) dst->data + i12*nb2 + i13*nb3);
-                ggml_vk_sync_buffers(subctx);
-                ggml_vk_buffer_read_async(ctx, subctx, d_D, d_offset, d, sizeof(float) * d_ne);
-            }
-        }
+    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 = {
+        (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),
+    };
+    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});
 }
 
 static void ggml_vk_mul_mat_vec_p021_f16_f32(ggml_backend_vk_context * ctx, vk_context * subctx, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
@@ -3011,7 +3324,7 @@ static bool ggml_vk_can_mul_mat(const ggml_tensor * src0, const ggml_tensor * sr
            ((ne0 >= 32 && ne1 >= 32 && ne10 >= 32) || src0->backend == GGML_BACKEND_TYPE_GPU);
 }
 
-static void ggml_vk_mul_mat(ggml_backend_vk_context * ctx, vk_context * subctx, const struct ggml_tensor * src0, const struct ggml_tensor * src1, struct ggml_tensor * dst) {
+static void ggml_vk_mul_mat(ggml_backend_vk_context * ctx, vk_context * subctx, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
 #ifdef GGML_VULKAN_DEBUG
     std::cerr << "ggml_vk_mul_mat(" << src0 << ", " << src1 << ", " << dst << ")" << std::endl;
 #endif
@@ -3026,9 +3339,347 @@ static void ggml_vk_mul_mat(ggml_backend_vk_context * ctx, vk_context * subctx,
     }
 }
 
