whispercpp 1.3.3 → 1.3.4

data/ext/sources/ggml/src/ggml-opencl/ggml-opencl.cpp

@@ -25,6 +25,7 @@
 #include <vector>
 #include <string>
 #include <cmath>
+#include <map>
 #include <memory>
 #include <charconv>
 #include <mutex>
@@ -33,6 +34,7 @@
 #undef MAX
 #define MIN(a, b) ((a) < (b) ? (a) : (b))
 #define MAX(a, b) ((a) > (b) ? (a) : (b))
+#define CEIL_DIV(M, N) (((M) + (N)-1) / (N))
 
 #define UNUSED(x) (void)(x)
 
@@ -331,8 +333,10 @@ struct ggml_backend_opencl_context {
 
     cl_int alignment;
     size_t max_alloc_size;
+    size_t max_workgroup_size;
     bool fp16_support;
     bool has_vector_subgroup_broadcast;
+    bool disable_fusion;
     ggml_cl_compiler_version adreno_cl_compiler_version;
 
     int adreno_wave_size;
@@ -343,6 +347,7 @@ struct ggml_backend_opencl_context {
     cl_command_queue queue;
 
     cl_program program_add;
+    cl_program program_add_id;
     cl_program program_clamp;
     cl_program program_cpy;
     cl_program program_cvt;
@@ -351,6 +356,7 @@ struct ggml_backend_opencl_context {
     cl_program program_gemv_noshuffle_general;
     cl_program program_gemv_noshuffle;
     cl_program program_get_rows;
+    cl_program program_set_rows;
     cl_program program_glu;
     cl_program program_im2col_f16;
     cl_program program_im2col_f32;
@@ -361,12 +367,16 @@ struct ggml_backend_opencl_context {
     cl_program program_mul_mv_q4_0_f32_1d_8x_flat;
     cl_program program_mul_mv_q4_0_f32_1d_16x_flat;
     cl_program program_mul_mv_q6_K;
+    cl_program program_mul_mv_q8_0_f32, program_mul_mv_q8_0_f32_flat;
+    cl_program program_mul_mv_mxfp4_f32;
+    cl_program program_mul_mv_mxfp4_f32_flat;
     cl_program program_mul_mv_f16_f16;
     cl_program program_mul_mv_f16_f32_1row;
     cl_program program_mul_mv_f16_f32_l4;
     cl_program program_mul_mv_f16_f32;
     cl_program program_mul_mv_f32_f32;
     cl_program program_mul;
+    cl_program program_mul_mat_f16_f32_tiled;
     cl_program program_div;
     cl_program program_sub;
     cl_program program_norm;
@@ -388,29 +398,48 @@ struct ggml_backend_opencl_context {
     cl_program program_tanh;
     cl_program program_upscale;
     cl_program program_concat;
+    cl_program program_conv_2d_f16;
+    cl_program program_conv_2d_f32;
+    cl_program program_conv_2d_f16_f32;
     cl_program program_tsembd;
     cl_program program_mul_mv_id_q4_0_f32_8x_flat;
-
-    cl_kernel kernel_add, kernel_add_row;
-    cl_kernel kernel_mul, kernel_mul_row;
-    cl_kernel kernel_div, kernel_div_row;
-    cl_kernel kernel_sub, kernel_sub_row;
+    cl_program program_mul_mv_id_q8_0_f32, program_mul_mv_id_q8_0_f32_flat;
+    cl_program program_mul_mv_id_mxfp4_f32;
+    cl_program program_mul_mv_id_mxfp4_f32_flat;
+    cl_program program_mul_mm_f32_f32_l4_lm;
+    cl_program program_mul_mm_f16_f32_l4_lm;
+
+    cl_kernel kernel_add, kernel_add_row, kernel_add_f16, kernel_add_row_f16;
+    cl_kernel kernel_mul, kernel_mul_row, kernel_mul_f16, kernel_mul_row_f16;
+    cl_kernel kernel_div, kernel_div_row, kernel_div_f16, kernel_div_row_f16;
+    cl_kernel kernel_sub, kernel_sub_row, kernel_sub_f16, kernel_sub_row_f16;
+    cl_kernel kernel_add_id;
     cl_kernel kernel_scale;
     cl_kernel kernel_silu, kernel_silu_4;
     cl_kernel kernel_gelu, kernel_gelu_4;
+    cl_kernel kernel_gelu_erf, kernel_gelu_erf_4;
     cl_kernel kernel_gelu_quick, kernel_gelu_quick_4;
     cl_kernel kernel_relu;
     cl_kernel kernel_sigmoid_f32, kernel_sigmoid_f16;
     cl_kernel kernel_clamp;
-    cl_kernel kernel_geglu, kernel_reglu, kernel_swiglu,
-              kernel_geglu_f16, kernel_reglu_f16, kernel_swiglu_f16;
-    cl_kernel kernel_norm;
-    cl_kernel kernel_rms_norm;
-    cl_kernel kernel_group_norm;
+    cl_kernel kernel_geglu, kernel_reglu, kernel_swiglu, kernel_swiglu_oai, kernel_geglu_erf, kernel_geglu_quick,
+              kernel_geglu_f16, kernel_reglu_f16, kernel_swiglu_f16, kernel_geglu_erf_f16, kernel_geglu_quick_f16;
+    cl_kernel kernel_norm, kernel_norm_mul_add;
+    cl_kernel kernel_rms_norm, kernel_rms_norm_mul;
+    cl_kernel kernel_group_norm, kernel_group_norm_mul_add;
     cl_kernel kernel_diag_mask_inf, kernel_diag_mask_inf_8;
     cl_kernel kernel_soft_max, kernel_soft_max_4;
     cl_kernel kernel_soft_max_f16, kernel_soft_max_4_f16;
+    std::map<std::pair<int, int>, cl_kernel> kernels_flash_attn_f16;
+    std::map<std::pair<int, int>, cl_kernel> kernels_flash_attn_f16_q1;
+    std::map<std::pair<int, int>, cl_kernel> kernels_flash_attn_f32;
+    std::map<std::pair<int, int>, cl_kernel> kernels_flash_attn_f32_q1;
+    std::map<std::pair<int, int>, cl_kernel> kernels_flash_attn_f32_f16;
+    std::map<std::pair<int, int>, cl_kernel> kernels_flash_attn_f32_f16_q1;
+    std::map<std::pair<int, int>, int>       kernels_flash_attn_bm;
+    std::map<std::pair<int, int>, int>       kernels_flash_attn_bn;
     cl_kernel kernel_get_rows_f32, kernel_get_rows_f16, kernel_get_rows_q4_0;
+    cl_kernel kernel_set_rows_f32_i64, kernel_set_rows_f32_i32, kernel_set_rows_f16_i64, kernel_set_rows_f16_i32;
     cl_kernel kernel_rope_norm_f32, kernel_rope_norm_f16, kernel_rope_neox_f32, kernel_rope_neox_f16;
     cl_kernel kernel_rope_multi_f32, kernel_rope_multi_f16, kernel_rope_vision_f32, kernel_rope_vision_f16;
     cl_kernel kernel_cpy_f16_f16, kernel_cpy_f16_f32, kernel_cpy_f32_f16, kernel_cpy_f32_f32;
@@ -419,12 +448,17 @@ struct ggml_backend_opencl_context {
     cl_kernel kernel_mul_mat_f16_f32_1row;
     cl_kernel kernel_mul_mat_f16_f32;
     cl_kernel kernel_mul_mat_f16_f32_l4;
+    cl_kernel kernel_mul_mat_f16_f32_tiled;
     cl_kernel kernel_mul_mat_q4_0_f32, kernel_mul_mat_q4_0_f32_v;
     cl_kernel kernel_convert_block_q4_0, kernel_restore_block_q4_0;
+    cl_kernel kernel_convert_block_mxfp4, kernel_restore_block_mxfp4;
+    cl_kernel kernel_convert_block_q8_0, kernel_restore_block_q8_0;
     cl_kernel kernel_mul_mat_q4_0_f32_8x_flat;
     cl_kernel kernel_convert_block_q4_0_noshuffle;
     cl_kernel kernel_mul_mat_q4_0_f32_1d_8x_flat, kernel_mul_mat_q4_0_f32_1d_16x_flat;
     cl_kernel kernel_mul_mv_q6_K_f32;
+    cl_kernel kernel_mul_mv_mxfp4_f32, kernel_mul_mv_mxfp4_f32_flat;
+    cl_kernel kernel_mul_mv_q8_0_f32, kernel_mul_mv_q8_0_f32_flat;
     cl_kernel kernel_im2col_f32, kernel_im2col_f16;
     cl_kernel kernel_argsort_f32_i32;
     cl_kernel kernel_sum_rows_f32;
@@ -436,8 +470,16 @@ struct ggml_backend_opencl_context {
     cl_kernel kernel_upscale_bilinear;
     cl_kernel kernel_concat_f32_contiguous;
     cl_kernel kernel_concat_f32_non_contiguous;
+    cl_kernel kernel_conv_2d_f16;
+    cl_kernel kernel_conv_2d_f32;
+    cl_kernel kernel_conv_2d_f16_f32;
     cl_kernel kernel_timestep_embedding;
     cl_kernel kernel_mul_mv_id_q4_0_f32_8x_flat;
+    cl_kernel kernel_mul_mv_id_q8_0_f32, kernel_mul_mv_id_q8_0_f32_flat;
+    cl_kernel kernel_mul_mv_id_mxfp4_f32;
+    cl_kernel kernel_mul_mv_id_mxfp4_f32_flat;
+    cl_kernel kernel_mul_mm_f32_f32_l4_lm;
+    cl_kernel kernel_mul_mm_f16_f32_l4_lm;
 
     std::vector<ProfilingInfo> profiling_info;
 
@@ -528,6 +570,16 @@ struct ggml_backend_opencl_context {
         fclose(ftrace);
     }
 
+    size_t get_kernel_workgroup_size(cl_kernel kernel) const {
+        size_t workgroup_size = 0;
+        size_t ret_size = 0;
+        CL_CHECK(
+            clGetKernelWorkGroupInfo(kernel, device, CL_KERNEL_WORK_GROUP_SIZE,
+                sizeof(size_t), &workgroup_size, &ret_size));
+        GGML_ASSERT(sizeof(size_t) == ret_size);
+        return workgroup_size;
+    }
+
     void enqueue_ndrange_kernel(cl_kernel kernel, cl_uint work_dim, size_t *global_work_size, size_t *local_work_size, const ggml_tensor * tensor) {
 #ifdef GGML_OPENCL_PROFILING
         cl_event evt;
@@ -548,6 +600,7 @@ struct ggml_backend_opencl_context {
     cl_kernel kernel_transpose_32;
     cl_kernel kernel_transpose_32_16;
     cl_kernel kernel_transpose_16;
+    cl_kernel kernel_transpose_16_4x1;
 
     cl_mem A_s_d_max;            // max scale buffer size for transpose
     cl_mem A_q_d_max;            // max weight buffer size for transpose
@@ -573,6 +626,7 @@ struct ggml_backend_opencl_context {
         if (ref_count == 0) {
 #ifdef GGML_OPENCL_PROFILING
             write_profiling_info();
+            profiling_info.clear();
 #endif
         }
     }
@@ -647,8 +701,26 @@ static void load_cl_kernels(ggml_backend_opencl_context *backend_ctx, ggml_cl_ve
         backend_ctx->program_add =
             build_program_from_source(backend_ctx->context, backend_ctx->device, kernel_src.c_str(), compile_opts);
 
-        CL_CHECK((backend_ctx->kernel_add     = clCreateKernel(backend_ctx->program_add, "kernel_add", &err), err));
-        CL_CHECK((backend_ctx->kernel_add_row = clCreateKernel(backend_ctx->program_add, "kernel_add_row", &err), err));
+        CL_CHECK((backend_ctx->kernel_add         = clCreateKernel(backend_ctx->program_add, "kernel_add", &err), err));
+        CL_CHECK((backend_ctx->kernel_add_row     = clCreateKernel(backend_ctx->program_add, "kernel_add_row", &err), err));
+        CL_CHECK((backend_ctx->kernel_add_f16     = clCreateKernel(backend_ctx->program_add, "kernel_add_f16", &err), err));
+        CL_CHECK((backend_ctx->kernel_add_row_f16 = clCreateKernel(backend_ctx->program_add, "kernel_add_row_f16", &err), err));
+        GGML_LOG_CONT(".");
+    }
+
+    // add_id
+    {
+#ifdef GGML_OPENCL_EMBED_KERNELS
+        const std::string kernel_src {
+            #include "add_id.cl.h"
+        };
+#else
+        const std::string kernel_src = read_file("add_id.cl");
+#endif
+        backend_ctx->program_add_id =
+            build_program_from_source(backend_ctx->context, backend_ctx->device, kernel_src.c_str(), compile_opts);
+
+        CL_CHECK((backend_ctx->kernel_add_id = clCreateKernel(backend_ctx->program_add_id, "kernel_add_id", &err), err));
         GGML_LOG_CONT(".");
     }
 
@@ -702,6 +774,10 @@ static void load_cl_kernels(ggml_backend_opencl_context *backend_ctx, ggml_cl_ve
         CL_CHECK((backend_ctx->kernel_convert_block_q4_0_noshuffle = clCreateKernel(backend_ctx->program_cvt, "kernel_convert_block_q4_0_noshuffle", &err), err));
         CL_CHECK((backend_ctx->kernel_convert_block_q4_0  = clCreateKernel(backend_ctx->program_cvt, "kernel_convert_block_q4_0", &err), err));
         CL_CHECK((backend_ctx->kernel_restore_block_q4_0  = clCreateKernel(backend_ctx->program_cvt, "kernel_restore_block_q4_0", &err), err));
+        CL_CHECK((backend_ctx->kernel_convert_block_mxfp4 = clCreateKernel(backend_ctx->program_cvt, "kernel_convert_block_mxfp4", &err), err));
+        CL_CHECK((backend_ctx->kernel_restore_block_mxfp4 = clCreateKernel(backend_ctx->program_cvt, "kernel_restore_block_mxfp4", &err), err));
+        CL_CHECK((backend_ctx->kernel_convert_block_q8_0  = clCreateKernel(backend_ctx->program_cvt, "kernel_convert_block_q8_0", &err), err));
+        CL_CHECK((backend_ctx->kernel_restore_block_q8_0  = clCreateKernel(backend_ctx->program_cvt, "kernel_restore_block_q8_0", &err), err));
         GGML_LOG_CONT(".");
     }
 
@@ -736,6 +812,8 @@ static void load_cl_kernels(ggml_backend_opencl_context *backend_ctx, ggml_cl_ve
 
         CL_CHECK((backend_ctx->kernel_gelu         = clCreateKernel(backend_ctx->program_gelu, "kernel_gelu", &err), err));
         CL_CHECK((backend_ctx->kernel_gelu_4       = clCreateKernel(backend_ctx->program_gelu, "kernel_gelu_4", &err), err));
+        CL_CHECK((backend_ctx->kernel_gelu_erf     = clCreateKernel(backend_ctx->program_gelu, "kernel_gelu_erf", &err), err));
+        CL_CHECK((backend_ctx->kernel_gelu_erf_4   = clCreateKernel(backend_ctx->program_gelu, "kernel_gelu_erf_4", &err), err));
         CL_CHECK((backend_ctx->kernel_gelu_quick   = clCreateKernel(backend_ctx->program_gelu, "kernel_gelu_quick", &err), err));
         CL_CHECK((backend_ctx->kernel_gelu_quick_4 = clCreateKernel(backend_ctx->program_gelu, "kernel_gelu_quick_4", &err), err));
         GGML_LOG_CONT(".");
@@ -753,12 +831,17 @@ static void load_cl_kernels(ggml_backend_opencl_context *backend_ctx, ggml_cl_ve
         backend_ctx->program_glu =
             build_program_from_source(backend_ctx->context, backend_ctx->device, kernel_src.c_str(), compile_opts);
 
-        CL_CHECK((backend_ctx->kernel_geglu      = clCreateKernel(backend_ctx->program_glu, "kernel_geglu", &err), err));
-        CL_CHECK((backend_ctx->kernel_reglu      = clCreateKernel(backend_ctx->program_glu, "kernel_reglu", &err), err));
-        CL_CHECK((backend_ctx->kernel_swiglu     = clCreateKernel(backend_ctx->program_glu, "kernel_swiglu", &err), err));
-        CL_CHECK((backend_ctx->kernel_geglu_f16  = clCreateKernel(backend_ctx->program_glu, "kernel_geglu_f16", &err), err));
-        CL_CHECK((backend_ctx->kernel_reglu_f16  = clCreateKernel(backend_ctx->program_glu, "kernel_reglu_f16", &err), err));
-        CL_CHECK((backend_ctx->kernel_swiglu_f16 = clCreateKernel(backend_ctx->program_glu, "kernel_swiglu_f16", &err), err));
+        CL_CHECK((backend_ctx->kernel_geglu           = clCreateKernel(backend_ctx->program_glu, "kernel_geglu", &err), err));
+        CL_CHECK((backend_ctx->kernel_reglu           = clCreateKernel(backend_ctx->program_glu, "kernel_reglu", &err), err));
+        CL_CHECK((backend_ctx->kernel_swiglu          = clCreateKernel(backend_ctx->program_glu, "kernel_swiglu", &err), err));
+        CL_CHECK((backend_ctx->kernel_swiglu_oai      = clCreateKernel(backend_ctx->program_glu, "kernel_swiglu_oai", &err), err));
+        CL_CHECK((backend_ctx->kernel_geglu_erf       = clCreateKernel(backend_ctx->program_glu, "kernel_geglu_erf", &err), err));
+        CL_CHECK((backend_ctx->kernel_geglu_quick     = clCreateKernel(backend_ctx->program_glu, "kernel_geglu_quick", &err), err));
+        CL_CHECK((backend_ctx->kernel_geglu_f16       = clCreateKernel(backend_ctx->program_glu, "kernel_geglu_f16", &err), err));
+        CL_CHECK((backend_ctx->kernel_reglu_f16       = clCreateKernel(backend_ctx->program_glu, "kernel_reglu_f16", &err), err));
+        CL_CHECK((backend_ctx->kernel_swiglu_f16      = clCreateKernel(backend_ctx->program_glu, "kernel_swiglu_f16", &err), err));
+        CL_CHECK((backend_ctx->kernel_geglu_erf_f16   = clCreateKernel(backend_ctx->program_glu, "kernel_geglu_erf_f16", &err), err));
+        CL_CHECK((backend_ctx->kernel_geglu_quick_f16 = clCreateKernel(backend_ctx->program_glu, "kernel_geglu_quick_f16", &err), err));
         GGML_LOG_CONT(".");
     }
 
@@ -916,6 +999,70 @@ static void load_cl_kernels(ggml_backend_opencl_context *backend_ctx, ggml_cl_ve
         GGML_LOG_CONT(".");
     }
 
+    // mul_mv_q8_0_f32
+    {
+#ifdef GGML_OPENCL_EMBED_KERNELS
+        const std::string kernel_src {
+            #include "mul_mv_q8_0_f32.cl.h"
+        };
+#else
+        const std::string kernel_src = read_file("mul_mv_q8_0_f32.cl");
+#endif
+        backend_ctx->program_mul_mv_q8_0_f32 =
+            build_program_from_source(backend_ctx->context, backend_ctx->device, kernel_src.c_str(), compile_opts);
+
+        CL_CHECK((backend_ctx->kernel_mul_mv_q8_0_f32 = clCreateKernel(backend_ctx->program_mul_mv_q8_0_f32, "kernel_mul_mv_q8_0_f32", &err), err));
+        GGML_LOG_CONT(".");
+    }
+
+    // mul_mv_q8_0_f32_flat
+    {
+#ifdef GGML_OPENCL_EMBED_KERNELS
+        const std::string kernel_src {
+            #include "mul_mv_q8_0_f32_flat.cl.h"
+        };
+#else
+        const std::string kernel_src = read_file("mul_mv_q8_0_f32_flat.cl");
+#endif
+        backend_ctx->program_mul_mv_q8_0_f32_flat =
+            build_program_from_source(backend_ctx->context, backend_ctx->device, kernel_src.c_str(), compile_opts);
+
+        CL_CHECK((backend_ctx->kernel_mul_mv_q8_0_f32_flat = clCreateKernel(backend_ctx->program_mul_mv_q8_0_f32_flat, "kernel_mul_mv_q8_0_f32_flat", &err), err));
+        GGML_LOG_CONT(".");
+    }
+
+    // mul_mv_mxfp4_f32
+    {
+#ifdef GGML_OPENCL_EMBED_KERNELS
+        const std::string kernel_src {
+            #include "mul_mv_mxfp4_f32.cl.h"
+        };
+#else
+        const std::string kernel_src = read_file("mul_mv_mxfp4_f32.cl");
+#endif
+        backend_ctx->program_mul_mv_mxfp4_f32 =
+            build_program_from_source(backend_ctx->context, backend_ctx->device, kernel_src.c_str(), compile_opts);
+
+        CL_CHECK((backend_ctx->kernel_mul_mv_mxfp4_f32 = clCreateKernel(backend_ctx->program_mul_mv_mxfp4_f32, "kernel_mul_mv_mxfp4_f32", &err), err));
+        GGML_LOG_CONT(".");
+    }
+
+    // mul_mv_mxfp4_f32_flat
+    {
+#ifdef GGML_OPENCL_EMBED_KERNELS
+        const std::string kernel_src {
+            #include "mul_mv_mxfp4_f32_flat.cl.h"
+        };
+#else
+        const std::string kernel_src = read_file("mul_mv_mxfp4_f32_flat.cl");
+#endif
+        backend_ctx->program_mul_mv_mxfp4_f32_flat =
+            build_program_from_source(backend_ctx->context, backend_ctx->device, kernel_src.c_str(), compile_opts);
+
+        CL_CHECK((backend_ctx->kernel_mul_mv_mxfp4_f32_flat = clCreateKernel(backend_ctx->program_mul_mv_mxfp4_f32_flat, "kernel_mul_mv_mxfp4_f32_flat", &err), err));
+        GGML_LOG_CONT(".");
+    }
+
     // mul_mv_f16_f16
     {
 #ifdef GGML_OPENCL_EMBED_KERNELS
@@ -996,6 +1143,54 @@ static void load_cl_kernels(ggml_backend_opencl_context *backend_ctx, ggml_cl_ve
         GGML_LOG_CONT(".");
     }
 
+    // mul_mat_f16_f32_tiled
+    {
+#ifdef GGML_OPENCL_EMBED_KERNELS
+        const std::string kernel_src {
+            #include "mul_mat_f16_f32.cl.h"
+        };
+#else
+        const std::string kernel_src = read_file("mul_mat_f16_f32.cl");
+#endif
+        backend_ctx->program_mul_mat_f16_f32_tiled =
+            build_program_from_source(backend_ctx->context, backend_ctx->device, kernel_src.c_str(), compile_opts);
+
+        CL_CHECK((backend_ctx->kernel_mul_mat_f16_f32_tiled = clCreateKernel(backend_ctx->program_mul_mat_f16_f32_tiled, "mul_mat_f16_f32", &err), err));
+        GGML_LOG_CONT(".");
+    }
+
+    // mul_mm_f32_f32_l4_lm
+    {
+#ifdef GGML_OPENCL_EMBED_KERNELS
+        const std::string kernel_src {
+            #include "mul_mm_f32_f32_l4_lm.cl.h"
+        };
+#else
+        const std::string kernel_src = read_file("mul_mm_f32_f32_l4_lm.cl");
+#endif
+        backend_ctx->program_mul_mm_f32_f32_l4_lm =
+            build_program_from_source(backend_ctx->context, backend_ctx->device, kernel_src.c_str(), compile_opts);
+
+        CL_CHECK((backend_ctx->kernel_mul_mm_f32_f32_l4_lm = clCreateKernel(backend_ctx->program_mul_mm_f32_f32_l4_lm, "kernel_mul_mm_f32_f32_l4_lm", &err), err));
+        GGML_LOG_CONT(".");
+    }
+
+    // mul_mm_f16_f32_l4_lm
+    {
+#ifdef GGML_OPENCL_EMBED_KERNELS
+        const std::string kernel_src {
+            #include "mul_mm_f16_f32_l4_lm.cl.h"
+        };
+#else
+        const std::string kernel_src = read_file("mul_mm_f16_f32_l4_lm.cl");
+#endif
+        backend_ctx->program_mul_mm_f16_f32_l4_lm =
+            build_program_from_source(backend_ctx->context, backend_ctx->device, kernel_src.c_str(), compile_opts);
+
+        CL_CHECK((backend_ctx->kernel_mul_mm_f16_f32_l4_lm = clCreateKernel(backend_ctx->program_mul_mm_f16_f32_l4_lm, "kernel_mul_mm_f16_f32_l4_lm", &err), err));
+        GGML_LOG_CONT(".");
+    }
+
     // mul
     {
 #ifdef GGML_OPENCL_EMBED_KERNELS
@@ -1008,8 +1203,10 @@ static void load_cl_kernels(ggml_backend_opencl_context *backend_ctx, ggml_cl_ve
         backend_ctx->program_mul =
             build_program_from_source(backend_ctx->context, backend_ctx->device, kernel_src.c_str(), compile_opts);
 
-        CL_CHECK((backend_ctx->kernel_mul     = clCreateKernel(backend_ctx->program_mul, "kernel_mul", &err), err));
-        CL_CHECK((backend_ctx->kernel_mul_row = clCreateKernel(backend_ctx->program_mul, "kernel_mul_row", &err), err));
+        CL_CHECK((backend_ctx->kernel_mul         = clCreateKernel(backend_ctx->program_mul, "kernel_mul", &err), err));
+        CL_CHECK((backend_ctx->kernel_mul_row     = clCreateKernel(backend_ctx->program_mul, "kernel_mul_row", &err), err));
+        CL_CHECK((backend_ctx->kernel_mul_f16     = clCreateKernel(backend_ctx->program_mul, "kernel_mul_f16", &err), err));
+        CL_CHECK((backend_ctx->kernel_mul_row_f16 = clCreateKernel(backend_ctx->program_mul, "kernel_mul_row_f16", &err), err));
         GGML_LOG_CONT(".");
     }
 
@@ -1025,7 +1222,8 @@ static void load_cl_kernels(ggml_backend_opencl_context *backend_ctx, ggml_cl_ve
         backend_ctx->program_norm =
             build_program_from_source(backend_ctx->context, backend_ctx->device, kernel_src.c_str(), compile_opts);
 
-        CL_CHECK((backend_ctx->kernel_norm = clCreateKernel(backend_ctx->program_norm, "kernel_norm", &err), err));
+        CL_CHECK((backend_ctx->kernel_norm         = clCreateKernel(backend_ctx->program_norm, "kernel_norm", &err), err));
+        CL_CHECK((backend_ctx->kernel_norm_mul_add = clCreateKernel(backend_ctx->program_norm, "kernel_norm_mul_add", &err), err));
         GGML_LOG_CONT(".");
     }
 
@@ -1057,7 +1255,8 @@ static void load_cl_kernels(ggml_backend_opencl_context *backend_ctx, ggml_cl_ve
         backend_ctx->program_rms_norm =
             build_program_from_source(backend_ctx->context, backend_ctx->device, kernel_src.c_str(), compile_opts);
 
-        CL_CHECK((backend_ctx->kernel_rms_norm = clCreateKernel(backend_ctx->program_rms_norm, "kernel_rms_norm", &err), err));
+        CL_CHECK((backend_ctx->kernel_rms_norm     = clCreateKernel(backend_ctx->program_rms_norm, "kernel_rms_norm", &err), err));
+        CL_CHECK((backend_ctx->kernel_rms_norm_mul = clCreateKernel(backend_ctx->program_rms_norm, "kernel_rms_norm_mul", &err), err));
         GGML_LOG_CONT(".");
     }
 
@@ -1181,6 +1380,73 @@ static void load_cl_kernels(ggml_backend_opencl_context *backend_ctx, ggml_cl_ve
         GGML_LOG_CONT(".");
     }
 
+    // flash_attn
+    {
+        #ifdef GGML_OPENCL_EMBED_KERNELS
+                const std::string kernel_src_f16 {
+                    #include "flash_attn_f16.cl.h"
+                };
+                const std::string kernel_src_f32 {
+                    #include "flash_attn_f32.cl.h"
+                };
+                const std::string kernel_src_f32_f16 {
+                    #include "flash_attn_f32_f16.cl.h"
+                };
+        #else
+                const std::string kernel_src_f16 = read_file("flash_attn_f16.cl");
+                const std::string kernel_src_f32 = read_file("flash_attn_f32.cl");
+                const std::string kernel_src_f32_f16 = read_file("flash_attn_f32_f16.cl");
+        #endif
+
+        if (!kernel_src_f16.empty() && !kernel_src_f32.empty() && !kernel_src_f32_f16.empty()) {
+            const struct { int dk; int dv; int bm; int bn; } fa_dims[] = {
+                { 40,  40, 32, 32}, { 64,  64, 64, 64}, { 80,  80, 64, 32}, { 96,  96, 64, 32},
+                {112, 112, 32, 32}, {128, 128, 32, 32}, {192, 128, 16, 16},
+                {192, 192, 16, 16}, {256, 256, 16, 16},
+            };
+
+            for (size_t i = 0; i < sizeof(fa_dims)/sizeof(fa_dims[0]); ++i) {
+                const int dk = fa_dims[i].dk;
+                const int dv = fa_dims[i].dv;
+                const int bm = fa_dims[i].bm;
+                const int bn = fa_dims[i].bn;
+                std::string OPTS = compile_opts +
+                    " -D DK=" + std::to_string(dk) +
+                    " -D DV=" + std::to_string(dv) +
+                    " -D BLOCK_M=" + std::to_string(bm) +
+                    " -D BLOCK_N=" + std::to_string(bn);
+
+                cl_program prog_f16 = build_program_from_source(backend_ctx->context, backend_ctx->device, kernel_src_f16.c_str(), OPTS);
+                cl_kernel k_f16, k_f16_q1;
+                CL_CHECK((k_f16 = clCreateKernel(prog_f16, "flash_attn_f16", &err), err));
+                CL_CHECK((k_f16_q1 = clCreateKernel(prog_f16, "flash_attn_f16_q1", &err), err));
+                backend_ctx->kernels_flash_attn_f16[{dk, dv}] = k_f16;
+                backend_ctx->kernels_flash_attn_f16_q1[{dk, dv}] = k_f16_q1;
+                CL_CHECK(clReleaseProgram(prog_f16));
+
+                cl_program prog_f32 = build_program_from_source(backend_ctx->context, backend_ctx->device, kernel_src_f32.c_str(), OPTS);
+                cl_kernel k_f32, k_f32_q1;
+                CL_CHECK((k_f32 = clCreateKernel(prog_f32, "flash_attn_f32", &err), err));
+                CL_CHECK((k_f32_q1 = clCreateKernel(prog_f32, "flash_attn_f32_q1", &err), err));
+                backend_ctx->kernels_flash_attn_f32[{dk, dv}] = k_f32;
+                backend_ctx->kernels_flash_attn_f32_q1[{dk, dv}] = k_f32_q1;
+                CL_CHECK(clReleaseProgram(prog_f32));
+
+                cl_program prog_f32_f16 = build_program_from_source(backend_ctx->context, backend_ctx->device, kernel_src_f32_f16.c_str(), OPTS);
+                cl_kernel k_f32_f16, k_f32_f16_q1;
+                CL_CHECK((k_f32_f16 = clCreateKernel(prog_f32_f16, "flash_attn_f32_f16", &err), err));
+                CL_CHECK((k_f32_f16_q1 = clCreateKernel(prog_f32_f16, "flash_attn_f32_f16_q1", &err), err));
+                backend_ctx->kernels_flash_attn_f32_f16[{dk, dv}] = k_f32_f16;
+                backend_ctx->kernels_flash_attn_f32_f16_q1[{dk, dv}] = k_f32_f16_q1;
+                CL_CHECK(clReleaseProgram(prog_f32_f16));
+
+                backend_ctx->kernels_flash_attn_bm[{dk, dv}] = bm;
+                backend_ctx->kernels_flash_attn_bn[{dk, dv}] = bn;
+            }
+            GGML_LOG_CONT(".");
+        }
+    }
+
     // argsort
     {
 #ifdef GGML_OPENCL_EMBED_KERNELS
@@ -1206,11 +1472,16 @@ static void load_cl_kernels(ggml_backend_opencl_context *backend_ctx, ggml_cl_ve
 #else
         const std::string kernel_src = read_file("div.cl");
 #endif
+        std::string compile_opts = std::string("-cl-std=") + opencl_c_std +
+                               " -cl-mad-enable -cl-finite-math-only ";
+
         backend_ctx->program_div =
             build_program_from_source(backend_ctx->context, backend_ctx->device, kernel_src.c_str(), compile_opts);
 
-        CL_CHECK((backend_ctx->kernel_div     = clCreateKernel(backend_ctx->program_div, "kernel_div", &err), err));
-        CL_CHECK((backend_ctx->kernel_div_row = clCreateKernel(backend_ctx->program_div, "kernel_div_row", &err), err));
+        CL_CHECK((backend_ctx->kernel_div         = clCreateKernel(backend_ctx->program_div, "kernel_div", &err), err));
+        CL_CHECK((backend_ctx->kernel_div_row     = clCreateKernel(backend_ctx->program_div, "kernel_div_row", &err), err));
+        CL_CHECK((backend_ctx->kernel_div_f16     = clCreateKernel(backend_ctx->program_div, "kernel_div_f16", &err), err));
+        CL_CHECK((backend_ctx->kernel_div_row_f16 = clCreateKernel(backend_ctx->program_div, "kernel_div_row_f16", &err), err));
         GGML_LOG_CONT(".");
     }
 
@@ -1226,8 +1497,10 @@ static void load_cl_kernels(ggml_backend_opencl_context *backend_ctx, ggml_cl_ve
         backend_ctx->program_sub =
             build_program_from_source(backend_ctx->context, backend_ctx->device, kernel_src.c_str(), compile_opts);
 
-        CL_CHECK((backend_ctx->kernel_sub     = clCreateKernel(backend_ctx->program_sub, "kernel_sub", &err), err));
-        CL_CHECK((backend_ctx->kernel_sub_row = clCreateKernel(backend_ctx->program_sub, "kernel_sub_row", &err), err));
+        CL_CHECK((backend_ctx->kernel_sub         = clCreateKernel(backend_ctx->program_sub, "kernel_sub", &err), err));
+        CL_CHECK((backend_ctx->kernel_sub_row     = clCreateKernel(backend_ctx->program_sub, "kernel_sub_row", &err), err));
+        CL_CHECK((backend_ctx->kernel_sub_f16     = clCreateKernel(backend_ctx->program_sub, "kernel_sub_f16", &err), err));
+        CL_CHECK((backend_ctx->kernel_sub_row_f16 = clCreateKernel(backend_ctx->program_sub, "kernel_sub_row_f16", &err), err));
         GGML_LOG_CONT(".");
     }
 
@@ -1276,7 +1549,8 @@ static void load_cl_kernels(ggml_backend_opencl_context *backend_ctx, ggml_cl_ve
         backend_ctx->program_group_norm =
             build_program_from_source(backend_ctx->context, backend_ctx->device, kernel_src.c_str(), compile_opts);
 
-        CL_CHECK((backend_ctx->kernel_group_norm = clCreateKernel(backend_ctx->program_group_norm, "kernel_group_norm", &err), err));
+        CL_CHECK((backend_ctx->kernel_group_norm         = clCreateKernel(backend_ctx->program_group_norm, "kernel_group_norm", &err), err));
+        CL_CHECK((backend_ctx->kernel_group_norm_mul_add = clCreateKernel(backend_ctx->program_group_norm, "kernel_group_norm_mul_add", &err), err));
         GGML_LOG_CONT(".");
     }
 
@@ -1424,6 +1698,66 @@ static void load_cl_kernels(ggml_backend_opencl_context *backend_ctx, ggml_cl_ve
         }
     }
 
+    // set_rows
+    {
+#ifdef GGML_OPENCL_EMBED_KERNELS
+        const std::string kernel_src {
+            #include "set_rows.cl.h"
+        };
+#else
+        const std::string kernel_src = read_file("set_rows.cl");
+#endif
+        backend_ctx->program_set_rows =
+            build_program_from_source(backend_ctx->context, backend_ctx->device, kernel_src.c_str(), compile_opts);
+
+        CL_CHECK((backend_ctx->kernel_set_rows_f32_i64 = clCreateKernel(backend_ctx->program_set_rows, "kernel_set_rows_f32_i64", &err), err));
+        CL_CHECK((backend_ctx->kernel_set_rows_f32_i32 = clCreateKernel(backend_ctx->program_set_rows, "kernel_set_rows_f32_i32", &err), err));
+        CL_CHECK((backend_ctx->kernel_set_rows_f16_i64 = clCreateKernel(backend_ctx->program_set_rows, "kernel_set_rows_f16_i64", &err), err));
+        CL_CHECK((backend_ctx->kernel_set_rows_f16_i32 = clCreateKernel(backend_ctx->program_set_rows, "kernel_set_rows_f16_i32", &err), err));
+        GGML_LOG_CONT(".");
+    }
+
+     // conv2d
+     {
+        #ifdef GGML_OPENCL_EMBED_KERNELS
+                const std::string kernel_src {
+                    #include "conv2d.cl.h"
+                };
+                const std::string kernel_src_f16_f32 {
+                    #include "conv2d_f16_f32.cl.h"
+                };
+        #else
+                const std::string kernel_src = read_file("conv2d.cl");
+                const std::string kernel_src_f16_f32 = read_file("conv2d_f16_f32.cl");
+        #endif
+                if (!kernel_src.empty()) {
+                    backend_ctx->program_conv_2d_f16 =
+                        build_program_from_source(backend_ctx->context, backend_ctx->device, kernel_src.c_str(), (std::string(compile_opts) + " -DUSE_FP16=1").c_str());
+                    CL_CHECK((backend_ctx->kernel_conv_2d_f16 = clCreateKernel(backend_ctx->program_conv_2d_f16, "kernel_conv_2d", &err), err));
+                    GGML_LOG_CONT(".");
+                    backend_ctx->program_conv_2d_f32 =
+                        build_program_from_source(backend_ctx->context, backend_ctx->device, kernel_src.c_str(), compile_opts);
+                    CL_CHECK((backend_ctx->kernel_conv_2d_f32 = clCreateKernel(backend_ctx->program_conv_2d_f32, "kernel_conv_2d", &err), err));
+                    GGML_LOG_CONT(".");
+                } else {
+                    GGML_LOG_WARN("ggml_opencl: conv2d kernel source not found or empty. This op will not be available.\n");
+                    backend_ctx->program_conv_2d_f16 = nullptr;
+                    backend_ctx->kernel_conv_2d_f16 = nullptr;
+                    backend_ctx->program_conv_2d_f32 = nullptr;
+                    backend_ctx->kernel_conv_2d_f32 = nullptr;
+                }
+                if (!kernel_src_f16_f32.empty()) {
+                    backend_ctx->program_conv_2d_f16_f32 =
+                        build_program_from_source(backend_ctx->context, backend_ctx->device, kernel_src_f16_f32.c_str(), compile_opts);
+                    CL_CHECK((backend_ctx->kernel_conv_2d_f16_f32 = clCreateKernel(backend_ctx->program_conv_2d_f16_f32, "kernel_conv_2d", &err), err));
+                    GGML_LOG_CONT(".");
+                } else {
+                    GGML_LOG_WARN("ggml_opencl: conv2d_f16_f32 kernel source not found or empty. This op will not be available.\n");
+                    backend_ctx->program_conv_2d_f16_f32 = nullptr;
+                    backend_ctx->kernel_conv_2d_f16_f32 = nullptr;
+                }
+    }
+
     // mul_mv_id_q4_0_f32_8x_flat
     {
 #ifdef GGML_OPENCL_EMBED_KERNELS
@@ -1440,6 +1774,70 @@ static void load_cl_kernels(ggml_backend_opencl_context *backend_ctx, ggml_cl_ve
         GGML_LOG_CONT(".");
     }
 