-// static void ggml_vk_mul_mat_id(ggml_backend_vk_context * ctx, vk_context * subctx, const struct ggml_tensor * src0, const struct ggml_tensor * src1, struct 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;
+#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
+
+    const uint64_t ne00 = src0->ne[0];
+    const uint64_t ne01 = src0->ne[1];
+    const uint64_t ne02 = src0->ne[2];
+    const uint64_t ne03 = src0->ne[3];
+
+    const uint64_t ne10 = src1->ne[0];
+    const uint64_t ne11 = src1->ne[1];
+    const uint64_t ne12 = src1->ne[2];
+    const uint64_t ne13 = src1->ne[3];
+
+    const uint32_t nb11 = src1->nb[1];
+
+    const uint64_t ne20 = dst->ne[0];
+    const uint64_t ne21 = dst->ne[1];
+
+    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);
+
+    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;
+
+    bool src0_uma = false;
+    bool src1_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);
+        src0_uma = d_Qx != nullptr;
+        src1_uma = d_Qy != nullptr;
+    }
+
+    const bool x_non_contig = !ggml_vk_dim01_contiguous(src0);
+    const bool y_non_contig = !ggml_vk_dim01_contiguous(src1);
+
+    const bool y_f32_kernel = src1->type == GGML_TYPE_F32 && !y_non_contig;
+
+    vk_matmul_pipeline mmp = ggml_vk_get_mul_mat_mat_id_pipeline(ctx, src0->type, y_non_contig ? GGML_TYPE_F16 : src1->type);
+
+    const bool qx_needs_dequant = mmp == nullptr || x_non_contig;
+    const bool qy_needs_dequant = (src1->type != GGML_TYPE_F16 && !y_f32_kernel) || y_non_contig;
+
+    if (mmp == nullptr) {
+        GGML_ASSERT(false);
+    }
+
+    // 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 uint32_t split_k = ggml_vk_guess_split_k(ne01, ne11, ne10);
+
+    vk_pipeline pipeline = ggml_vk_guess_matmul_pipeline(ctx, mmp, ne01, ne11, 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 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;
+    uint64_t y_buf_offset = 0;
+    if (!src0_uma) {
+        d_Qx = extra_src0->buffer_gpu.lock();
+        qx_buf_offset = extra_src0->offset;
+        GGML_ASSERT(d_Qx != nullptr);
+    }
+    if (!src1_uma) {
+        d_Qy = extra_src1->buffer_gpu.lock();
+        qy_buf_offset = extra_src1->offset;
+        GGML_ASSERT(d_Qy != nullptr);
+    }
+    if (qx_needs_dequant) {
+        d_X = ctx->prealloc_x;
+        GGML_ASSERT(d_X->size >= x_sz * ne02 * ne03);
+    } else {
+        d_X = d_Qx;
+        x_buf_offset = qx_buf_offset;
+        GGML_ASSERT(qx_sz == x_sz);
+    }
+    if (qy_needs_dequant) {
+        d_Y = ctx->prealloc_y;
+        GGML_ASSERT(d_Y->size >= y_sz * ne02 * ne03);
+    } else {
+        d_Y = d_Qy;
+        y_buf_offset = qy_buf_offset;
+        GGML_ASSERT(qy_sz == y_sz);
+    }
+
+    vk_pipeline to_fp16_vk_0 = nullptr;
+    vk_pipeline to_fp16_vk_1 = nullptr;
+
+    if (x_non_contig) {
+        to_fp16_vk_0 = ggml_vk_get_cpy_pipeline(ctx, src0->type, GGML_TYPE_F16);
+    } else {
+        to_fp16_vk_0 = ggml_vk_get_to_fp16(ctx, src0->type);
+    }
+    if (y_non_contig) {
+        to_fp16_vk_1 = ggml_vk_get_cpy_pipeline(ctx, src1->type, GGML_TYPE_F16);
+    } else {
+        to_fp16_vk_1 = ggml_vk_get_to_fp16(ctx, src1->type);
+    }
+    GGML_ASSERT(!qx_needs_dequant || to_fp16_vk_0 != nullptr);  // NOLINT
+    GGML_ASSERT(!qy_needs_dequant || to_fp16_vk_1 != nullptr);  // NOLINT
+
+    // Allocate descriptor sets
+    ggml_pipeline_allocate_descriptor_sets(ctx, pipeline, 1);
+    if (qx_needs_dequant) {
+        ggml_pipeline_allocate_descriptor_sets(ctx, to_fp16_vk_0, 1);
+    }
+    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 });
+    } else if (qx_needs_dequant) {
+        const std::vector<uint32_t> pc = { (uint32_t)ne01, (uint32_t)ne10, (uint32_t)ne10, (uint32_t)ne10, (uint32_t)(ggml_nelements(src0)) };
+        ggml_vk_sync_buffers(subctx);
+        ggml_vk_dispatch_pipeline(ctx, subctx, to_fp16_vk_0, { { d_Qx, qx_buf_offset, qx_sz * ne02 * ne03 }, { d_X, 0, x_sz * ne02 * ne03 } }, pc.size() * sizeof(uint32_t), pc.data(), { (uint32_t)(x_ne * ne02 * ne03), 1, 1});
+    }
+    if (y_non_contig) {
+        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
+    ggml_vk_matmul(
+        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
+    );  // 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) {
+#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;
+#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
+
+    const uint64_t ne00 = src0->ne[0];
+    const uint64_t ne01 = src0->ne[1];
+    const uint64_t ne02 = src0->ne[2];
+    const uint64_t ne03 = src0->ne[3];
+
+    const uint64_t ne10 = src1->ne[0];
+    const uint64_t ne11 = src1->ne[1];
+    const uint64_t ne12 = src1->ne[2];
+    const uint64_t ne13 = src1->ne[3];
+
+    GGML_ASSERT(ne11 == 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;
+
+    vk_buffer d_Qx;
+    size_t qx_buf_offset = 0;
+    vk_buffer d_Qy;
+    size_t qy_buf_offset = 0;
+
+    bool src0_uma = false;
+    bool src1_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);
+        src0_uma = d_Qx != nullptr;
+        src1_uma = d_Qy != nullptr;
+    }
+
+    const bool x_non_contig = !ggml_vk_dim01_contiguous(src0);
+    const bool y_non_contig = !ggml_vk_dim01_contiguous(src1);
+
+    const bool f16_f32_kernel = src1->type == GGML_TYPE_F32;
+
+    const bool qx_needs_dequant = x_non_contig;
+    const bool qy_needs_dequant = (src1->type != GGML_TYPE_F16 && !f16_f32_kernel) || y_non_contig;
+
+    // Not implemented
+    GGML_ASSERT(y_non_contig || !qy_needs_dequant);  // NOLINT
+
+    const uint64_t x_ne = ne01 * ne00;
+    const uint64_t y_ne = ne11 * ne10;
+    const uint64_t d_ne = ne11 * ne01;
+
+    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 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);
+    vk_buffer d_X;
+    uint64_t x_buf_offset = 0;
+    vk_buffer d_Y;
+    uint64_t y_buf_offset = 0;
+    if(!src0_uma) {
+        d_Qx = extra_src0->buffer_gpu.lock();
+        qx_buf_offset = extra_src0->offset;
+        GGML_ASSERT(d_Qx != nullptr);
+    }
+    if(!src1_uma) {
+        d_Qy = extra_src1->buffer_gpu.lock();
+        qy_buf_offset = extra_src1->offset;
+        GGML_ASSERT(d_Qy != nullptr);
+    }
+    if (qx_needs_dequant) {
+        d_X = ctx->prealloc_x;
+    } else {
+        d_X = d_Qx;
+        x_buf_offset = qx_buf_offset;
+        GGML_ASSERT(qx_sz == x_sz);
+    }
+    if (qy_needs_dequant) {
+        d_Y = ctx->prealloc_y;
+    } else {
+        d_Y = d_Qy;
+        y_buf_offset = qy_buf_offset;
+        GGML_ASSERT(qy_sz == y_sz);
+    }
+
+    vk_pipeline to_fp16_vk_0 = nullptr;
+    vk_pipeline to_fp16_vk_1 = nullptr;
+    if (x_non_contig) {
+        to_fp16_vk_0 = ggml_vk_get_cpy_pipeline(ctx, src0->type, src0->type);
+    }
+    if (y_non_contig) {
+        to_fp16_vk_1 = ggml_vk_get_cpy_pipeline(ctx, src1->type, src1->type);
+    } 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);
+    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);
+
+    // Allocate descriptor sets
+    if (qx_needs_dequant) {
+        ggml_pipeline_allocate_descriptor_sets(ctx, to_fp16_vk_0, 1);
+    }
+    if (qy_needs_dequant) {
+        ggml_pipeline_allocate_descriptor_sets(ctx, to_fp16_vk_1, y_non_contig ? 1 : ne12 * ne13);
+    }
+    ggml_pipeline_allocate_descriptor_sets(ctx, dmmv, ne12 * ne13);
+
+    if (x_non_contig) {
+        GGML_ASSERT(x_sz == ggml_vk_align_size(ggml_type_size(src0->type) * x_ne, ctx->device->properties.limits.minStorageBufferOffsetAlignment));
+        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 });
+    }
+    if (y_non_contig) {
+        GGML_ASSERT(y_sz == ggml_type_size(src1->type) * y_ne);
+        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 = {
+        (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
+    };
+    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});
+}*/
 