+    // mul_mv_id_q8_0_f32
+    {
+#ifdef GGML_OPENCL_EMBED_KERNELS
+        const std::string kernel_src {
+            #include "mul_mv_id_q8_0_f32.cl.h"
+        };
+#else
+        const std::string kernel_src = read_file("mul_mv_id_q8_0_f32.cl");
+#endif
+        backend_ctx->program_mul_mv_id_q8_0_f32 =
+            build_program_from_source(backend_ctx->context, backend_ctx->device, kernel_src.c_str(), compile_opts);
+
+        CL_CHECK((backend_ctx->kernel_mul_mv_id_q8_0_f32 = clCreateKernel(backend_ctx->program_mul_mv_id_q8_0_f32, "kernel_mul_mv_id_q8_0_f32", &err), err));
+        GGML_LOG_CONT(".");
+    }
+
+    // mul_mv_id_q8_0_f32_flat
+    {
+#ifdef GGML_OPENCL_EMBED_KERNELS
+        const std::string kernel_src {
+            #include "mul_mv_id_q8_0_f32_flat.cl.h"
+        };
+#else
+        const std::string kernel_src = read_file("mul_mv_id_q8_0_f32_flat.cl");
+#endif
+        backend_ctx->program_mul_mv_id_q8_0_f32_flat =
+            build_program_from_source(backend_ctx->context, backend_ctx->device, kernel_src.c_str(), compile_opts);
+
+        CL_CHECK((backend_ctx->kernel_mul_mv_id_q8_0_f32_flat = clCreateKernel(backend_ctx->program_mul_mv_id_q8_0_f32_flat, "kernel_mul_mv_id_q8_0_f32_flat", &err), err));
+        GGML_LOG_CONT(".");
+    }
+
+    // mul_mv_id_mxfp4_f32
+    {
+#ifdef GGML_OPENCL_EMBED_KERNELS
+        const std::string kernel_src {
+            #include "mul_mv_id_mxfp4_f32.cl.h"
+        };
+#else
+        const std::string kernel_src = read_file("mul_mv_id_mxfp4_f32.cl");
+#endif
+        backend_ctx->program_mul_mv_id_mxfp4_f32 =
+            build_program_from_source(backend_ctx->context, backend_ctx->device, kernel_src.c_str(), compile_opts);
+
+        CL_CHECK((backend_ctx->kernel_mul_mv_id_mxfp4_f32 = clCreateKernel(backend_ctx->program_mul_mv_id_mxfp4_f32, "kernel_mul_mv_id_mxfp4_f32", &err), err));
+        GGML_LOG_CONT(".");
+    }
+
+    // mul_mv_id_mxfp4_f32_flat
+    {
+#ifdef GGML_OPENCL_EMBED_KERNELS
+        const std::string kernel_src {
+            #include "mul_mv_id_mxfp4_f32_flat.cl.h"
+        };
+#else
+        const std::string kernel_src = read_file("mul_mv_id_mxfp4_f32_flat.cl");
+#endif
+        backend_ctx->program_mul_mv_id_mxfp4_f32_flat =
+            build_program_from_source(backend_ctx->context, backend_ctx->device, kernel_src.c_str(), compile_opts);
+
+        CL_CHECK((backend_ctx->kernel_mul_mv_id_mxfp4_f32_flat = clCreateKernel(backend_ctx->program_mul_mv_id_mxfp4_f32_flat, "kernel_mul_mv_id_mxfp4_f32_flat", &err), err));
+        GGML_LOG_CONT(".");
+    }
+
     // Adreno kernels
 #ifdef GGML_OPENCL_USE_ADRENO_KERNELS
     // transpose
@@ -1457,6 +1855,7 @@ static void load_cl_kernels(ggml_backend_opencl_context *backend_ctx, ggml_cl_ve
         CL_CHECK((backend_ctx->kernel_transpose_32_16 = clCreateKernel(backend_ctx->program_transpose, "kernel_transpose_32_16", &err), err));
         CL_CHECK((backend_ctx->kernel_transpose_32    = clCreateKernel(backend_ctx->program_transpose, "kernel_transpose_32", &err), err));
         CL_CHECK((backend_ctx->kernel_transpose_16    = clCreateKernel(backend_ctx->program_transpose, "kernel_transpose_16", &err), err));
+        CL_CHECK((backend_ctx->kernel_transpose_16_4x1    = clCreateKernel(backend_ctx->program_transpose, "kernel_transpose_16_4x1", &err), err));
         GGML_LOG_CONT(".");
     }
 
@@ -1895,8 +2294,8 @@ static ggml_backend_opencl_context * ggml_cl2_init(ggml_backend_dev_t dev) {
 
     backend_ctx->adreno_cl_compiler_version = get_adreno_cl_compiler_version(driver_version);
     backend_ctx->has_vector_subgroup_broadcast =
-        backend_ctx->adreno_cl_compiler_version.major >= 47 ||
-        backend_ctx->adreno_cl_compiler_version.major == 17;
+        (backend_ctx->adreno_cl_compiler_version.type == E031 && backend_ctx->adreno_cl_compiler_version.major >= 47) ||
+        (backend_ctx->adreno_cl_compiler_version.type == DX   && backend_ctx->adreno_cl_compiler_version.major >= 17);
     GGML_LOG_INFO("ggml_opencl: vector subgroup broadcast support: %s\n",
         backend_ctx->has_vector_subgroup_broadcast ? "true" : "false");
 
@@ -1933,6 +2332,9 @@ static ggml_backend_opencl_context * ggml_cl2_init(ggml_backend_dev_t dev) {
     clGetDeviceInfo(device, CL_DEVICE_MAX_MEM_ALLOC_SIZE, sizeof(size_t), &backend_ctx->max_alloc_size, NULL);
     GGML_LOG_INFO("ggml_opencl: max mem alloc size: %zu MB\n", backend_ctx->max_alloc_size/1024/1024);
 
+    clGetDeviceInfo(device, CL_DEVICE_MAX_WORK_GROUP_SIZE, sizeof(size_t), &backend_ctx->max_workgroup_size, NULL);
+    GGML_LOG_INFO("ggml_opencl: device max workgroup size: %lu\n", backend_ctx->max_workgroup_size);
+
     // Check SVM.
     cl_device_svm_capabilities svm_caps;
     CL_CHECK(clGetDeviceInfo(device, CL_DEVICE_SVM_CAPABILITIES, sizeof(cl_device_svm_capabilities), &svm_caps, 0));
@@ -2009,6 +2411,8 @@ static ggml_backend_opencl_context * ggml_cl2_init(ggml_backend_dev_t dev) {
     CL_CHECK((backend_ctx->B_d_max   = clCreateBuffer(context, 0, max_B_d_bytes,   NULL, &err), err));
 #endif // GGML_OPENCL_USE_ADRENO_KERNELS
 
+    backend_ctx->disable_fusion = getenv("GGML_OPENCL_DISABLE_FUSION") != nullptr;
+
     dev_ctx->backend_ctx = backend_ctx.release();
     return dev_ctx->backend_ctx;
 }
@@ -2098,6 +2502,84 @@ struct ggml_tensor_extra_cl_q4_0 {
     }
 };
 
+struct ggml_tensor_extra_cl_mxfp4 {
+    // Quantized values.
+    cl_mem q = nullptr;
+    // Quantized values in image1d_buffer_t.
+    cl_mem q_img = nullptr;
+    // Scales in E8M0.
+    cl_mem e = nullptr;
+    // Scales in image1d_buffer_t.
+    cl_mem e_img = nullptr;
+    // Size of quantized values.
+    size_t size_q = 0;
+    // Size of scales.
+    size_t size_e = 0;
+
+    ~ggml_tensor_extra_cl_mxfp4() {
+        reset();
+    }
+
+    void reset() {
+        // q and d are subbuffers into the bigger buffer allocated in ggml_backend_buffer.
+        // They must be properly released so that the original buffer can be
+        // properly released to avoid memory leak.
+        if (q != nullptr) {
+            CL_CHECK(clReleaseMemObject(q));
+            q = nullptr;
+        }
+        if (e != nullptr) {
+            CL_CHECK(clReleaseMemObject(e));
+            e = nullptr;
+        }
+        if (q != nullptr) {
+            CL_CHECK(clReleaseMemObject(q_img));
+            q = nullptr;
+        }
+        // Currently, q_img and d_img are not used. They can be image1d_buffer_t
+        // that wraps around q and d to utilize image access path.
+        q_img = nullptr;
+        e_img = nullptr;
+        size_q = 0;
+        size_e = 0;
+    }
+};
+
+struct ggml_tensor_extra_cl_q8_0 {
+    cl_mem q = nullptr;
+    cl_mem q_img = nullptr;
+
+    cl_mem d = nullptr;
+    cl_mem d_img = nullptr;
+
+    size_t size_q = 0;
+    size_t size_d = 0;
+
+    ~ggml_tensor_extra_cl_q8_0() {
+        reset();
+    }
+
+    void reset() {
+        // q and d are subbuffers into the bigger buffer allocated in ggml_backend_buffer.
+        // They must be properly released so that the original buffer can be
+        // properly released to avoid memory leak.
+        if (q != nullptr) {
+            CL_CHECK(clReleaseMemObject(q));
+            q = nullptr;
+        }
+        if (d != nullptr) {
+            CL_CHECK(clReleaseMemObject(d));
+            d = nullptr;
+        }
+        // Currently, q_img and d_img are not used. They can be image1d_buffer_t
+        // that wraps around q and d to utilize image access path.
+        q_img = nullptr;
+        d_img = nullptr;
+        size_q = 0;
+        size_d = 0;
+    }
+};
+
 //------------------------------------------------------------------------------
 // Backend API
 //------------------------------------------------------------------------------
@@ -2178,37 +2660,127 @@ static void sync_with_other_backends(ggml_backend_t backend) {
     sync_with_other_backends(backend_ctx);
 }
 
-static ggml_status ggml_backend_opencl_graph_compute(ggml_backend_t backend, ggml_cgraph * cgraph) {
-    for (int i = 0; i < cgraph->n_nodes; i++) {
-        ggml_tensor * node = cgraph->nodes[i];
+static bool ggml_opencl_can_fuse(const struct ggml_cgraph * cgraph, int node_idx, std::initializer_list<enum ggml_op> ops) {
+    if (!ggml_can_fuse(cgraph, node_idx, ops)) {
+        return false;
+    }
 
-        // NOTE: this may oversynchronize by synchronizing with
-        //       backends/devices which don't compute 'cgraph's
-        //       dependencies.
-        sync_with_other_backends(backend);
+    if (ops.size() == 2 && ops.begin()[0] == GGML_OP_RMS_NORM && ops.begin()[1] == GGML_OP_MUL) {
+        const ggml_tensor *rms_norm = cgraph->nodes[node_idx];
+        const ggml_tensor *mul      = cgraph->nodes[node_idx+1];
 
-        if (node->op == GGML_OP_RESHAPE || node->op == GGML_OP_TRANSPOSE || node->op == GGML_OP_VIEW || node->op == GGML_OP_PERMUTE || node->op == GGML_OP_NONE) {
-            continue;
+        GGML_ASSERT(rms_norm->src[0]->type == GGML_TYPE_F32);
+        GGML_ASSERT(rms_norm->type == GGML_TYPE_F32);
+
+        // rms_norm only supports f32
+        if (mul->src[0]->type != GGML_TYPE_F32 ||
+            mul->src[1]->type != GGML_TYPE_F32 ||
+            mul->type != GGML_TYPE_F32) {
+            return false;
         }
 
-        bool ok = ggml_cl_compute_forward(backend, node);
-        if (!ok) {
-            GGML_LOG_ERROR("%s: error: op not supported %s (%s)\n", __func__, node->name, ggml_op_name(node->op));
+        // if rms_norm is the B operand, then we don't handle broadcast
+        if (rms_norm == mul->src[1] &&
+            !ggml_are_same_shape(mul->src[0], rms_norm->src[1])) {
+            return false;
         }
-        GGML_ASSERT(ok);
-    }
 
-    return GGML_STATUS_SUCCESS;
-}
+        // rms_norm assumes contiguous rows
+        if (!ggml_is_contiguous_rows(mul->src[0]) || !ggml_is_contiguous_rows(mul->src[1])) {
+            return false;
+        }
+    } else if (ops.size() == 3 && ops.begin()[0] == GGML_OP_NORM && ops.begin()[1] == GGML_OP_MUL && ops.begin()[2] == GGML_OP_ADD) {
+        const ggml_tensor *norm = cgraph->nodes[node_idx];
+        const ggml_tensor *mul  = cgraph->nodes[node_idx+1];
+        const ggml_tensor *add  = cgraph->nodes[node_idx+2];
+        const ggml_tensor *w    = mul->src[0] == norm ? mul->src[1] : mul->src[0];
+        const ggml_tensor *b    = add->src[0] == mul  ? add->src[1] : add->src[0];
+
+        // norm fusion only supports F32
+        if (norm->src[0]->type != GGML_TYPE_F32 || w->type != GGML_TYPE_F32 || b->type != GGML_TYPE_F32) {
+            return false;
+        }
 
-static bool ggml_opencl_supports_op(ggml_backend_dev_t dev, const struct ggml_tensor * op) {
-    GGML_UNUSED(dev);
+        if (norm->src[0]->ne[0] % 4 != 0) {
+            return false;
+        }
 