 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
@@ -3178,14 +3829,15 @@ static vk_pipeline ggml_vk_op_get_pipeline(ggml_backend_vk_context * ctx, const
         }
         return nullptr;
     case GGML_OP_SOFT_MAX:
-#pragma message("TODO: add ggml_vk_soft_max() F16 src1 and src2 support")
-#pragma message("ref:  https://github.com/ggerganov/llama.cpp/pull/5021")
-        GGML_ASSERT(!src1 || src1->type == GGML_TYPE_F32);
-        GGML_ASSERT(!src2 || src2->type == GGML_TYPE_F32);
+        GGML_ASSERT(!src1 || src1->type == GGML_TYPE_F32 || src1->type == GGML_TYPE_F16);
+        GGML_ASSERT(!src2 || src2->type == GGML_TYPE_F32 || src2->type == GGML_TYPE_F16);
 
         if (src0->type == GGML_TYPE_F32 && (src1 == nullptr || src1->type == GGML_TYPE_F32) && (src2 == nullptr || src2->type == GGML_TYPE_F32) && dst->type == GGML_TYPE_F32) {
             return ctx->device->pipeline_soft_max_f32;
         }
+        if (src0->type == GGML_TYPE_F32 && src1->type == GGML_TYPE_F16 && src2->type == GGML_TYPE_F16 && dst->type == GGML_TYPE_F32) {
+            return ctx->device->pipeline_soft_max_f32_f16;
+        }
         return nullptr;
     case GGML_OP_ROPE:
         {
@@ -3233,6 +3885,21 @@ static ggml_vk_func_t ggml_vk_op_get_func(ggml_op op) {
     }
 }
 
+static bool ggml_vk_op_supports_incontiguous(ggml_op op) {
+    switch (op) {
+    case GGML_OP_CPY:
+    case GGML_OP_GET_ROWS:
+    case GGML_OP_ADD:
+    case GGML_OP_MUL:
+    case GGML_OP_SCALE:
+    case GGML_OP_SQR:
+    case GGML_OP_CLAMP:
+        return true;
+    default:
+        return false;
+    }
+}
+
 template<typename PC>
 static void ggml_vk_op_f32(ggml_backend_vk_context * ctx, vk_context * subctx, const ggml_tensor * src0, const ggml_tensor * src1, const ggml_tensor * src2, ggml_tensor * dst, ggml_op op, const PC&& pc) {
 #ifdef GGML_VULKAN_DEBUG
@@ -3284,6 +3951,8 @@ static void ggml_vk_op_f32(ggml_backend_vk_context * ctx, vk_context * subctx, c
         return;
     }
 
+    const bool op_supports_incontiguous = ggml_vk_op_supports_incontiguous(op);
+
     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 = use_src1 ? (ggml_tensor_extra_gpu *) src1->extra : nullptr;
@@ -3345,7 +4014,7 @@ static void ggml_vk_op_f32(ggml_backend_vk_context * ctx, vk_context * subctx, c
         GGML_ASSERT(d_Z != nullptr);
     }
 
-    if (op == GGML_OP_CPY || op == GGML_OP_GET_ROWS) {
+    if (op_supports_incontiguous) {
         x_sz = ggml_nbytes(src0);
         y_sz = use_src1 ? ggml_nbytes(src1) : 0;
         d_sz = ggml_nbytes(dst);
@@ -3364,7 +4033,7 @@ static void ggml_vk_op_f32(ggml_backend_vk_context * ctx, vk_context * subctx, c
     std::array<uint32_t, 3> elements;
 
     // Single call if dimension 2 is contiguous
-    if (op == GGML_OP_CPY || op == GGML_OP_GET_ROWS || (ggml_is_contiguous(src0) && (src1 == nullptr || ggml_is_contiguous(src1)))) {
+    if (op_supports_incontiguous || (ggml_is_contiguous(src0) && (src1 == nullptr || ggml_is_contiguous(src1)))) {
         ggml_pipeline_allocate_descriptor_sets(ctx, pipeline, 1);
 
         switch (dst->op) {
@@ -3385,7 +4054,7 @@ static void ggml_vk_op_f32(ggml_backend_vk_context * ctx, vk_context * subctx, c
             break;
         }
 
-        if (op != GGML_OP_CPY && op != GGML_OP_GET_ROWS) {
+        if (!op_supports_incontiguous) {
             if (x_sz != VK_WHOLE_SIZE) {
                 x_sz *= ne02 * ne03;
             }
@@ -3403,14 +4072,14 @@ static void ggml_vk_op_f32(ggml_backend_vk_context * ctx, vk_context * subctx, c
             if (use_src1) {
                 subbuf_y = { d_Y, y_buf_offset, y_sz };
             } else {
-                subbuf_y = { ctx->prealloc_y, 0, ctx->prealloc_y->size };
+                subbuf_y = { d_X, 0, d_X->size };
             }
 
             vk_subbuffer subbuf_z;
             if (use_src2) {
                 subbuf_z = { d_Z, z_buf_offset, z_sz };
             } else {
-                subbuf_z = { ctx->prealloc_y, 0, ctx->prealloc_y->size };
+                subbuf_z = { d_X, 0, d_X->size };
             }
 
             ggml_vk_sync_buffers(subctx);
@@ -3582,7 +4251,9 @@ static void ggml_vk_cpy(ggml_backend_vk_context * ctx, vk_context * subctx, cons
 }
 
 static void ggml_vk_norm(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_NORM, { (uint32_t)src0->ne[0], (uint32_t)src0->ne[1], 0.0f, 0.0f });
+    float * op_params = (float *)dst->op_params;
+
+    ggml_vk_op_f32<vk_op_push_constants>(ctx, subctx, src0, nullptr, nullptr, dst, GGML_OP_NORM, { (uint32_t)src0->ne[0], (uint32_t)src0->ne[1], op_params[0], 0.0f });
 }
 
 static void ggml_vk_rms_norm(ggml_backend_vk_context * ctx, vk_context * subctx, const ggml_tensor * src0, ggml_tensor * dst) {
@@ -3617,7 +4288,7 @@ static void ggml_vk_soft_max(ggml_backend_vk_context * ctx, vk_context * subctx,
 
     ggml_vk_op_f32<vk_op_soft_max_push_constants>(ctx, subctx, src0, src1, src2, dst, GGML_OP_SOFT_MAX, {
         ncols,
-        nrows_y,
+        src1 != nullptr ? nrows_y : (uint32_t)0,
         src2 != nullptr ? (uint32_t)1 : (uint32_t)0,
         scale, max_bias,
         m0, m1,
@@ -3834,7 +4505,10 @@ static void ggml_vk_test_matmul(ggml_backend_vk_context * ctx, size_t m, size_t
     vk_context * subctx = ggml_vk_create_context(ctx, ctx->device->compute_queue);
     for (size_t i = 0; i < num_it; i++) {
         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);
+        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
+        );
         ggml_vk_ctx_end(subctx);
     }
 