-    switch (op->op) {
-        case GGML_OP_NONE:
-            return true;
-        case GGML_OP_GET_ROWS:
-            switch (op->src[0]->type) {
+        if (!ggml_is_contiguous(norm->src[0]) || !ggml_is_contiguous(w) || !ggml_is_contiguous(b)) {
+            return false;
+        }
+    } else if (ops.size() == 3 && ops.begin()[0] == GGML_OP_GROUP_NORM && ops.begin()[1] == GGML_OP_MUL && ops.begin()[2] == GGML_OP_ADD) {
+        const ggml_tensor *gn = cgraph->nodes[node_idx];
+        const ggml_tensor *mul = cgraph->nodes[node_idx+1];
+        const ggml_tensor *add = cgraph->nodes[node_idx+2];
+        const ggml_tensor *w   = mul->src[0] == gn ? mul->src[1] : mul->src[0];
+        const ggml_tensor *b   = add->src[0] == mul ? add->src[1] : add->src[0];
+
+        if (gn->src[0]->type != GGML_TYPE_F32 || w->type != GGML_TYPE_F32 || b->type != GGML_TYPE_F32) {
+            return false;
+        }
+
+        if (!ggml_is_contiguous(gn->src[0]) || !ggml_is_contiguous(w) || !ggml_is_contiguous(b)) {
+            return false;
+        }
+    }
+
+    return true;
+}
+
+static void ggml_opencl_op_rms_norm_fused(ggml_backend_t backend, ggml_tensor * rms_norm_tensor, ggml_tensor * mul_tensor);
+static void ggml_opencl_op_norm_fused(ggml_backend_t backend, ggml_tensor * norm_tensor, ggml_tensor * mul_tensor, ggml_tensor * add_tensor);
+static void ggml_opencl_op_group_norm_fused(ggml_backend_t backend, ggml_tensor * gn_tensor, ggml_tensor * mul_tensor, ggml_tensor * add_tensor);
+
+static ggml_status ggml_backend_opencl_graph_compute(ggml_backend_t backend, ggml_cgraph * cgraph) {
+    ggml_backend_opencl_context *backend_ctx = (ggml_backend_opencl_context *)backend->context;
+
+    for (int i = 0; i < cgraph->n_nodes; i++) {
+        ggml_tensor * node = cgraph->nodes[i];
+
+        // NOTE: this may oversynchronize by synchronizing with
+        //       backends/devices which don't compute 'cgraph's
+        //       dependencies.
+        sync_with_other_backends(backend);
+
+        if (ggml_is_empty(node) || node->op == GGML_OP_RESHAPE || node->op == GGML_OP_TRANSPOSE || node->op == GGML_OP_VIEW || node->op == GGML_OP_PERMUTE || node->op == GGML_OP_NONE) {
+            continue;
+        }
+
+        if (!backend_ctx->disable_fusion && ggml_opencl_can_fuse(cgraph, i, { GGML_OP_NORM, GGML_OP_MUL, GGML_OP_ADD })) {
+            ggml_opencl_op_norm_fused(backend, node, cgraph->nodes[i+1], cgraph->nodes[i+2]);
+            i += 2;
+            continue;
+        }
+        if (!backend_ctx->disable_fusion && ggml_opencl_can_fuse(cgraph, i, { GGML_OP_GROUP_NORM, GGML_OP_MUL, GGML_OP_ADD })) {
+            ggml_opencl_op_group_norm_fused(backend, node, cgraph->nodes[i+1], cgraph->nodes[i+2]);
+            i += 2;
+            continue;
+        }
+        if (!backend_ctx->disable_fusion && ggml_opencl_can_fuse(cgraph, i, { GGML_OP_RMS_NORM, GGML_OP_MUL })) {
+            ggml_opencl_op_rms_norm_fused(backend, node, cgraph->nodes[i+1]);
+            i++;
+            continue;
+        }
+
+        bool ok = ggml_cl_compute_forward(backend, node);
+        if (!ok) {
+            GGML_LOG_ERROR("%s: error: op not supported %s (%s)\n", __func__, node->name, ggml_op_name(node->op));
+        }
+        GGML_ASSERT(ok);
+    }
+
+    return GGML_STATUS_SUCCESS;
+}
+
+static bool ggml_opencl_supports_op(ggml_backend_dev_t dev, const struct ggml_tensor * op) {
+    ggml_backend_opencl_device_context * dev_ctx     = (ggml_backend_opencl_device_context *)dev->context;
+    ggml_backend_opencl_context *        backend_ctx = dev_ctx->backend_ctx;
+
+    switch (op->op) {
+        case GGML_OP_NONE:
+            return true;
+        case GGML_OP_GET_ROWS:
+            switch (op->src[0]->type) {
                 case GGML_TYPE_F32:
                 case GGML_TYPE_F16:
                     return true;
@@ -2222,6 +2794,22 @@ static bool ggml_opencl_supports_op(ggml_backend_dev_t dev, const struct ggml_te
                 default:
                     return false;
             }
+        case GGML_OP_SET_ROWS:
+            {
+                // TODO: add support
+                // ref: https://github.com/ggml-org/llama.cpp/pull/14274
+#pragma message("TODO: implement BF16, Q4_0, Q4_1, Q5_0, Q5_1, Q8_0, IQ4_NL support (https://github.com/ggml-org/llama.cpp/pull/14661)")
+                if (op->src[0]->type != GGML_TYPE_F32) {
+                    return false;
+                }
+                switch (op->type) {
+                    case GGML_TYPE_F16:
+                    case GGML_TYPE_F32:
+                        return (op->src[1]->type == GGML_TYPE_I64 || op->src[1]->type == GGML_TYPE_I32);
+                    default:
+                        return false;
+                }
+            }
         case GGML_OP_CPY:
         case GGML_OP_DUP:
         case GGML_OP_CONT:
@@ -2245,17 +2833,30 @@ static bool ggml_opencl_supports_op(ggml_backend_dev_t dev, const struct ggml_te
                 default:
                     return false;
             }
-        case GGML_OP_ADD:
         case GGML_OP_SCALE:
+            return op->src[0]->type == GGML_TYPE_F32 && ggml_is_contiguous(op->src[0]);
+        case GGML_OP_ADD:
+            if (op->type == GGML_TYPE_F16) {
+                const bool src0_ok = op->src[0]->type == GGML_TYPE_F16 || op->src[0]->type == GGML_TYPE_F32;
+                const bool src1_ok = op->src[1]->type == GGML_TYPE_F16 || op->src[1]->type == GGML_TYPE_F32;
+                if (src0_ok && src1_ok) {
+                    return true;
+                }
+            }
         case GGML_OP_MUL:
         case GGML_OP_DIV:
         case GGML_OP_SUB:
+            return (op->src[0]->type == op->src[1]->type) &&
+                   (op->src[0]->type == op->type) &&
+                   (op->src[0]->type == GGML_TYPE_F32 || op->src[0]->type == GGML_TYPE_F16);
+        case GGML_OP_ADD_ID:
             return op->src[0]->type == GGML_TYPE_F32;
         case GGML_OP_UNARY:
             switch (ggml_get_unary_op(op)) {
                 case GGML_UNARY_OP_GELU:
                 case GGML_UNARY_OP_SILU:
                 case GGML_UNARY_OP_RELU:
+                case GGML_UNARY_OP_GELU_ERF:
                 case GGML_UNARY_OP_GELU_QUICK:
                    return ggml_is_contiguous(op->src[0]) && op->src[0]->type == GGML_TYPE_F32;
                 case GGML_UNARY_OP_SIGMOID:
@@ -2271,6 +2872,9 @@ static bool ggml_opencl_supports_op(ggml_backend_dev_t dev, const struct ggml_te
                 case GGML_GLU_OP_GEGLU:
                 case GGML_GLU_OP_REGLU:
                 case GGML_GLU_OP_SWIGLU:
+                case GGML_GLU_OP_SWIGLU_OAI:
+                case GGML_GLU_OP_GEGLU_ERF:
+                case GGML_GLU_OP_GEGLU_QUICK:
                     return ggml_is_contiguous_1(op->src[0]) && (op->type == GGML_TYPE_F32 || op->type == GGML_TYPE_F16);
                 default:
                     return false;
@@ -2279,15 +2883,22 @@ static bool ggml_opencl_supports_op(ggml_backend_dev_t dev, const struct ggml_te
             return op->src[0]->type == GGML_TYPE_F32;
         case GGML_OP_SOFT_MAX:
         case GGML_OP_NORM:
-        case GGML_OP_RMS_NORM:
             return true;
+        case GGML_OP_RMS_NORM:
+            return op->ne[0] % 4 == 0 && ggml_is_contiguous_rows(op->src[0]);
         case GGML_OP_REPEAT:
             return op->src[0]->type == GGML_TYPE_F32 && op->type == GGML_TYPE_F32; // Assuming F32 for now, can be expanded
         case GGML_OP_PAD:
             return op->src[0]->type == GGML_TYPE_F32 && op->type == GGML_TYPE_F32 &&
-                   op->src[0]->ne[3] == 1 && op->ne[3] == 1;
+                   op->src[0]->ne[3] == 1 && op->ne[3] == 1 &&
+                   (ggml_get_op_params_i32(op, 0) == 0) && (ggml_get_op_params_i32(op, 2) == 0) &&
+                   (ggml_get_op_params_i32(op, 4) == 0) && (ggml_get_op_params_i32(op, 6) == 0);
         case GGML_OP_UPSCALE:
             return op->src[0]->type == GGML_TYPE_F32 && op->type == GGML_TYPE_F32;
+        case GGML_OP_CONV_2D:
+            return (op->src[0]->type == GGML_TYPE_F16 && op->src[1]->type == GGML_TYPE_F16 && op->type == GGML_TYPE_F16) ||
+                   (op->src[0]->type == GGML_TYPE_F32 && op->src[1]->type == GGML_TYPE_F32 && op->type == GGML_TYPE_F32) ||
+                   (op->src[0]->type == GGML_TYPE_F16 && op->src[1]->type == GGML_TYPE_F32 && op->type == GGML_TYPE_F32);
         case GGML_OP_CONCAT:
             return op->src[0]->type == GGML_TYPE_F32 && op->src[1]->type == GGML_TYPE_F32 && op->type == GGML_TYPE_F32;
         case GGML_OP_TIMESTEP_EMBEDDING:
@@ -2299,13 +2910,17 @@ static bool ggml_opencl_supports_op(ggml_backend_dev_t dev, const struct ggml_te
                 return true;
             } else if (op->src[0]->type == GGML_TYPE_F32) {
                 return op->src[1]->type == GGML_TYPE_F32;
-            } else if (op->src[0]->type == GGML_TYPE_Q4_0 ||
+            } else if (op->src[0]->type == GGML_TYPE_Q4_0 || op->src[0]->type == GGML_TYPE_MXFP4 ||
                        op->src[0]->type == GGML_TYPE_Q6_K) {
                 return op->src[1]->type == GGML_TYPE_F32 && ggml_is_contiguous(op->src[0]) && ggml_is_contiguous(op->src[1]);
+            } else if (op->src[0]->type == GGML_TYPE_Q8_0) {
+                return op->src[1]->type == GGML_TYPE_F32;
             }
             return false;
         case GGML_OP_MUL_MAT_ID:
-            if (op->src[0]->type == GGML_TYPE_Q4_0) {
+            if (op->src[0]->type == GGML_TYPE_Q4_0 ||
+                op->src[0]->type == GGML_TYPE_Q8_0 ||
+                op->src[0]->type == GGML_TYPE_MXFP4) {
                 if (op->src[1]->type == GGML_TYPE_F32) {
                     return ggml_is_contiguous(op->src[0]) && ggml_is_contiguous(op->src[1]);
                 }
@@ -2340,10 +2955,54 @@ static bool ggml_opencl_supports_op(ggml_backend_dev_t dev, const struct ggml_te
         }
         case GGML_OP_IM2COL:
             return true;
-        case GGML_OP_ARGSORT:
-            return op->src[0]->type == GGML_TYPE_F32;
+        case GGML_OP_ARGSORT: {
+            cl_kernel kernel = backend_ctx->kernel_argsort_f32_i32;
+            int max_workgroup_size = backend_ctx->get_kernel_workgroup_size(kernel);
+
+            int cols = 1;
+            while (cols < op->ne[0]) {
+                cols *= 2;
+            }
+
+            return cols <= max_workgroup_size && op->src[0]->type == GGML_TYPE_F32;
+        }
         case GGML_OP_SUM_ROWS:
             return op->src[0]->type == GGML_TYPE_F32 && ggml_is_contiguous(op->src[0]);
+        case GGML_OP_FLASH_ATTN_EXT:
+            {
+                const ggml_tensor * q = op->src[0];
+                const ggml_tensor * k = op->src[1];
+                const ggml_tensor * v = op->src[2];
+
+                const int dk = q->ne[0];
+                const int dv = v->ne[0];
+
+                const struct { int dk; int dv; } supported_dims[] = {
+                    { 40,  40}, { 64,  64}, { 80,  80}, { 96,  96},
+                    {112, 112}, {128, 128}, {192, 128},
+                    {192, 192}, {256, 256},
+                };
+
+                bool dims_supported = false;
+                for (size_t i = 0; i < sizeof(supported_dims)/sizeof(supported_dims[0]); ++i) {
+                    if (supported_dims[i].dk == dk && supported_dims[i].dv == dv) {
+                        dims_supported = true;
+                        break;
+                    }
+                }
+                if (!dims_supported) {
+                    return false;
+                }
+
+                const bool is_f32_f32 = q->type == GGML_TYPE_F32 && k->type == GGML_TYPE_F32 &&
+                                        v->type == GGML_TYPE_F32 && op->type == GGML_TYPE_F32;
+                const bool is_f16_f16 = q->type == GGML_TYPE_F16 && k->type == GGML_TYPE_F16 &&
+                                        v->type == GGML_TYPE_F16 && op->type == GGML_TYPE_F16;
+                const bool is_f32_f16 = q->type == GGML_TYPE_F32 && k->type == GGML_TYPE_F16 &&
+                                        v->type == GGML_TYPE_F16 && op->type == GGML_TYPE_F32;
+
+                return is_f32_f32 || is_f16_f16 || is_f32_f16;
+            }
         default:
             return false;
     }
@@ -2371,6 +3030,7 @@ static ggml_backend_i ggml_backend_opencl_i = {
     /* .graph_compute           = */ ggml_backend_opencl_graph_compute,
     /* .event_record            = */ NULL,
     /* .event_wait              = */ NULL,
+    /* .graph_optimize          = */ NULL,
 };
 
 ggml_backend_t ggml_backend_opencl_init(void) {
@@ -2378,10 +3038,10 @@ ggml_backend_t ggml_backend_opencl_init(void) {
     ggml_backend_opencl_context *backend_ctx = ggml_cl2_init(dev);
 
     ggml_backend_t backend = new ggml_backend {
-        /* .guid      = */ ggml_backend_opencl_guid(),
-        /* .interface = */ ggml_backend_opencl_i,
-        /* .device    = */ dev,
-        /* .context   = */ backend_ctx
+        /* .guid    = */ ggml_backend_opencl_guid(),
+        /* .iface   = */ ggml_backend_opencl_i,
+        /* .device  = */ dev,
+        /* .context = */ backend_ctx
     };
 
     return backend;
@@ -2426,6 +3086,18 @@ struct ggml_backend_opencl_buffer_context {
         for (ggml_tensor_extra_cl_q4_0 * e : temp_tensor_extras_q4_0_in_use) {
             delete e;
         }
+        for (ggml_tensor_extra_cl_mxfp4 * e : temp_tensor_extras_mxfp4) {
+            delete e;
+        }
+        for (ggml_tensor_extra_cl_mxfp4 * e : temp_tensor_extras_mxfp4_in_use) {
+            delete e;
+        }
+        for (ggml_tensor_extra_cl_q8_0 * e : temp_tensor_extras_q8_0) {
+            delete e;
+        }
+        for (ggml_tensor_extra_cl_q8_0 * e : temp_tensor_extras_q8_0_in_use) {
+            delete e;
+        }
     }
 
     ggml_tensor_extra_cl * ggml_opencl_alloc_temp_tensor_extra() {
@@ -2458,6 +3130,36 @@ struct ggml_backend_opencl_buffer_context {
         return extra;
     }
 
+    ggml_tensor_extra_cl_mxfp4 * ggml_opencl_alloc_temp_tensor_extra_mxfp4() {
+        ggml_tensor_extra_cl_mxfp4 * extra;
+        if (temp_tensor_extras_mxfp4.empty()) {
+            extra = new ggml_tensor_extra_cl_mxfp4();
+        } else {
+            extra = temp_tensor_extras_mxfp4.back();
+            temp_tensor_extras_mxfp4.pop_back();
+        }
+
+        temp_tensor_extras_mxfp4_in_use.push_back(extra);
+
+        extra->reset();
+        return extra;
+    }
+
+    ggml_tensor_extra_cl_q8_0 * ggml_opencl_alloc_temp_tensor_extra_q8_0() {
+        ggml_tensor_extra_cl_q8_0 * extra;
+        if (temp_tensor_extras_q8_0.empty()) {
+            extra = new ggml_tensor_extra_cl_q8_0();
+        } else {
+            extra = temp_tensor_extras_q8_0.back();
+            temp_tensor_extras_q8_0.pop_back();
+        }
+
+        temp_tensor_extras_q8_0_in_use.push_back(extra);
+
+        extra->reset();
+        return extra;
+    }
+
     void reset() {
         for (ggml_tensor_extra_cl * e : temp_tensor_extras_in_use) {
             temp_tensor_extras.push_back(e);
@@ -2468,6 +3170,16 @@ struct ggml_backend_opencl_buffer_context {
             temp_tensor_extras_q4_0.push_back(e);
         }
         temp_tensor_extras_q4_0_in_use.clear();
+
+        for (ggml_tensor_extra_cl_mxfp4 * e : temp_tensor_extras_mxfp4_in_use) {
+            temp_tensor_extras_mxfp4.push_back(e);
+        }
+        temp_tensor_extras_mxfp4_in_use.clear();
+
+        for (ggml_tensor_extra_cl_q8_0 * e : temp_tensor_extras_q8_0_in_use) {
+            temp_tensor_extras_q8_0.push_back(e);
+        }
+        temp_tensor_extras_q8_0_in_use.clear();
     }
 
     // Pools for extras. Available extras are in `temp_tensor_extras`. Extras
@@ -2479,6 +3191,10 @@ struct ggml_backend_opencl_buffer_context {
     std::vector<ggml_tensor_extra_cl *> temp_tensor_extras_in_use;
     std::vector<ggml_tensor_extra_cl_q4_0 *> temp_tensor_extras_q4_0;
     std::vector<ggml_tensor_extra_cl_q4_0 *> temp_tensor_extras_q4_0_in_use;
+    std::vector<ggml_tensor_extra_cl_mxfp4 *> temp_tensor_extras_mxfp4;
+    std::vector<ggml_tensor_extra_cl_mxfp4 *> temp_tensor_extras_mxfp4_in_use;
+    std::vector<ggml_tensor_extra_cl_q8_0 *> temp_tensor_extras_q8_0;
+    std::vector<ggml_tensor_extra_cl_q8_0 *> temp_tensor_extras_q8_0_in_use;
 
     // The buffer_context is initially created by ggml_backend_buft_alloc_buffer
     // before any tensor is initialized (at the beginning of alloc_tensor_range).
@@ -2691,7 +3407,10 @@ static void ggml_backend_opencl_buffer_set_tensor(ggml_backend_buffer_t buffer,
         // cl_mem qT_d = clCreateBuffer(context, CL_MEM_READ_WRITE, q_size_bytes, NULL, &err);
         CL_CHECK(err);
 
-        // size_t d_size_bytes = M * (K / 32) / 2 * sizeof(float);
+        bool K_tile_trans = true;
+        if ((K / 32) % 4 != 0){
+            K_tile_trans =false;
+        }
         size_t d_size_bytes = M * (K / 32) * 2;
         region.origin = 0;
         region.size = d_size_bytes;
@@ -2732,10 +3451,15 @@ static void ggml_backend_opencl_buffer_set_tensor(ggml_backend_buffer_t buffer,
         qT_d_image1D = clCreateImage(context, 0, &img_fmt_1d, &img_desc_1d, NULL, &err);
         CL_CHECK(err);
 
-        img_fmt_1d = { CL_RGBA, CL_HALF_FLOAT };
         memset(&img_desc_1d, 0, sizeof(img_desc_1d));
+        if (K_tile_trans) {
+            img_fmt_1d = { CL_RGBA, CL_HALF_FLOAT };
+            img_desc_1d.image_width = M * K / 32 / 4;
+        } else {
+            img_fmt_1d = { CL_R, CL_HALF_FLOAT };
+            img_desc_1d.image_width = M * K / 32;
+        }
         img_desc_1d.image_type = CL_MEM_OBJECT_IMAGE1D_BUFFER;
-        img_desc_1d.image_width = M * K / 32 / 4;
         img_desc_1d.buffer = extra->d;
         d_d_image1D = clCreateImage(context, 0, &img_fmt_1d, &img_desc_1d, NULL, &err);
         CL_CHECK(err);
@@ -2771,6 +3495,10 @@ static void ggml_backend_opencl_buffer_set_tensor(ggml_backend_buffer_t buffer,
         int width_s = K / 32 / 4;
 
         kernel = backend_ctx->kernel_transpose_16;
+        if (!K_tile_trans) {
+            kernel = backend_ctx->kernel_transpose_16_4x1;
+            width_s = K / 32;
+        }
         CL_CHECK(clSetKernelArg(kernel, 0, sizeof(cl_mem), &d_d_image1D));
         CL_CHECK(clSetKernelArg(kernel, 1, sizeof(cl_mem), &dT_d_image1D));
         CL_CHECK(clSetKernelArg(kernel, 2, sizeof(int), &height_s));
@@ -2809,6 +3537,135 @@ static void ggml_backend_opencl_buffer_set_tensor(ggml_backend_buffer_t buffer,
         }
     #endif // GGML_OPENCL_USE_ADRENO_KERNELS
 
+        return;
+
+    }
+    if (tensor->type == GGML_TYPE_MXFP4) {
+        ggml_tensor_extra_cl * extra_orig = (ggml_tensor_extra_cl *)tensor->extra;
+        GGML_ASSERT(extra_orig && "Tesnors in OpenCL backend should have been allocated and initialized");
+
+        // Allocate the new extra and create aliases from the original.
+        ggml_backend_opencl_buffer_context * ctx = (ggml_backend_opencl_buffer_context *) buffer->context;
+        ggml_tensor_extra_cl_mxfp4 * extra = ctx->ggml_opencl_alloc_temp_tensor_extra_mxfp4();
+
+        size_t size_e = ggml_nelements(tensor)/ggml_blck_size(tensor->type)*sizeof(char);
+        size_t size_q = ggml_nelements(tensor)/ggml_blck_size(tensor->type)*ggml_blck_size(tensor->type)/2;
+        GGML_ASSERT(size_e + size_q == ggml_nbytes(tensor) && "Incorrect tensor size");
+
+        cl_int err;
+        cl_mem data_device = clCreateBuffer(context, CL_MEM_READ_WRITE,
+            ggml_nbytes(tensor), NULL, &err);
+        CL_CHECK(err);
+        CL_CHECK(clEnqueueWriteBuffer(
+            queue, data_device, CL_TRUE, 0,
+            ggml_nbytes(tensor), data, 0, NULL, NULL));
+
+        // The original tensor memory is divided into scales and quants, i.e.,
+        // we first store scales, then quants.
+        cl_buffer_region region;
+
+        // Create subbuffer for scales.
+        region.origin = align_to(extra_orig->offset + tensor->view_offs + offset, backend_ctx->alignment);
+        region.size = size_e;
+        extra->e = clCreateSubBuffer(
+            extra_orig->data_device, CL_MEM_READ_WRITE,
+            CL_BUFFER_CREATE_TYPE_REGION, &region, &err);
+        CL_CHECK(err);
+        auto previous_origin = region.origin;
+
+        // Create subbuffer for quants.
+        region.origin = align_to(previous_origin + size_e, backend_ctx->alignment);
+        region.size = size_q;
+        extra->q = clCreateSubBuffer(
+            extra_orig->data_device, CL_MEM_READ_WRITE,
+            CL_BUFFER_CREATE_TYPE_REGION, &region, &err);
+        CL_CHECK(err);
+
+        cl_kernel kernel = backend_ctx->kernel_convert_block_mxfp4;
+
+        CL_CHECK(clSetKernelArg(kernel, 0, sizeof(cl_mem), &data_device));
+        CL_CHECK(clSetKernelArg(kernel, 1, sizeof(cl_mem), &extra->q));
+        CL_CHECK(clSetKernelArg(kernel, 2, sizeof(cl_mem), &extra->e));
+
+        size_t global_work_size[] = {(size_t)ggml_nelements(tensor)/ggml_blck_size(tensor->type), 1, 1};
+        size_t local_work_size[] = {64, 1, 1};
+
+        cl_event evt;
+        CL_CHECK(clEnqueueNDRangeKernel(queue, kernel, 3, NULL, global_work_size, local_work_size, 0, NULL, &evt));
+        CL_CHECK(clWaitForEvents(1, &evt));
+        CL_CHECK(clReleaseMemObject(data_device));
+
+        // Create image for Q
+        cl_image_format img_format_q = {CL_RG, CL_UNSIGNED_INT32};
+        cl_image_desc img_desc_q = {
+            CL_MEM_OBJECT_IMAGE1D_BUFFER,
+            static_cast<size_t>(ggml_nelements(tensor)/32*2),
+            0, 0, 0, 0, 0, 0, 0,
+            { extra->q }
+        };
+        extra->q_img = clCreateImage(context, CL_MEM_READ_ONLY, &img_format_q, &img_desc_q, NULL, &err);
+
+        tensor->extra = extra;
+
+        return;
+    }
+    if (tensor->type == GGML_TYPE_Q8_0) {
+        ggml_tensor_extra_cl * extra_orig = (ggml_tensor_extra_cl *)tensor->extra;
+        GGML_ASSERT(extra_orig && "Tesnors in OpenCL backend should have been allocated and initialized");
+
+        // Allocate the new extra and create aliases from the original.
+        ggml_backend_opencl_buffer_context * ctx = (ggml_backend_opencl_buffer_context *) buffer->context;
+        ggml_tensor_extra_cl_q8_0 * extra = ctx->ggml_opencl_alloc_temp_tensor_extra_q8_0();
+
+        size_t size_d = ggml_nelements(tensor)/ggml_blck_size(tensor->type)*sizeof(ggml_fp16_t);
+        size_t size_q = ggml_nelements(tensor)/ggml_blck_size(tensor->type)*(ggml_blck_size(tensor->type)*sizeof(char));
+        GGML_ASSERT(size_d + size_q == ggml_nbytes(tensor) && "Incorrect tensor size");
+
+        cl_int err;
+        cl_mem data_device = clCreateBuffer(context, CL_MEM_READ_WRITE,
+            ggml_nbytes(tensor), NULL, &err);
+        CL_CHECK(err);
+        CL_CHECK(clEnqueueWriteBuffer(
+            queue, data_device, CL_TRUE, 0,
+            ggml_nbytes(tensor), data, 0, NULL, NULL));
+
+        // The original tensor memory is divided into scales and quants, i.e.,
+        // we first store scales, then quants.
+        cl_buffer_region region;
+
+        // Create subbuffer for scales.
+        region.origin = align_to(extra_orig->offset + tensor->view_offs + offset, backend_ctx->alignment);
+        region.size = size_d;
+        extra->d = clCreateSubBuffer(
+            extra_orig->data_device, CL_MEM_READ_WRITE,
+            CL_BUFFER_CREATE_TYPE_REGION, &region, &err);
+        CL_CHECK(err);
+        auto previous_origin = region.origin;
+
+        // Create subbuffer for quants.
+        region.origin = align_to(previous_origin + size_d, backend_ctx->alignment);
+        region.size = size_q;
+        extra->q = clCreateSubBuffer(
+            extra_orig->data_device, CL_MEM_READ_WRITE,
+            CL_BUFFER_CREATE_TYPE_REGION, &region, &err);
+        CL_CHECK(err);
+
+        cl_kernel kernel = backend_ctx->kernel_convert_block_q8_0;
+
+        CL_CHECK(clSetKernelArg(kernel, 0, sizeof(cl_mem), &data_device));
+        CL_CHECK(clSetKernelArg(kernel, 1, sizeof(cl_mem), &extra->q));
+        CL_CHECK(clSetKernelArg(kernel, 2, sizeof(cl_mem), &extra->d));
+
+        size_t global_work_size[] = {(size_t)ggml_nelements(tensor)/ggml_blck_size(tensor->type), 1, 1};
+        size_t local_work_size[] = {64, 1, 1};
+
+        cl_event evt;
+        CL_CHECK(clEnqueueNDRangeKernel(queue, kernel, 3, NULL, global_work_size, local_work_size, 0, NULL, &evt));
+        CL_CHECK(clWaitForEvents(1, &evt));
+        CL_CHECK(clReleaseMemObject(data_device));
+
+        tensor->extra = extra;
+
         return;
     }
 #endif // GGML_OPENCL_SOA_Q
@@ -2866,25 +3723,76 @@ static void ggml_backend_opencl_buffer_get_tensor(ggml_backend_buffer_t buffer,
             size, data, 0, NULL, NULL));
         CL_CHECK(clReleaseMemObject(data_device));
         return;
-    }
-#endif // GGML_OPENCL_SOA_Q
+    } else if (tensor->type == GGML_TYPE_MXFP4) {
+        ggml_tensor_extra_cl_mxfp4 * extra = (ggml_tensor_extra_cl_mxfp4 *)tensor->extra;
 