@@ -4339,7 +5013,10 @@ static void ggml_vk_test_dequant_matmul(ggml_backend_vk_context * ctx, size_t m,
     vk_context * subctx = ggml_vk_create_context(ctx, ctx->device->compute_queue);
     for (size_t i = 0; i < num_it; i++) {
         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);
+        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
+        );
         ggml_vk_ctx_end(subctx);
     }
 
@@ -4590,6 +5267,8 @@ static void ggml_vk_preallocate_buffers(ggml_backend_vk_context * ctx) {
     ggml_vk_test_dequant(ctx, 7680, GGML_TYPE_Q5_K);
     ggml_vk_test_dequant(ctx, 7680, GGML_TYPE_Q6_K);
 
+    ggml_vk_test_matmul<ggml_fp16_t, ggml_fp16_t>(ctx, 512, 512, 100, 32, 100, 1, 2);
+
     ggml_vk_test_matmul<float, float>(ctx, 128, 512, 512, 2, 100, 1, 0);
     ggml_vk_test_matmul<float, float>(ctx, 128, 512, 512, 2, 100, 1, 1);
     ggml_vk_test_matmul<float, float>(ctx, 128, 512, 512, 2, 100, 1, 2);
@@ -4739,7 +5418,7 @@ static void ggml_vk_preallocate_buffers(ggml_backend_vk_context * ctx) {
 }
 
 static void ggml_vk_build_graph(ggml_backend_vk_context * ctx, ggml_tensor * node, bool last_node){
-    if (ctx->disable || node->backend != GGML_BACKEND_TYPE_GPU) {
+    if (ctx->disable || node->backend != GGML_BACKEND_TYPE_GPU || ggml_is_empty(node)) {
         return;
     }
 
@@ -5539,7 +6218,7 @@ GGML_CALL static ggml_status ggml_backend_vk_graph_compute(ggml_backend_t backen
     int last_node = cgraph->n_nodes - 1;
 
     // If the last op in the cgraph isn't backend GPU, the command buffer doesn't get closed properly
-    while (last_node > 0 && cgraph->nodes[last_node]->backend != GGML_BACKEND_TYPE_GPU) {
+    while (last_node > 0 && (cgraph->nodes[last_node]->backend != GGML_BACKEND_TYPE_GPU || ggml_is_empty(cgraph->nodes[last_node]))) {
         last_node -= 1;
     }
 
@@ -5577,6 +6256,8 @@ GGML_CALL static ggml_status ggml_backend_vk_graph_compute(ggml_backend_t backen
 }
 
 GGML_CALL static bool ggml_backend_vk_supports_op(ggml_backend_t backend, const ggml_tensor * op) {
+    // ggml_backend_vk_context * ctx = (ggml_backend_vk_context *) backend->context;
+
     switch (op->op) {
         case GGML_OP_UNARY:
             switch (ggml_get_unary_op(op)) {
@@ -6199,7 +6880,7 @@ static void ggml_vk_check_results_0(ggml_backend_vk_context * ctx, ggml_compute_
             tensor_clone = ggml_soft_max(ggml_ctx, src0_clone);
         }
     } else if (tensor->op == GGML_OP_DIAG_MASK_INF) {
-        tensor_clone = ggml_diag_mask_inf(ggml_ctx, src0_clone, *(float *)tensor->op_params);
+        tensor_clone = ggml_diag_mask_inf(ggml_ctx, src0_clone, *(int *)tensor->op_params);
     } else if (tensor->op == GGML_OP_ROPE) {
         const int n_dims      = ((int32_t *) tensor->op_params)[1];
         const int mode        = ((int32_t *) tensor->op_params)[2];
@@ -6247,6 +6928,8 @@ static void ggml_vk_check_results_0(ggml_backend_vk_context * ctx, ggml_compute_
         tensor_clone = ggml_transpose(ggml_ctx, src0_clone);
     } else if (tensor->op == GGML_OP_GET_ROWS) {
         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 {
         std::cerr << "Missing vk_check_results OP: " << ggml_op_name(tensor->op) << std::endl;
         GGML_ASSERT(false);
@@ -6280,7 +6963,7 @@ static void ggml_vk_check_results_0(ggml_backend_vk_context * ctx, ggml_compute_
         free(src1_buffer);
     }
     if (src2 != nullptr) {
-        free(src1_buffer);
+        free(src2_buffer);
     }
 
     ggml_free(ggml_ctx);