-    ggml_tensor_extra_cl * extra = (ggml_tensor_extra_cl *) tensor->extra;
+        cl_int err;
+        cl_mem data_device = clCreateBuffer(context, CL_MEM_READ_WRITE,
+            ggml_nbytes(tensor), NULL, &err);
+        CL_CHECK(err);
 
-    CL_CHECK(clEnqueueReadBuffer(
-        queue, extra->data_device, CL_TRUE, extra->offset + tensor->view_offs + offset,
-        size, data, 0, NULL, NULL));
+        cl_kernel kernel = backend_ctx->kernel_restore_block_mxfp4;
+        CL_CHECK(clSetKernelArg(kernel, 0, sizeof(cl_mem), &extra->q));
+        CL_CHECK(clSetKernelArg(kernel, 1, sizeof(cl_mem), &extra->e));
+        CL_CHECK(clSetKernelArg(kernel, 2, sizeof(cl_mem), &data_device));
 
-    GGML_UNUSED(buffer);
-}
+        size_t global_work_size[] = {(size_t)ggml_nelements(tensor)/ggml_blck_size(tensor->type), 1, 1};
+        size_t local_work_size[] = {1, 1, 1};
 
-static void ggml_backend_opencl_buffer_clear(ggml_backend_buffer_t buffer, uint8_t value) {
-    ggml_backend_dev_t dev = buffer->buft->device;
-    ggml_backend_opencl_context *backend_ctx = ggml_cl2_init(dev);
-    cl_command_queue queue = backend_ctx->queue;
+        cl_event evt;
+        CL_CHECK(clEnqueueNDRangeKernel(queue, kernel, 3, NULL,
+            global_work_size, local_work_size, 0, NULL, &evt));
+        CL_CHECK(clWaitForEvents(1, &evt));
+        CL_CHECK(clEnqueueReadBuffer(
+            queue, data_device, CL_TRUE, offset,
+            size, data, 0, NULL, NULL));
+        CL_CHECK(clReleaseMemObject(data_device));
+        return;
+    }
+    if (tensor->type == GGML_TYPE_Q8_0) {
+        ggml_tensor_extra_cl_q8_0 * extra = (ggml_tensor_extra_cl_q8_0 *)tensor->extra;
 
-    ggml_backend_opencl_buffer_context * ctx = (ggml_backend_opencl_buffer_context *) buffer->context;
-    for (cl_mem buf : ctx->buffer) {
+        cl_int err;
+        cl_mem data_device = clCreateBuffer(context, CL_MEM_READ_WRITE,
+            ggml_nbytes(tensor), NULL, &err);
+        CL_CHECK(err);
+
+        cl_kernel kernel = backend_ctx->kernel_restore_block_q8_0;
+        CL_CHECK(clSetKernelArg(kernel, 0, sizeof(cl_mem), &extra->q));
+        CL_CHECK(clSetKernelArg(kernel, 1, sizeof(cl_mem), &extra->d));
+        CL_CHECK(clSetKernelArg(kernel, 2, sizeof(cl_mem), &data_device));
+
+        size_t global_work_size[] = {(size_t)ggml_nelements(tensor)/ggml_blck_size(tensor->type), 1, 1};
+        size_t local_work_size[] = {1, 1, 1};
+
+        cl_event evt;
+        CL_CHECK(clEnqueueNDRangeKernel(queue, kernel, 3, NULL,
+            global_work_size, local_work_size, 0, NULL, &evt));
+        CL_CHECK(clWaitForEvents(1, &evt));
+        CL_CHECK(clEnqueueReadBuffer(
+            queue, data_device, CL_TRUE, offset,
+            size, data, 0, NULL, NULL));
+        CL_CHECK(clReleaseMemObject(data_device));
+        return;
+    }
+#endif // GGML_OPENCL_SOA_Q
+
+    ggml_tensor_extra_cl * extra = (ggml_tensor_extra_cl *) tensor->extra;
+
+    CL_CHECK(clEnqueueReadBuffer(
+        queue, extra->data_device, CL_TRUE, extra->offset + tensor->view_offs + offset,
+        size, data, 0, NULL, NULL));
+
+    GGML_UNUSED(buffer);
+}
+
+static void ggml_backend_opencl_buffer_clear(ggml_backend_buffer_t buffer, uint8_t value) {
+    ggml_backend_dev_t dev = buffer->buft->device;
+    ggml_backend_opencl_context *backend_ctx = ggml_cl2_init(dev);
+    cl_command_queue queue = backend_ctx->queue;
+
+    ggml_backend_opencl_buffer_context * ctx = (ggml_backend_opencl_buffer_context *) buffer->context;
+    for (cl_mem buf : ctx->buffer) {
         CL_CHECK(clEnqueueFillBuffer(queue, buf, &value, sizeof(value), 0, buffer->size, 0, NULL, NULL));
     }
     CL_CHECK(clFinish(queue));
@@ -3178,6 +4086,19 @@ static void dump_tensor(ggml_backend_t backend, const struct ggml_tensor * tenso
         CL_CHECK(clEnqueueReadBuffer(queue, extra->q, CL_TRUE, 0, size_q, buf_q, 0, NULL, NULL));
         CL_CHECK(clEnqueueReadBuffer(queue, extra->d, CL_TRUE, 0, size_d, buf_d, 0, NULL, NULL));
         CL_CHECK(clFinish(queue));
+    } else if (tensor->type == GGML_TYPE_MXFP4) {
+        ggml_tensor_extra_cl_mxfp4 * extra = (ggml_tensor_extra_cl_mxfp4 *) tensor->extra;
+        GGML_ASSERT(extra);
+
+        size_t size_q = ggml_nelements(tensor)/QK_MXFP4 * QK_MXFP4/2;
+        size_t size_e = ggml_nelements(tensor)/QK_MXFP4 * sizeof(char);
+        GGML_ASSERT(size_q + size_e == ggml_nbytes(tensor));
+        buf_q = malloc(size_q);
+        buf_d = malloc(size_e);
+
+        CL_CHECK(clEnqueueReadBuffer(queue, extra->q, CL_TRUE, 0, size_q, buf_q, 0, NULL, NULL));
+        CL_CHECK(clEnqueueReadBuffer(queue, extra->d, CL_TRUE, 0, size_e, buf_d, 0, NULL, NULL));
+        CL_CHECK(clFinish(queue));
     } else {
         // Read out the tensor from GPU memory.
         ggml_tensor_extra_cl * extra = (ggml_tensor_extra_cl *) tensor->extra;
@@ -3199,7 +4120,7 @@ static void dump_tensor(ggml_backend_t backend, const struct ggml_tensor * tenso
 
     // Open file and dump.
     char fname[512];
-    sprintf(fname, "./tensor-dumps/%s.txt", tensor->name);
+    snprintf(fname, sizeof(fname), "./tensor-dumps/%s.txt", tensor->name);
     FILE * f = fopen(fname, "w");
     if (!f) {
         printf("Failed to open %s\n", fname);
@@ -3358,6 +4279,120 @@ static void ggml_cl_get_rows(ggml_backend_t backend, const ggml_tensor * src0, c
     backend_ctx->enqueue_ndrange_kernel(kernel, 3, global_work_size, local_work_size, dst);
 }
 
+static void ggml_cl_set_rows(ggml_backend_t backend, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
+    GGML_ASSERT(src0);
+    GGML_ASSERT(src0->extra);
+    GGML_ASSERT(src1);
+    GGML_ASSERT(src1->extra);
+    GGML_ASSERT(dst);
+    GGML_ASSERT(dst->extra);
+    GGML_ASSERT(src1->type == GGML_TYPE_I64 || src1->type == GGML_TYPE_I32);
+
+    // ne0 = ne00
+    // ne2 = ne02
+    // ne3 = ne03
+
+    const int      ne01 = src0->ne[1];
+    const int      ne02 = src0->ne[2];
+    const int      ne03 = src0->ne[3];
+
+    const cl_ulong nb01 = src0->nb[1];
+    const cl_ulong nb02 = src0->nb[2];
+    const cl_ulong nb03 = src0->nb[3];
+
+    const int      ne11 = src1->ne[1];
+    const int      ne12 = src1->ne[2];
+
+    const cl_ulong nb10 = src1->nb[0];
+    const cl_ulong nb11 = src1->nb[1];
+    const cl_ulong nb12 = src1->nb[2];
+
+    const int      ne0  = dst->ne[0];
+
+    const cl_ulong nb1  = dst->nb[1];
+    const cl_ulong nb2  = dst->nb[2];
+    const cl_ulong nb3  = dst->nb[3];
+
+    const int nblk0 = ne0/ggml_blck_size(dst->type);
+
+    ggml_backend_opencl_context *backend_ctx = (ggml_backend_opencl_context *)backend->context;
+
+    ggml_tensor_extra_cl * extra0 = (ggml_tensor_extra_cl *)src0->extra;
+    ggml_tensor_extra_cl * extra1 = (ggml_tensor_extra_cl *)src1->extra;
+    ggml_tensor_extra_cl * extrad = (ggml_tensor_extra_cl *)dst->extra;
+
+    cl_ulong offset0 = extra0->offset + src0->view_offs;
+    cl_ulong offset1 = extra1->offset + src1->view_offs;
+    cl_ulong offsetd = extrad->offset + dst->view_offs;
+
+    cl_kernel kernel;
+
+    switch (dst->type) {
+        case GGML_TYPE_F32:
+            if (src1->type == GGML_TYPE_I64) {
+                kernel = backend_ctx->kernel_set_rows_f32_i64;
+            } else {
+                kernel = backend_ctx->kernel_set_rows_f32_i32;
+            }
+            break;
+        case GGML_TYPE_F16:
+            if (src1->type == GGML_TYPE_I64) {
+                kernel = backend_ctx->kernel_set_rows_f16_i64;
+            } else {
+                kernel = backend_ctx->kernel_set_rows_f16_i32;
+            }
+            break;
+        default:
+            GGML_ABORT("not implemented");
+    }
+
+    CL_CHECK(clSetKernelArg(kernel,  0, sizeof(cl_mem),   &extra0->data_device));
+    CL_CHECK(clSetKernelArg(kernel,  1, sizeof(cl_ulong), &offset0));
+    CL_CHECK(clSetKernelArg(kernel,  2, sizeof(cl_mem),   &extra1->data_device));
+    CL_CHECK(clSetKernelArg(kernel,  3, sizeof(cl_ulong), &offset1));
+    CL_CHECK(clSetKernelArg(kernel,  4, sizeof(cl_mem),   &extrad->data_device));
+    CL_CHECK(clSetKernelArg(kernel,  5, sizeof(cl_ulong), &offsetd));
+    CL_CHECK(clSetKernelArg(kernel,  6, sizeof(int),      &ne01));
+    CL_CHECK(clSetKernelArg(kernel,  7, sizeof(cl_ulong), &nb01));
+    CL_CHECK(clSetKernelArg(kernel,  8, sizeof(cl_ulong), &nb02));
+    CL_CHECK(clSetKernelArg(kernel,  9, sizeof(cl_ulong), &nb03));
+    CL_CHECK(clSetKernelArg(kernel, 10, sizeof(int),      &ne11));
+    CL_CHECK(clSetKernelArg(kernel, 11, sizeof(int),      &ne12));
+    CL_CHECK(clSetKernelArg(kernel, 12, sizeof(cl_ulong), &nb10));
+    CL_CHECK(clSetKernelArg(kernel, 13, sizeof(cl_ulong), &nb11));
+    CL_CHECK(clSetKernelArg(kernel, 14, sizeof(cl_ulong), &nb12));
+    CL_CHECK(clSetKernelArg(kernel, 15, sizeof(int),      &nblk0));
+    CL_CHECK(clSetKernelArg(kernel, 16, sizeof(cl_ulong), &nb1));
+    CL_CHECK(clSetKernelArg(kernel, 17, sizeof(cl_ulong), &nb2));
+    CL_CHECK(clSetKernelArg(kernel, 18, sizeof(cl_ulong), &nb3));
+
+    int nth0 = 64;
+    if (backend_ctx->gpu_family == INTEL) {
+        nth0 = 32;
+    } else if (backend_ctx->gpu_family == ADRENO) {
+        nth0 = 64;
+    }
+
+    int max_workgroup_size = backend_ctx->get_kernel_workgroup_size(kernel);
+    while (nth0 < nblk0 && nth0 < max_workgroup_size) {
+        nth0 *= 2;
+    }
+
+    int rows_per_workgroup = 1;
+    if (nth0 > nblk0) {
+        rows_per_workgroup = nth0 / nblk0;
+        nth0 = nblk0;
+    }
+
+    size_t global_work_size[] = {
+        (size_t)(ne01 + rows_per_workgroup - 1)/rows_per_workgroup*nth0,
+        (size_t)ne02*rows_per_workgroup,
+        (size_t)ne03};
+    size_t local_work_size[] = {(size_t)nth0, (size_t)rows_per_workgroup, 1};
+
+    backend_ctx->enqueue_ndrange_kernel(kernel, 3, global_work_size, local_work_size, dst);
+}
+
 static void ggml_cl_add(ggml_backend_t backend, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
     GGML_ASSERT(src0);
     GGML_ASSERT(src0->extra);
@@ -3366,35 +4401,35 @@ static void ggml_cl_add(ggml_backend_t backend, const ggml_tensor * src0, const
     GGML_ASSERT(dst);
     GGML_ASSERT(dst->extra);
 
-    const int  ne00 = src0 ? src0->ne[0] : 0;
-    const int  ne01 = src0 ? src0->ne[1] : 0;
-    const int  ne02 = src0 ? src0->ne[2] : 0;
-    const int  ne03 = src0 ? src0->ne[3] : 0;
+    const int ne00 = src0->ne[0];
+    const int ne01 = src0->ne[1];
+    const int ne02 = src0->ne[2];
+    const int ne03 = src0->ne[3];
 
-    const cl_ulong nb00 = src0 ? src0->nb[0] : 0;
-    const cl_ulong nb01 = src0 ? src0->nb[1] : 0;
-    const cl_ulong nb02 = src0 ? src0->nb[2] : 0;
-    const cl_ulong nb03 = src0 ? src0->nb[3] : 0;
+    const cl_ulong nb00 = src0->nb[0];
+    const cl_ulong nb01 = src0->nb[1];
+    const cl_ulong nb02 = src0->nb[2];
+    const cl_ulong nb03 = src0->nb[3];
 
-    const int  ne10 = src1 ? src1->ne[0] : 0;
-    const int  ne11 = src1 ? src1->ne[1] : 0;
-    const int  ne12 = src1 ? src1->ne[2] : 0;
-    const int  ne13 = src1 ? src1->ne[3] : 0; UNUSED(ne13);
+    const int ne10 = src1->ne[0];
+    const int ne11 = src1->ne[1];
+    const int ne12 = src1->ne[2];
+    const int ne13 = src1->ne[3];
 
-    const cl_ulong nb10 = src1 ? src1->nb[0] : 0;
-    const cl_ulong nb11 = src1 ? src1->nb[1] : 0;
-    const cl_ulong nb12 = src1 ? src1->nb[2] : 0;
-    const cl_ulong nb13 = src1 ? src1->nb[3] : 0; UNUSED(nb13);
+    const cl_ulong nb10 = src1->nb[0];
+    const cl_ulong nb11 = src1->nb[1];
+    const cl_ulong nb12 = src1->nb[2];
+    const cl_ulong nb13 = src1->nb[3];
 
-    const int  ne0  = dst ? dst->ne[0] : 0;
-    const int  ne1  = dst ? dst->ne[1] : 0;
-    const int  ne2  = dst ? dst->ne[2] : 0;
-    const int  ne3  = dst ? dst->ne[3] : 0;
+    const int ne0  = dst->ne[0];
+    const int ne1  = dst->ne[1];
+    const int ne2  = dst->ne[2];
+    const int ne3  = dst->ne[3];
 
-    const cl_ulong nb0  = dst ? dst->nb[0] : 0;
-    const cl_ulong nb1  = dst ? dst->nb[1] : 0;
-    const cl_ulong nb2  = dst ? dst->nb[2] : 0;
-    const cl_ulong nb3  = dst ? dst->nb[3] : 0;
+    const cl_ulong nb0  = dst->nb[0];
+    const cl_ulong nb1  = dst->nb[1];
+    const cl_ulong nb2  = dst->nb[2];
+    const cl_ulong nb3  = dst->nb[3];
 
     ggml_backend_opencl_context *backend_ctx = (ggml_backend_opencl_context *)backend->context;
 
@@ -3406,59 +4441,114 @@ static void ggml_cl_add(ggml_backend_t backend, const ggml_tensor * src0, const
     cl_ulong offset1 = extra1->offset + src1->view_offs;
     cl_ulong offsetd = extrad->offset + dst->view_offs;
 
-    bool bcast_row = false;
     cl_kernel kernel;
 
-    if (ggml_nelements(src1) == ne10 && ggml_is_contiguous(src1) && ne00 % 4 == 0 && ne10 % 4 == 0) {
-        GGML_ASSERT(ggml_is_contiguous(src0));
+    const bool bcast_row = ggml_nelements(src1) == ne10 && ggml_is_contiguous(src1) && ne00 % 4 == 0 && ne10 % 4 == 0;
 
-        // src1 is a row
+    if (bcast_row) {
+        GGML_ASSERT(ggml_is_contiguous(src0));
         GGML_ASSERT(ne11 == 1);
+    }
 
-        bcast_row = true;
-        int ne = ne00 / 4;
-        kernel = backend_ctx->kernel_add_row;
-
-        CL_CHECK(clSetKernelArg(kernel, 0, sizeof(cl_mem),   &extra0->data_device));
-        CL_CHECK(clSetKernelArg(kernel, 1, sizeof(cl_ulong), &offset0));
-        CL_CHECK(clSetKernelArg(kernel, 2, sizeof(cl_mem),   &extra1->data_device));
-        CL_CHECK(clSetKernelArg(kernel, 3, sizeof(cl_ulong), &offset1));
-        CL_CHECK(clSetKernelArg(kernel, 4, sizeof(cl_mem),   &extrad->data_device));
-        CL_CHECK(clSetKernelArg(kernel, 5, sizeof(cl_ulong), &offsetd));
-        CL_CHECK(clSetKernelArg(kernel, 6, sizeof(int),      &ne));
+    if (dst->type == GGML_TYPE_F32) {
+        GGML_ASSERT(src0->type == GGML_TYPE_F32 && src1->type == GGML_TYPE_F32);
+        if (bcast_row) {
+            kernel = backend_ctx->kernel_add_row;
+            const int ne = ne00 / 4;
+            CL_CHECK(clSetKernelArg(kernel, 0, sizeof(cl_mem),   &extra0->data_device));
+            CL_CHECK(clSetKernelArg(kernel, 1, sizeof(cl_ulong), &offset0));
+            CL_CHECK(clSetKernelArg(kernel, 2, sizeof(cl_mem),   &extra1->data_device));
+            CL_CHECK(clSetKernelArg(kernel, 3, sizeof(cl_ulong), &offset1));
+            CL_CHECK(clSetKernelArg(kernel, 4, sizeof(cl_mem),   &extrad->data_device));
+            CL_CHECK(clSetKernelArg(kernel, 5, sizeof(cl_ulong), &offsetd));
+            CL_CHECK(clSetKernelArg(kernel, 6, sizeof(int),      &ne));
+        } else {
+            kernel = backend_ctx->kernel_add;
+            CL_CHECK(clSetKernelArg(kernel,  0, sizeof(cl_mem),   &extra0->data_device));
+            CL_CHECK(clSetKernelArg(kernel,  1, sizeof(cl_ulong), &offset0));
+            CL_CHECK(clSetKernelArg(kernel,  2, sizeof(cl_mem),   &extra1->data_device));
+            CL_CHECK(clSetKernelArg(kernel,  3, sizeof(cl_ulong), &offset1));
+            CL_CHECK(clSetKernelArg(kernel,  4, sizeof(cl_mem),   &extrad->data_device));
+            CL_CHECK(clSetKernelArg(kernel,  5, sizeof(cl_ulong), &offsetd));
+            CL_CHECK(clSetKernelArg(kernel,  6, sizeof(int),      &ne00));
+            CL_CHECK(clSetKernelArg(kernel,  7, sizeof(int),      &ne01));
+            CL_CHECK(clSetKernelArg(kernel,  8, sizeof(int),      &ne02));
+            CL_CHECK(clSetKernelArg(kernel,  9, sizeof(int),      &ne03));
+            CL_CHECK(clSetKernelArg(kernel, 10, sizeof(cl_ulong), &nb00));
+            CL_CHECK(clSetKernelArg(kernel, 11, sizeof(cl_ulong), &nb01));
+            CL_CHECK(clSetKernelArg(kernel, 12, sizeof(cl_ulong), &nb02));
+            CL_CHECK(clSetKernelArg(kernel, 13, sizeof(cl_ulong), &nb03));
+            CL_CHECK(clSetKernelArg(kernel, 14, sizeof(int),      &ne10));
+            CL_CHECK(clSetKernelArg(kernel, 15, sizeof(int),      &ne11));
+            CL_CHECK(clSetKernelArg(kernel, 16, sizeof(int),      &ne12));
+            CL_CHECK(clSetKernelArg(kernel, 17, sizeof(int),      &ne13));
+            CL_CHECK(clSetKernelArg(kernel, 18, sizeof(cl_ulong), &nb10));
+            CL_CHECK(clSetKernelArg(kernel, 19, sizeof(cl_ulong), &nb11));
+            CL_CHECK(clSetKernelArg(kernel, 20, sizeof(cl_ulong), &nb12));
+            CL_CHECK(clSetKernelArg(kernel, 21, sizeof(cl_ulong), &nb13));
+            CL_CHECK(clSetKernelArg(kernel, 22, sizeof(int),      &ne0));
+            CL_CHECK(clSetKernelArg(kernel, 23, sizeof(int),      &ne1));
+            CL_CHECK(clSetKernelArg(kernel, 24, sizeof(int),      &ne2));
+            CL_CHECK(clSetKernelArg(kernel, 25, sizeof(int),      &ne3));
+            CL_CHECK(clSetKernelArg(kernel, 26, sizeof(cl_ulong), &nb0));
+            CL_CHECK(clSetKernelArg(kernel, 27, sizeof(cl_ulong), &nb1));
+            CL_CHECK(clSetKernelArg(kernel, 28, sizeof(cl_ulong), &nb2));
+            CL_CHECK(clSetKernelArg(kernel, 29, sizeof(cl_ulong), &nb3));
+        }
+    } else if (dst->type == GGML_TYPE_F16) {
+        GGML_ASSERT(src0->type == GGML_TYPE_F16 || src0->type == GGML_TYPE_F32);
+        GGML_ASSERT(src1->type == GGML_TYPE_F16 || src1->type == GGML_TYPE_F32);
+        const int type_src0 = (src0->type == GGML_TYPE_F32);
+        const int type_src1 = (src1->type == GGML_TYPE_F32);
+        if (bcast_row) {
+            kernel = backend_ctx->kernel_add_row_f16;
+            const int ne = ne00 / 4;
+            CL_CHECK(clSetKernelArg(kernel, 0, sizeof(cl_mem),   &extra0->data_device));
+            CL_CHECK(clSetKernelArg(kernel, 1, sizeof(cl_ulong), &offset0));
+            CL_CHECK(clSetKernelArg(kernel, 2, sizeof(cl_mem),   &extra1->data_device));
+            CL_CHECK(clSetKernelArg(kernel, 3, sizeof(cl_ulong), &offset1));
+            CL_CHECK(clSetKernelArg(kernel, 4, sizeof(cl_mem),   &extrad->data_device));
+            CL_CHECK(clSetKernelArg(kernel, 5, sizeof(cl_ulong), &offsetd));
+            CL_CHECK(clSetKernelArg(kernel, 6, sizeof(int),      &ne));
+            CL_CHECK(clSetKernelArg(kernel, 7, sizeof(int),      &type_src0));
+            CL_CHECK(clSetKernelArg(kernel, 8, sizeof(int),      &type_src1));
+        } else {
+            kernel = backend_ctx->kernel_add_f16;
+            CL_CHECK(clSetKernelArg(kernel,  0, sizeof(cl_mem),   &extra0->data_device));
+            CL_CHECK(clSetKernelArg(kernel,  1, sizeof(cl_ulong), &offset0));
+            CL_CHECK(clSetKernelArg(kernel,  2, sizeof(cl_mem),   &extra1->data_device));
+            CL_CHECK(clSetKernelArg(kernel,  3, sizeof(cl_ulong), &offset1));
+            CL_CHECK(clSetKernelArg(kernel,  4, sizeof(cl_mem),   &extrad->data_device));
+            CL_CHECK(clSetKernelArg(kernel,  5, sizeof(cl_ulong), &offsetd));
+            CL_CHECK(clSetKernelArg(kernel,  6, sizeof(int),      &ne00));
+            CL_CHECK(clSetKernelArg(kernel,  7, sizeof(int),      &ne01));
+            CL_CHECK(clSetKernelArg(kernel,  8, sizeof(int),      &ne02));
+            CL_CHECK(clSetKernelArg(kernel,  9, sizeof(int),      &ne03));
+            CL_CHECK(clSetKernelArg(kernel, 10, sizeof(cl_ulong), &nb00));
+            CL_CHECK(clSetKernelArg(kernel, 11, sizeof(cl_ulong), &nb01));
+            CL_CHECK(clSetKernelArg(kernel, 12, sizeof(cl_ulong), &nb02));
+            CL_CHECK(clSetKernelArg(kernel, 13, sizeof(cl_ulong), &nb03));
+            CL_CHECK(clSetKernelArg(kernel, 14, sizeof(int),      &ne10));
+            CL_CHECK(clSetKernelArg(kernel, 15, sizeof(int),      &ne11));
+            CL_CHECK(clSetKernelArg(kernel, 16, sizeof(int),      &ne12));
+            CL_CHECK(clSetKernelArg(kernel, 17, sizeof(int),      &ne13));
+            CL_CHECK(clSetKernelArg(kernel, 18, sizeof(cl_ulong), &nb10));
+            CL_CHECK(clSetKernelArg(kernel, 19, sizeof(cl_ulong), &nb11));
+            CL_CHECK(clSetKernelArg(kernel, 20, sizeof(cl_ulong), &nb12));
+            CL_CHECK(clSetKernelArg(kernel, 21, sizeof(cl_ulong), &nb13));
+            CL_CHECK(clSetKernelArg(kernel, 22, sizeof(int),      &ne0));
+            CL_CHECK(clSetKernelArg(kernel, 23, sizeof(int),      &ne1));
+            CL_CHECK(clSetKernelArg(kernel, 24, sizeof(int),      &ne2));
+            CL_CHECK(clSetKernelArg(kernel, 25, sizeof(int),      &ne3));
+            CL_CHECK(clSetKernelArg(kernel, 26, sizeof(cl_ulong), &nb0));
+            CL_CHECK(clSetKernelArg(kernel, 27, sizeof(cl_ulong), &nb1));
+            CL_CHECK(clSetKernelArg(kernel, 28, sizeof(cl_ulong), &nb2));
+            CL_CHECK(clSetKernelArg(kernel, 29, sizeof(cl_ulong), &nb3));
+            CL_CHECK(clSetKernelArg(kernel, 30, sizeof(int),      &type_src0));
+            CL_CHECK(clSetKernelArg(kernel, 31, sizeof(int),      &type_src1));
+        }
     } else {
-        kernel = backend_ctx->kernel_add;
-
-        CL_CHECK(clSetKernelArg(kernel,  0, sizeof(cl_mem),   &extra0->data_device));
-        CL_CHECK(clSetKernelArg(kernel,  1, sizeof(cl_ulong), &offset0));
-        CL_CHECK(clSetKernelArg(kernel,  2, sizeof(cl_mem),   &extra1->data_device));
-        CL_CHECK(clSetKernelArg(kernel,  3, sizeof(cl_ulong), &offset1));
-        CL_CHECK(clSetKernelArg(kernel,  4, sizeof(cl_mem),   &extrad->data_device));
-        CL_CHECK(clSetKernelArg(kernel,  5, sizeof(cl_ulong), &offsetd));
-        CL_CHECK(clSetKernelArg(kernel,  6, sizeof(int),      &ne00));
-        CL_CHECK(clSetKernelArg(kernel,  7, sizeof(int),      &ne01));
-        CL_CHECK(clSetKernelArg(kernel,  8, sizeof(int),      &ne02));
-        CL_CHECK(clSetKernelArg(kernel,  9, sizeof(int),      &ne03));
-        CL_CHECK(clSetKernelArg(kernel, 10, sizeof(cl_ulong), &nb00));
-        CL_CHECK(clSetKernelArg(kernel, 11, sizeof(cl_ulong), &nb01));
-        CL_CHECK(clSetKernelArg(kernel, 12, sizeof(cl_ulong), &nb02));
-        CL_CHECK(clSetKernelArg(kernel, 13, sizeof(cl_ulong), &nb03));
-        CL_CHECK(clSetKernelArg(kernel, 14, sizeof(int),      &ne10));
-        CL_CHECK(clSetKernelArg(kernel, 15, sizeof(int),      &ne11));
-        CL_CHECK(clSetKernelArg(kernel, 16, sizeof(int),      &ne12));
-        CL_CHECK(clSetKernelArg(kernel, 17, sizeof(int),      &ne13));
-        CL_CHECK(clSetKernelArg(kernel, 18, sizeof(cl_ulong), &nb10));
-        CL_CHECK(clSetKernelArg(kernel, 19, sizeof(cl_ulong), &nb11));
-        CL_CHECK(clSetKernelArg(kernel, 20, sizeof(cl_ulong), &nb12));
-        CL_CHECK(clSetKernelArg(kernel, 21, sizeof(cl_ulong), &nb13));
-        CL_CHECK(clSetKernelArg(kernel, 22, sizeof(int),      &ne0));
-        CL_CHECK(clSetKernelArg(kernel, 23, sizeof(int),      &ne1));
-        CL_CHECK(clSetKernelArg(kernel, 24, sizeof(int),      &ne2));
-        CL_CHECK(clSetKernelArg(kernel, 25, sizeof(int),      &ne3));
-        CL_CHECK(clSetKernelArg(kernel, 26, sizeof(cl_ulong), &nb0));
-        CL_CHECK(clSetKernelArg(kernel, 27, sizeof(cl_ulong), &nb1));
-        CL_CHECK(clSetKernelArg(kernel, 28, sizeof(cl_ulong), &nb2));
-        CL_CHECK(clSetKernelArg(kernel, 29, sizeof(cl_ulong), &nb3));
+        GGML_ASSERT(false && "unsupported data types for add");
     }
 
     if (bcast_row) {
@@ -3468,19 +4558,88 @@ static void ggml_cl_add(ggml_backend_t backend, const ggml_tensor * src0, const
 
         size_t * local_work_size_ptr = local_work_size;
         if (n % 64 != 0 && !backend_ctx->non_uniform_workgroups) {
-            local_work_size_ptr = nullptr;  // Let driver choose the work-group sizes.
+            local_work_size_ptr = nullptr;
         }
 
-        backend_ctx->enqueue_ndrange_kernel(kernel, 3, global_work_size, local_work_size_ptr, dst);
+        backend_ctx->enqueue_ndrange_kernel(kernel, 1, global_work_size, local_work_size_ptr, dst);
     } else {
         unsigned int nth = MIN(64, ne0);
-        size_t global_work_size[] = {ne01*nth, (size_t)ne02, (size_t)ne03};
+        size_t global_work_size[] = {(size_t)ne01*nth, (size_t)ne02, (size_t)ne03};
         size_t local_work_size[] = {nth, 1, 1};
 
         backend_ctx->enqueue_ndrange_kernel(kernel, 3, global_work_size, local_work_size, dst);
     }
 }
 
+static void ggml_cl_add_id(ggml_backend_t backend, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
+    GGML_ASSERT(src0);
+    GGML_ASSERT(src0->extra);
+    GGML_ASSERT(src1);
+    GGML_ASSERT(src1->extra);
+    GGML_ASSERT(dst);
+    GGML_ASSERT(dst->extra);
+
+    const ggml_tensor * src2 = dst->src[2];
+    GGML_ASSERT(src2);
+    GGML_ASSERT(src2->extra);
+
+    GGML_ASSERT(src0->type == GGML_TYPE_F32);
+    GGML_ASSERT(src1->type == GGML_TYPE_F32);
+    GGML_ASSERT(src2->type == GGML_TYPE_I32);
+    GGML_ASSERT(dst->type  == GGML_TYPE_F32);
+
+    GGML_ASSERT(ggml_is_contiguous_rows(src0));
+
+    const int ne00 = src0->ne[0];
+    const int ne01 = src0->ne[1];
+    const int ne02 = src0->ne[2];
+
+    const cl_ulong nb01 = src0->nb[1];
+    const cl_ulong nb02 = src0->nb[2];
+
+    const cl_ulong nb11 = src1->nb[1];
+
+    const cl_ulong nb21 = src2->nb[1];
+
+    const int ne0 = dst->ne[0];
+    const int ne1 = dst->ne[1];
+
+    ggml_backend_opencl_context *backend_ctx = (ggml_backend_opencl_context *)backend->context;
+
+    ggml_tensor_extra_cl * extra0 = (ggml_tensor_extra_cl *)src0->extra;
+    ggml_tensor_extra_cl * extra1 = (ggml_tensor_extra_cl *)src1->extra;
+    ggml_tensor_extra_cl * extra2 = (ggml_tensor_extra_cl *)src2->extra;
+    ggml_tensor_extra_cl * extrad = (ggml_tensor_extra_cl *)dst->extra;
+
+    cl_ulong offset0 = extra0->offset + src0->view_offs;
+    cl_ulong offset1 = extra1->offset + src1->view_offs;
+    cl_ulong offset2 = extra2->offset + src2->view_offs;
+    cl_ulong offsetd = extrad->offset + dst->view_offs;
+
+    cl_kernel kernel = backend_ctx->kernel_add_id;
+
+    CL_CHECK(clSetKernelArg(kernel,  0, sizeof(cl_mem),   &extra0->data_device));
+    CL_CHECK(clSetKernelArg(kernel,  1, sizeof(cl_ulong), &offset0));
+    CL_CHECK(clSetKernelArg(kernel,  2, sizeof(cl_mem),   &extra1->data_device));
+    CL_CHECK(clSetKernelArg(kernel,  3, sizeof(cl_ulong), &offset1));
+    CL_CHECK(clSetKernelArg(kernel,  4, sizeof(cl_mem),   &extra2->data_device));
+    CL_CHECK(clSetKernelArg(kernel,  5, sizeof(cl_ulong), &offset2));
+    CL_CHECK(clSetKernelArg(kernel,  6, sizeof(cl_mem),   &extrad->data_device));
+    CL_CHECK(clSetKernelArg(kernel,  7, sizeof(cl_ulong), &offsetd));
+    CL_CHECK(clSetKernelArg(kernel,  8, sizeof(cl_ulong), &nb01));
+    CL_CHECK(clSetKernelArg(kernel,  9, sizeof(cl_ulong), &nb02));
+    CL_CHECK(clSetKernelArg(kernel, 10, sizeof(cl_ulong), &nb11));
+    CL_CHECK(clSetKernelArg(kernel, 11, sizeof(cl_ulong), &nb21));
+    CL_CHECK(clSetKernelArg(kernel, 12, sizeof(int),      &ne0));
+    CL_CHECK(clSetKernelArg(kernel, 13, sizeof(int),      &ne1));
+
+    int nth = MIN(ne00, (int) backend_ctx->get_kernel_workgroup_size(kernel));
+    size_t global_work_size[] = { (size_t)ne01*nth, (size_t)ne02, 1 };
+    size_t local_work_size[] = { (size_t)nth, 1, 1 };
+
+    backend_ctx->enqueue_ndrange_kernel(kernel, 3, global_work_size, local_work_size, dst);
+}
+
 static void ggml_cl_mul(ggml_backend_t backend, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
     GGML_ASSERT(src0);
     GGML_ASSERT(src0->extra);
@@ -3489,35 +4648,39 @@ static void ggml_cl_mul(ggml_backend_t backend, const ggml_tensor * src0, const
     GGML_ASSERT(dst);
     GGML_ASSERT(dst->extra);
 
-    const int ne00 = src0 ? src0->ne[0] : 0;
-    const int ne01 = src0 ? src0->ne[1] : 0;
-    const int ne02 = src0 ? src0->ne[2] : 0;
-    const int ne03 = src0 ? src0->ne[3] : 0;
+    GGML_ASSERT(src0->type == src1->type);
+    GGML_ASSERT(src0->type == dst->type);
+    GGML_ASSERT(src0->type == GGML_TYPE_F32 || src0->type == GGML_TYPE_F16);
 
-    const cl_ulong nb00 = src0 ? src0->nb[0] : 0;
-    const cl_ulong nb01 = src0 ? src0->nb[1] : 0;
-    const cl_ulong nb02 = src0 ? src0->nb[2] : 0;
-    const cl_ulong nb03 = src0 ? src0->nb[3] : 0;
+    const int ne00 = src0->ne[0];
+    const int ne01 = src0->ne[1];
+    const int ne02 = src0->ne[2];
+    const int ne03 = src0->ne[3];
 
-    const int ne10 = src1 ? src1->ne[0] : 0;
-    const int ne11 = src1 ? src1->ne[1] : 0;
-    const int ne12 = src1 ? src1->ne[2] : 0;
-    const int ne13 = src1 ? src1->ne[3] : 0; UNUSED(ne13);
+    const cl_ulong nb00 = src0->nb[0];
+    const cl_ulong nb01 = src0->nb[1];
+    const cl_ulong nb02 = src0->nb[2];
+    const cl_ulong nb03 = src0->nb[3];
 
-    const cl_ulong nb10 = src1 ? src1->nb[0] : 0;
-    const cl_ulong nb11 = src1 ? src1->nb[1] : 0;
-    const cl_ulong nb12 = src1 ? src1->nb[2] : 0;
-    const cl_ulong nb13 = src1 ? src1->nb[3] : 0; UNUSED(nb13);
+    const int ne10 = src1->ne[0];
+    const int ne11 = src1->ne[1];
+    const int ne12 = src1->ne[2];
+    const int ne13 = src1->ne[3]; UNUSED(ne13);
 
-    const int ne0  = dst ? dst->ne[0] : 0;
-    const int ne1  = dst ? dst->ne[1] : 0;
-    const int ne2  = dst ? dst->ne[2] : 0;
-    const int ne3  = dst ? dst->ne[3] : 0;
+    const cl_ulong nb10 = src1->nb[0];
+    const cl_ulong nb11 = src1->nb[1];
+    const cl_ulong nb12 = src1->nb[2];
+    const cl_ulong nb13 = src1->nb[3]; UNUSED(nb13);
+
+    const int ne0  = dst->ne[0];
+    const int ne1  = dst->ne[1];
+    const int ne2  = dst->ne[2];
+    const int ne3  = dst->ne[3];
 
-    const cl_ulong nb0  = dst ? dst->nb[0] : 0;
-    const cl_ulong nb1  = dst ? dst->nb[1] : 0;
-    const cl_ulong nb2  = dst ? dst->nb[2] : 0;
-    const cl_ulong nb3  = dst ? dst->nb[3] : 0;
+    const cl_ulong nb0  = dst->nb[0];
+    const cl_ulong nb1  = dst->nb[1];
+    const cl_ulong nb2  = dst->nb[2];
+    const cl_ulong nb3  = dst->nb[3];
 
     ggml_backend_opencl_context *backend_ctx = (ggml_backend_opencl_context *)backend->context;
 
@@ -3540,7 +4703,12 @@ static void ggml_cl_mul(ggml_backend_t backend, const ggml_tensor * src0, const
 
         bcast_row = true;
         int ne = ne00 / 4;
-        kernel = backend_ctx->kernel_mul_row;
+
+        if (src0->type == GGML_TYPE_F32) {
+            kernel = backend_ctx->kernel_mul_row;
+        } else {
+            kernel = backend_ctx->kernel_mul_row_f16;
+        }
 
         CL_CHECK(clSetKernelArg(kernel, 0, sizeof(cl_mem),   &extra0->data_device));
         CL_CHECK(clSetKernelArg(kernel, 1, sizeof(cl_ulong), &offset0));
@@ -3550,7 +4718,11 @@ static void ggml_cl_mul(ggml_backend_t backend, const ggml_tensor * src0, const
         CL_CHECK(clSetKernelArg(kernel, 5, sizeof(cl_ulong), &offsetd));
         CL_CHECK(clSetKernelArg(kernel, 6, sizeof(int),      &ne));
     } else {
-        kernel = backend_ctx->kernel_mul;
+        if (src0->type == GGML_TYPE_F32) {
+            kernel = backend_ctx->kernel_mul;
+        } else {
+            kernel = backend_ctx->kernel_mul_f16;
+        }
 
         CL_CHECK(clSetKernelArg(kernel,  0, sizeof(cl_mem),   &extra0->data_device));
         CL_CHECK(clSetKernelArg(kernel,  1, sizeof(cl_ulong), &offset0));
@@ -3612,6 +4784,10 @@ static void ggml_cl_div(ggml_backend_t backend, const ggml_tensor * src0, const
     GGML_ASSERT(dst);
     GGML_ASSERT(dst->extra);
 
+    GGML_ASSERT(src0->type == src1->type);
+    GGML_ASSERT(src0->type == dst->type);
+    GGML_ASSERT(src0->type == GGML_TYPE_F32 || src0->type == GGML_TYPE_F16);
+
     const int ne00 = src0->ne[0];
     const int ne01 = src0->ne[1];
     const int ne02 = src0->ne[2];
@@ -3660,7 +4836,12 @@ static void ggml_cl_div(ggml_backend_t backend, const ggml_tensor * src0, const
 
         bcast_row = true;
         int ne = ne00 / 4;
-        kernel = backend_ctx->kernel_div_row;
+
+        if (src0->type == GGML_TYPE_F32) {
+            kernel = backend_ctx->kernel_div_row;
+        } else {
+            kernel = backend_ctx->kernel_div_row_f16;
+        }
 
         CL_CHECK(clSetKernelArg(kernel, 0, sizeof(cl_mem),   &extra0->data_device));
         CL_CHECK(clSetKernelArg(kernel, 1, sizeof(cl_ulong), &offset0));
@@ -3670,7 +4851,11 @@ static void ggml_cl_div(ggml_backend_t backend, const ggml_tensor * src0, const
         CL_CHECK(clSetKernelArg(kernel, 5, sizeof(cl_ulong), &offsetd));
         CL_CHECK(clSetKernelArg(kernel, 6, sizeof(int),      &ne));
     } else {
-        kernel = backend_ctx->kernel_div;
+        if (src0->type == GGML_TYPE_F32) {
+            kernel = backend_ctx->kernel_div;
+        } else {
+            kernel = backend_ctx->kernel_div_f16;
+        }
 
         CL_CHECK(clSetKernelArg(kernel,  0, sizeof(cl_mem),   &extra0->data_device));
         CL_CHECK(clSetKernelArg(kernel,  1, sizeof(cl_ulong), &offset0));
@@ -3720,6 +4905,10 @@ static void ggml_cl_sub(ggml_backend_t backend, const ggml_tensor * src0, const
     GGML_ASSERT(dst);
     GGML_ASSERT(dst->extra);
 
+    GGML_ASSERT(src0->type == src1->type);
+    GGML_ASSERT(src0->type == dst->type);
+    GGML_ASSERT(src0->type == GGML_TYPE_F32 || src0->type == GGML_TYPE_F16);
+
     const int ne00 = src0->ne[0];
     const int ne01 = src0->ne[1];
     const int ne02 = src0->ne[2];
@@ -3768,7 +4957,12 @@ static void ggml_cl_sub(ggml_backend_t backend, const ggml_tensor * src0, const
 
         bcast_row = true;
         int ne = ne00 / 4;
-        kernel = backend_ctx->kernel_sub_row;
+
+        if (src0->type == GGML_TYPE_F32) {
+            kernel = backend_ctx->kernel_sub_row;
+        } else {
+            kernel = backend_ctx->kernel_sub_row_f16;
+        }
 
         CL_CHECK(clSetKernelArg(kernel, 0, sizeof(cl_mem),   &extra0->data_device));
         CL_CHECK(clSetKernelArg(kernel, 1, sizeof(cl_ulong), &offset0));
@@ -3778,7 +4972,11 @@ static void ggml_cl_sub(ggml_backend_t backend, const ggml_tensor * src0, const
         CL_CHECK(clSetKernelArg(kernel, 5, sizeof(cl_ulong), &offsetd));
         CL_CHECK(clSetKernelArg(kernel, 6, sizeof(int),      &ne));
     } else {
-        kernel = backend_ctx->kernel_sub;
+        if (src0->type == GGML_TYPE_F32) {
+            kernel = backend_ctx->kernel_sub;
+        } else {
+            kernel = backend_ctx->kernel_sub_f16;
+        }
 
         CL_CHECK(clSetKernelArg(kernel,  0, sizeof(cl_mem),   &extra0->data_device));
         CL_CHECK(clSetKernelArg(kernel,  1, sizeof(cl_ulong), &offset0));
@@ -3858,6 +5056,44 @@ static void ggml_cl_gelu(ggml_backend_t backend, const ggml_tensor * src0, const
     backend_ctx->enqueue_ndrange_kernel(kernel, 3, global_work_size, local_work_size, dst);
 }
 
+static void ggml_cl_gelu_erf(ggml_backend_t backend, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
+    GGML_ASSERT(src0);
+    GGML_ASSERT(src0->extra);
+    GGML_ASSERT(dst);
+    GGML_ASSERT(dst->extra);
+
+    UNUSED(src1);
+
+    ggml_backend_opencl_context *backend_ctx = (ggml_backend_opencl_context *)backend->context;
+
+    ggml_tensor_extra_cl * extra0 = (ggml_tensor_extra_cl *)src0->extra;
+    ggml_tensor_extra_cl * extrad = (ggml_tensor_extra_cl *)dst->extra;
+
+    cl_ulong offset0 = extra0->offset + src0->view_offs;
+    cl_ulong offsetd = extrad->offset + dst->view_offs;
+
+    cl_kernel kernel;
+
+    int n = ggml_nelements(dst);
+
+    if (n % 4 == 0) {
+        kernel = backend_ctx->kernel_gelu_erf_4;
+        n /= 4;
+    } else {
+        kernel = backend_ctx->kernel_gelu_erf;
+    }
+
+    CL_CHECK(clSetKernelArg(kernel, 0, sizeof(cl_mem),   &extra0->data_device));
+    CL_CHECK(clSetKernelArg(kernel, 1, sizeof(cl_ulong), &offset0));
+    CL_CHECK(clSetKernelArg(kernel, 2, sizeof(cl_mem),   &extrad->data_device));
+    CL_CHECK(clSetKernelArg(kernel, 3, sizeof(cl_ulong), &offsetd));
+
+    size_t global_work_size[] = {(size_t)n, 1, 1};
+    size_t local_work_size[] = {64, 1, 1};
+
+    backend_ctx->enqueue_ndrange_kernel(kernel, 3, global_work_size, local_work_size, dst);
+}
+
 static void ggml_cl_gelu_quick(ggml_backend_t backend, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
     GGML_ASSERT(src0);
     GGML_ASSERT(src0->extra);
@@ -4188,6 +5424,251 @@ static void ggml_cl_rms_norm(ggml_backend_t backend, const ggml_tensor * src0, c
     backend_ctx->enqueue_ndrange_kernel(kernel, 3, global_work_size, local_work_size, dst);
 }
 
+static void ggml_opencl_op_rms_norm_fused(ggml_backend_t backend, ggml_tensor * rms_norm_tensor, ggml_tensor * mul_tensor) {
+    GGML_ASSERT(mul_tensor);
+    GGML_ASSERT(rms_norm_tensor);
+
+    // src0 is the src of rms_norm, src1 is the other src of mul (one being rms_norm)
+    const ggml_tensor * src0 = rms_norm_tensor->src[0];
+    const ggml_tensor * src1;
+    if (mul_tensor->src[0] == rms_norm_tensor) {
+        src1 = mul_tensor->src[1];
+    } else if (mul_tensor->src[1] == rms_norm_tensor) {
+        src1 = mul_tensor->src[0];
+    } else {
+        GGML_ASSERT(false && "Invalid args for rms_norm and mul");
+    }
+    const ggml_tensor * dst = mul_tensor;
+
+    GGML_ASSERT(src0);
+    GGML_ASSERT(src0->extra);
+    GGML_ASSERT(src1);
+    GGML_ASSERT(src1->extra);
+    GGML_ASSERT(dst);
+    GGML_ASSERT(dst->extra);
+
+    ggml_tensor_extra_cl * extra0 = (ggml_tensor_extra_cl *)src0->extra;
+    ggml_tensor_extra_cl * extra1 = (ggml_tensor_extra_cl *)src1->extra;
+    ggml_tensor_extra_cl * extrad = (ggml_tensor_extra_cl *)dst->extra;
+
+    cl_ulong offset0 = extra0->offset + src0->view_offs;
+    cl_ulong offset1 = extra1->offset + src0->view_offs;
+    cl_ulong offsetd = extrad->offset + dst->view_offs;
+
+    ggml_backend_opencl_context *backend_ctx = (ggml_backend_opencl_context *)backend->context;
+
+    float eps;
+    memcpy(&eps, rms_norm_tensor->op_params, sizeof(float));
+
+    const int ne00 = src0->ne[0];
+    const int ne01 = src0->ne[1];
+    const int ne02 = src0->ne[2];
+    const int ne03 = src0->ne[3];
+
+    const cl_ulong nb01 = src0->nb[1];
+    const cl_ulong nb02 = src0->nb[2];
+    const cl_ulong nb03 = src0->nb[3];
+
+    const int ne10 = src1->ne[0];
+    const int ne11 = src1->ne[1];
+    const int ne12 = src1->ne[2];
+    const int ne13 = src1->ne[3];
+
+    const cl_ulong nb11 = src1->nb[1];
+    const cl_ulong nb12 = src1->nb[2];
+    const cl_ulong nb13 = src1->nb[3];
+
+    const cl_ulong nb1 = dst->nb[1];
+    const cl_ulong nb2 = dst->nb[2];
+    const cl_ulong nb3 = dst->nb[3];
+
+    GGML_ASSERT(ne00 % 4 == 0);
+
+    size_t sgs;
+    if (backend_ctx->gpu_family == ADRENO) {
+        sgs = 64;
+    } else if (backend_ctx->gpu_family == INTEL) {
+        sgs = 32;
+    } else {
+        GGML_ASSERT(false && "Unsupported GPU");
+    }
+
+    cl_kernel kernel = backend_ctx->kernel_rms_norm_mul;
+
+    int nth = sgs;
+    int max_workgroup_size = backend_ctx->get_kernel_workgroup_size(kernel);
+    while (nth < ne00 && nth < max_workgroup_size) {
+        nth *= 2;
+    }
+    nth = MIN(nth, max_workgroup_size);
+    nth = MIN(nth, ne00);
+
+    size_t global_work_size[] = {(size_t)ne01*nth, (size_t)ne02, (size_t)ne03};
+    size_t local_work_size[] = {(size_t)nth, 1, 1};
+
+    CL_CHECK(clSetKernelArg(kernel,  0, sizeof(cl_mem),        &extra0->data_device));
+    CL_CHECK(clSetKernelArg(kernel,  1, sizeof(cl_ulong),      &offset0));
+    CL_CHECK(clSetKernelArg(kernel,  2, sizeof(cl_mem),        &extra1->data_device));
+    CL_CHECK(clSetKernelArg(kernel,  3, sizeof(cl_ulong),      &offset1));
+    CL_CHECK(clSetKernelArg(kernel,  4, sizeof(cl_mem),        &extrad->data_device));
+    CL_CHECK(clSetKernelArg(kernel,  5, sizeof(cl_ulong),      &offsetd));
+    CL_CHECK(clSetKernelArg(kernel,  6, sizeof(int),           &ne00));
+    CL_CHECK(clSetKernelArg(kernel,  7, sizeof(int),           &ne01));
+    CL_CHECK(clSetKernelArg(kernel,  8, sizeof(int),           &ne02));
+    CL_CHECK(clSetKernelArg(kernel,  9, sizeof(int),           &ne03));
+    CL_CHECK(clSetKernelArg(kernel, 10, sizeof(cl_ulong),      &nb01));
+    CL_CHECK(clSetKernelArg(kernel, 11, sizeof(cl_ulong),      &nb02));
+    CL_CHECK(clSetKernelArg(kernel, 12, sizeof(cl_ulong),      &nb03));
+    CL_CHECK(clSetKernelArg(kernel, 13, sizeof(int),           &ne10));
+    CL_CHECK(clSetKernelArg(kernel, 14, sizeof(int),           &ne11));
+    CL_CHECK(clSetKernelArg(kernel, 15, sizeof(int),           &ne12));
+    CL_CHECK(clSetKernelArg(kernel, 16, sizeof(int),           &ne13));
+    CL_CHECK(clSetKernelArg(kernel, 17, sizeof(cl_ulong),      &nb11));
+    CL_CHECK(clSetKernelArg(kernel, 18, sizeof(cl_ulong),      &nb12));
+    CL_CHECK(clSetKernelArg(kernel, 19, sizeof(cl_ulong),      &nb13));
+    CL_CHECK(clSetKernelArg(kernel, 20, sizeof(cl_ulong),      &nb1));
+    CL_CHECK(clSetKernelArg(kernel, 21, sizeof(cl_ulong),      &nb2));
+    CL_CHECK(clSetKernelArg(kernel, 22, sizeof(cl_ulong),      &nb3));
+    CL_CHECK(clSetKernelArg(kernel, 23, sizeof(float),         &eps));
+    CL_CHECK(clSetKernelArg(kernel, 24, sizeof(float)*nth/sgs, NULL));
+
+    backend_ctx->enqueue_ndrange_kernel(kernel, 3, global_work_size, local_work_size, dst);
+}
+
+static void ggml_opencl_op_norm_fused(ggml_backend_t backend, ggml_tensor * norm_tensor, ggml_tensor * mul_tensor, ggml_tensor * add_tensor) {
+    GGML_ASSERT(norm_tensor && mul_tensor && add_tensor);
+
+    const ggml_tensor * src0 = norm_tensor->src[0];
+    const ggml_tensor * src1 = mul_tensor->src[0] == norm_tensor ? mul_tensor->src[1] : mul_tensor->src[0];
+    const ggml_tensor * src2 = add_tensor->src[0] == mul_tensor ? add_tensor->src[1] : add_tensor->src[0];
+    const ggml_tensor * dst = add_tensor;
+
+    ggml_tensor_extra_cl * extra0 = (ggml_tensor_extra_cl *)src0->extra;
+    ggml_tensor_extra_cl * extra1 = (ggml_tensor_extra_cl *)src1->extra;
+    ggml_tensor_extra_cl * extra2 = (ggml_tensor_extra_cl *)src2->extra;
+    ggml_tensor_extra_cl * extrad = (ggml_tensor_extra_cl *)dst->extra;
+
+    cl_ulong offset0 = extra0->offset + src0->view_offs;
+    cl_ulong offset1 = extra1->offset + src1->view_offs;
+    cl_ulong offset2 = extra2->offset + src2->view_offs;
+    cl_ulong offsetd = extrad->offset + dst->view_offs;
+
+    ggml_backend_opencl_context *backend_ctx = (ggml_backend_opencl_context *)backend->context;
+
+    float eps;
+    memcpy(&eps, norm_tensor->op_params, sizeof(float));
+
+    const int ne00 = src0->ne[0], ne01 = src0->ne[1], ne02 = src0->ne[2], ne03 = src0->ne[3];
+    const cl_ulong nb01 = src0->nb[1], nb02 = src0->nb[2], nb03 = src0->nb[3];
+    const int ne10 = src1->ne[0], ne11 = src1->ne[1], ne12 = src1->ne[2], ne13 = src1->ne[3];
+    const cl_ulong nb11 = src1->nb[1], nb12 = src1->nb[2], nb13 = src1->nb[3];
+    const int ne20 = src2->ne[0], ne21 = src2->ne[1], ne22 = src2->ne[2], ne23 = src2->ne[3];
+    const cl_ulong nb21 = src2->nb[1], nb22 = src2->nb[2], nb23 = src2->nb[3];
+    const cl_ulong nbd1 = dst->nb[1], nbd2 = dst->nb[2], nbd3 = dst->nb[3];
+
+    size_t sgs;
+    if (backend_ctx->gpu_family == ADRENO) sgs = 64;
+    else if (backend_ctx->gpu_family == INTEL) sgs = 32;
+    else GGML_ASSERT(false && "Unsupported GPU");
+
+    cl_kernel kernel = backend_ctx->kernel_norm_mul_add;
+
+    int nth = sgs;
+    int max_workgroup_size = backend_ctx->get_kernel_workgroup_size(kernel);
+    while (nth < ne00/4 && nth < max_workgroup_size) nth *= 2;
+    nth = MIN(nth, max_workgroup_size);
+    nth = MIN(nth, ne00/4);
+
+    size_t gws[] = {(size_t)ne01*nth, (size_t)ne02, (size_t)ne03};
+    size_t lws[] = {(size_t)nth, 1, 1};
+    size_t num_subgroups = (nth + sgs - 1) / sgs;
+
+    CL_CHECK(clSetKernelArg(kernel, 0, sizeof(cl_mem), &extra0->data_device));
+    CL_CHECK(clSetKernelArg(kernel, 1, sizeof(cl_ulong), &offset0));
+    CL_CHECK(clSetKernelArg(kernel, 2, sizeof(cl_mem), &extra1->data_device));
+    CL_CHECK(clSetKernelArg(kernel, 3, sizeof(cl_ulong), &offset1));
+    CL_CHECK(clSetKernelArg(kernel, 4, sizeof(cl_mem), &extra2->data_device));
+    CL_CHECK(clSetKernelArg(kernel, 5, sizeof(cl_ulong), &offset2));
+    CL_CHECK(clSetKernelArg(kernel, 6, sizeof(cl_mem), &extrad->data_device));
+    CL_CHECK(clSetKernelArg(kernel, 7, sizeof(cl_ulong), &offsetd));
+    CL_CHECK(clSetKernelArg(kernel, 8, sizeof(int), &ne00));
+    CL_CHECK(clSetKernelArg(kernel, 9, sizeof(int), &ne01));
+    CL_CHECK(clSetKernelArg(kernel, 10, sizeof(int), &ne02));
+    CL_CHECK(clSetKernelArg(kernel, 11, sizeof(int), &ne03));
+    CL_CHECK(clSetKernelArg(kernel, 12, sizeof(cl_ulong), &nb01));
+    CL_CHECK(clSetKernelArg(kernel, 13, sizeof(cl_ulong), &nb02));
+    CL_CHECK(clSetKernelArg(kernel, 14, sizeof(cl_ulong), &nb03));
+    CL_CHECK(clSetKernelArg(kernel, 15, sizeof(int), &ne10));
+    CL_CHECK(clSetKernelArg(kernel, 16, sizeof(int), &ne11));
+    CL_CHECK(clSetKernelArg(kernel, 17, sizeof(int), &ne12));
+    CL_CHECK(clSetKernelArg(kernel, 18, sizeof(int), &ne13));
+    CL_CHECK(clSetKernelArg(kernel, 19, sizeof(cl_ulong), &nb11));
+    CL_CHECK(clSetKernelArg(kernel, 20, sizeof(cl_ulong), &nb12));
+    CL_CHECK(clSetKernelArg(kernel, 21, sizeof(cl_ulong), &nb13));
+    CL_CHECK(clSetKernelArg(kernel, 22, sizeof(int), &ne20));
+    CL_CHECK(clSetKernelArg(kernel, 23, sizeof(int), &ne21));
+    CL_CHECK(clSetKernelArg(kernel, 24, sizeof(int), &ne22));
+    CL_CHECK(clSetKernelArg(kernel, 25, sizeof(int), &ne23));
+    CL_CHECK(clSetKernelArg(kernel, 26, sizeof(cl_ulong), &nb21));
+    CL_CHECK(clSetKernelArg(kernel, 27, sizeof(cl_ulong), &nb22));
+    CL_CHECK(clSetKernelArg(kernel, 28, sizeof(cl_ulong), &nb23));
+    CL_CHECK(clSetKernelArg(kernel, 29, sizeof(cl_ulong), &nbd1));
+    CL_CHECK(clSetKernelArg(kernel, 30, sizeof(cl_ulong), &nbd2));
+    CL_CHECK(clSetKernelArg(kernel, 31, sizeof(cl_ulong), &nbd3));
+    CL_CHECK(clSetKernelArg(kernel, 32, sizeof(float), &eps));
+    CL_CHECK(clSetKernelArg(kernel, 33, sizeof(cl_float2) * num_subgroups, NULL));
+
+    backend_ctx->enqueue_ndrange_kernel(kernel, 3, gws, lws, dst);
+}
+
+static void ggml_opencl_op_group_norm_fused(ggml_backend_t backend, ggml_tensor * gn_tensor, ggml_tensor * mul_tensor, ggml_tensor * add_tensor) {
+    GGML_ASSERT(gn_tensor && mul_tensor && add_tensor);
+
+    const ggml_tensor * src0 = gn_tensor->src[0];
+    const ggml_tensor * src1 = mul_tensor->src[0] == gn_tensor ? mul_tensor->src[1] : mul_tensor->src[0];
+    const ggml_tensor * src2 = add_tensor->src[0] == mul_tensor ? add_tensor->src[1] : add_tensor->src[0];
+    const ggml_tensor * dst = add_tensor;
+
+    ggml_tensor_extra_cl * extra0 = (ggml_tensor_extra_cl *)src0->extra;
+    ggml_tensor_extra_cl * extra1 = (ggml_tensor_extra_cl *)src1->extra;
+    ggml_tensor_extra_cl * extra2 = (ggml_tensor_extra_cl *)src2->extra;
+    ggml_tensor_extra_cl * extrad = (ggml_tensor_extra_cl *)dst->extra;
+
+    cl_ulong offset0 = extra0->offset + src0->view_offs;
+    cl_ulong offset1 = extra1->offset + src1->view_offs;
+    cl_ulong offset2 = extra2->offset + src2->view_offs;
+    cl_ulong offsetd = extrad->offset + dst->view_offs;
+
+    ggml_backend_opencl_context *backend_ctx = (ggml_backend_opencl_context *)backend->context;
+
+    int groups;
+    float eps;
+    memcpy(&groups, gn_tensor->op_params, sizeof(int));
+    memcpy(&eps, (char *)gn_tensor->op_params + sizeof(int), sizeof(float));
+
+    cl_kernel kernel = backend_ctx->kernel_group_norm_mul_add;
+    int max_workgroup_size = backend_ctx->get_kernel_workgroup_size(kernel);
+    int ne = ggml_nelements(src0);
+    int group_size = ne / groups;
+
+    size_t lws[] = { (size_t)MIN(max_workgroup_size, group_size) };
+    size_t gws[] = { (size_t)groups * lws[0] };
+
+    CL_CHECK(clSetKernelArg(kernel, 0, sizeof(cl_mem), &extra0->data_device));
+    CL_CHECK(clSetKernelArg(kernel, 1, sizeof(cl_ulong), &offset0));
+    CL_CHECK(clSetKernelArg(kernel, 2, sizeof(cl_mem), &extra1->data_device));
+    CL_CHECK(clSetKernelArg(kernel, 3, sizeof(cl_ulong), &offset1));
+    CL_CHECK(clSetKernelArg(kernel, 4, sizeof(cl_mem), &extra2->data_device));
+    CL_CHECK(clSetKernelArg(kernel, 5, sizeof(cl_ulong), &offset2));
+    CL_CHECK(clSetKernelArg(kernel, 6, sizeof(cl_mem), &extrad->data_device));
+    CL_CHECK(clSetKernelArg(kernel, 7, sizeof(cl_ulong), &offsetd));
+    CL_CHECK(clSetKernelArg(kernel, 8, sizeof(int), &ne));
+    CL_CHECK(clSetKernelArg(kernel, 9, sizeof(int), &group_size));
+    CL_CHECK(clSetKernelArg(kernel, 10, sizeof(float), &eps));
+
+    backend_ctx->enqueue_ndrange_kernel(kernel, 1, gws, lws, dst);
+}
+
 static void ggml_cl_group_norm(ggml_backend_t backend, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
     GGML_ASSERT(src0);
     GGML_ASSERT(src0->extra);
@@ -4453,7 +5934,8 @@ static void ggml_cl_upscale(ggml_backend_t backend, const ggml_tensor * src0, gg
 
     ggml_backend_opencl_context *backend_ctx = (ggml_backend_opencl_context *)backend->context;
 
-    const ggml_scale_mode mode = (ggml_scale_mode) ggml_get_op_params_i32(dst, 0);
+    const int mode_flags        = (ggml_scale_mode) ggml_get_op_params_i32(dst, 0);
+    const ggml_scale_mode mode  = (ggml_scale_mode) (mode_flags & 0xFF);
     cl_kernel kernel = nullptr;
 
     if (mode == GGML_SCALE_MODE_NEAREST) {
@@ -4484,18 +5966,22 @@ static void ggml_cl_upscale(ggml_backend_t backend, const ggml_tensor * src0, gg
     const cl_ulong nb02 = src0->nb[2];
     const cl_ulong nb03 = src0->nb[3];
 
-    const int ne00_src = src0->ne[0];
-    const int ne01_src = src0->ne[1];
+    const int ne00 = src0->ne[0];
+    const int ne01 = src0->ne[1];
+    const int ne02 = src0->ne[2];
+    const int ne03 = src0->ne[3];
 
-    const int ne10_dst = dst->ne[0];
-    const int ne11_dst = dst->ne[1];
-    const int ne12_dst = dst->ne[2];
-    const int ne13_dst = dst->ne[3];
+    const int ne0 = dst->ne[0];
+    const int ne1 = dst->ne[1];
+    const int ne2 = dst->ne[2];
+    const int ne3 = dst->ne[3];
 
-    const float sf0 = (float)dst->ne[0] / src0->ne[0];
-    const float sf1 = (float)dst->ne[1] / src0->ne[1];
-    const float sf2 = (float)dst->ne[2] / src0->ne[2];
-    const float sf3 = (float)dst->ne[3] / src0->ne[3];
+    float sf0 = (float)ne0 / ne00;
+    float sf1 = (float)ne1 / ne01;
+    float sf2 = (float)ne2 / ne02;
+    float sf3 = (float)ne3 / ne03;
+
+    float pixel_offset = 0.5f;
 
     CL_CHECK(clSetKernelArg(kernel, 0, sizeof(cl_mem),    &extra_src0->data_device));
     CL_CHECK(clSetKernelArg(kernel, 1, sizeof(cl_ulong),  &off_src0));
@@ -4507,29 +5993,36 @@ static void ggml_cl_upscale(ggml_backend_t backend, const ggml_tensor * src0, gg
     CL_CHECK(clSetKernelArg(kernel, 7, sizeof(cl_ulong),  &nb03));
 
     if (mode == GGML_SCALE_MODE_NEAREST) {
-        CL_CHECK(clSetKernelArg(kernel, 8, sizeof(int),       &ne10_dst));
-        CL_CHECK(clSetKernelArg(kernel, 9, sizeof(int),       &ne11_dst));
-        CL_CHECK(clSetKernelArg(kernel, 10, sizeof(int),      &ne12_dst));
-        CL_CHECK(clSetKernelArg(kernel, 11, sizeof(int),      &ne13_dst));
+        CL_CHECK(clSetKernelArg(kernel, 8, sizeof(int),       &ne0));
+        CL_CHECK(clSetKernelArg(kernel, 9, sizeof(int),       &ne1));
+        CL_CHECK(clSetKernelArg(kernel, 10, sizeof(int),      &ne2));
+        CL_CHECK(clSetKernelArg(kernel, 11, sizeof(int),      &ne3));
         CL_CHECK(clSetKernelArg(kernel, 12, sizeof(float),    &sf0));
         CL_CHECK(clSetKernelArg(kernel, 13, sizeof(float),    &sf1));
         CL_CHECK(clSetKernelArg(kernel, 14, sizeof(float),    &sf2));
         CL_CHECK(clSetKernelArg(kernel, 15, sizeof(float),    &sf3));
     } else if (mode == GGML_SCALE_MODE_BILINEAR) {
-        CL_CHECK(clSetKernelArg(kernel, 8, sizeof(int),       &ne00_src));
-        CL_CHECK(clSetKernelArg(kernel, 9, sizeof(int),       &ne01_src));
-        CL_CHECK(clSetKernelArg(kernel, 10, sizeof(int),      &ne10_dst));
-        CL_CHECK(clSetKernelArg(kernel, 11, sizeof(int),      &ne11_dst));
-        CL_CHECK(clSetKernelArg(kernel, 12, sizeof(int),      &ne12_dst));
-        CL_CHECK(clSetKernelArg(kernel, 13, sizeof(int),      &ne13_dst));
+        if (mode_flags & GGML_SCALE_FLAG_ALIGN_CORNERS) {
+            sf0 = (float)(ne0 - 1) / (ne00 - 1);
+            sf1 = (float)(ne1 - 1) / (ne01 - 1);
+            pixel_offset = 0.0f;
+        }
+
+        CL_CHECK(clSetKernelArg(kernel, 8, sizeof(int),       &ne00));
+        CL_CHECK(clSetKernelArg(kernel, 9, sizeof(int),       &ne01));
+        CL_CHECK(clSetKernelArg(kernel, 10, sizeof(int),      &ne0));
+        CL_CHECK(clSetKernelArg(kernel, 11, sizeof(int),      &ne1));
+        CL_CHECK(clSetKernelArg(kernel, 12, sizeof(int),      &ne2));
+        CL_CHECK(clSetKernelArg(kernel, 13, sizeof(int),      &ne3));
         CL_CHECK(clSetKernelArg(kernel, 14, sizeof(float),    &sf0));
         CL_CHECK(clSetKernelArg(kernel, 15, sizeof(float),    &sf1));
         CL_CHECK(clSetKernelArg(kernel, 16, sizeof(float),    &sf2));
         CL_CHECK(clSetKernelArg(kernel, 17, sizeof(float),    &sf3));
+        CL_CHECK(clSetKernelArg(kernel, 18, sizeof(float),    &pixel_offset));
     }
 
 
-    size_t dst_total_elements = (size_t)ne10_dst * ne11_dst * ne12_dst * ne13_dst;
+    size_t dst_total_elements = (size_t)ne0 * ne1 * ne2 * ne3;
     if (dst_total_elements == 0) {
         return;
     }
@@ -4626,12 +6119,12 @@ static void ggml_cl_concat(ggml_backend_t backend, const ggml_tensor * src0, con
     } else {
         cl_kernel kernel = backend_ctx->kernel_concat_f32_non_contiguous;
 
-        long ne00 = src0->ne[0], ne01 = src0->ne[1], ne02 = src0->ne[2], ne03 = src0->ne[3];
+        cl_long ne00 = src0->ne[0], ne01 = src0->ne[1], ne02 = src0->ne[2], ne03 = src0->ne[3];
         cl_ulong nb00 = src0->nb[0], nb01 = src0->nb[1], nb02 = src0->nb[2], nb03 = src0->nb[3];
 
         cl_ulong nb10 = src1->nb[0], nb11 = src1->nb[1], nb12 = src1->nb[2], nb13 = src1->nb[3];
 
-        long d_ne0 = dst->ne[0], d_ne1 = dst->ne[1], d_ne2 = dst->ne[2], d_ne3 = dst->ne[3];
+        cl_long d_ne0 = dst->ne[0], d_ne1 = dst->ne[1], d_ne2 = dst->ne[2], d_ne3 = dst->ne[3];
         cl_ulong d_nb0 = dst->nb[0], d_nb1 = dst->nb[1], d_nb2 = dst->nb[2], d_nb3 = dst->nb[3];
 
 
@@ -4642,10 +6135,10 @@ static void ggml_cl_concat(ggml_backend_t backend, const ggml_tensor * src0, con
         CL_CHECK(clSetKernelArg(kernel, 4, sizeof(cl_mem),    &extrad_cl->data_device));
         CL_CHECK(clSetKernelArg(kernel, 5, sizeof(cl_ulong),  &off_dst));
 
-        CL_CHECK(clSetKernelArg(kernel, 6, sizeof(long),      &ne00));
-        CL_CHECK(clSetKernelArg(kernel, 7, sizeof(long),      &ne01));
-        CL_CHECK(clSetKernelArg(kernel, 8, sizeof(long),      &ne02));
-        CL_CHECK(clSetKernelArg(kernel, 9, sizeof(long),      &ne03));
+        CL_CHECK(clSetKernelArg(kernel, 6, sizeof(cl_long),      &ne00));
+        CL_CHECK(clSetKernelArg(kernel, 7, sizeof(cl_long),      &ne01));
+        CL_CHECK(clSetKernelArg(kernel, 8, sizeof(cl_long),      &ne02));
+        CL_CHECK(clSetKernelArg(kernel, 9, sizeof(cl_long),      &ne03));
         CL_CHECK(clSetKernelArg(kernel, 10, sizeof(cl_ulong),    &nb00));
         CL_CHECK(clSetKernelArg(kernel, 11, sizeof(cl_ulong),    &nb01));
         CL_CHECK(clSetKernelArg(kernel, 12, sizeof(cl_ulong),    &nb02));
@@ -4656,10 +6149,10 @@ static void ggml_cl_concat(ggml_backend_t backend, const ggml_tensor * src0, con
         CL_CHECK(clSetKernelArg(kernel, 16, sizeof(cl_ulong),    &nb12));
         CL_CHECK(clSetKernelArg(kernel, 17, sizeof(cl_ulong),    &nb13));
 
-        CL_CHECK(clSetKernelArg(kernel, 18, sizeof(long),     &d_ne0));
-        CL_CHECK(clSetKernelArg(kernel, 19, sizeof(long),     &d_ne1));
-        CL_CHECK(clSetKernelArg(kernel, 20, sizeof(long),     &d_ne2));
-        CL_CHECK(clSetKernelArg(kernel, 21, sizeof(long),     &d_ne3));
+        CL_CHECK(clSetKernelArg(kernel, 18, sizeof(cl_long),     &d_ne0));
+        CL_CHECK(clSetKernelArg(kernel, 19, sizeof(cl_long),     &d_ne1));
+        CL_CHECK(clSetKernelArg(kernel, 20, sizeof(cl_long),     &d_ne2));
+        CL_CHECK(clSetKernelArg(kernel, 21, sizeof(cl_long),     &d_ne3));
         CL_CHECK(clSetKernelArg(kernel, 22, sizeof(cl_ulong),    &d_nb0));
         CL_CHECK(clSetKernelArg(kernel, 23, sizeof(cl_ulong),    &d_nb1));
         CL_CHECK(clSetKernelArg(kernel, 24, sizeof(cl_ulong),    &d_nb2));
@@ -4718,6 +6211,270 @@ static void ggml_cl_timestep_embedding(ggml_backend_t backend, const ggml_tensor
     backend_ctx->enqueue_ndrange_kernel(kernel, 3, global_work_size, NULL, dst);
 }
 
+static void ggml_cl_flash_attn(ggml_backend_t backend, const ggml_tensor * q, const ggml_tensor * k, ggml_tensor * dst) {
+    const ggml_tensor * v = dst->src[2];
+    const ggml_tensor * mask = dst->src[3];
+    const ggml_tensor * sinks = dst->src[4];
+    GGML_ASSERT(q->extra);
+    GGML_ASSERT(k->extra);
+    GGML_ASSERT(v->extra);
+    GGML_ASSERT(dst->extra);
+    if (mask) {
+        GGML_ASSERT(mask->extra);
+    }
+    if (sinks) {
+        GGML_ASSERT(sinks->extra);
+    }
+
+    ggml_backend_opencl_context *backend_ctx = (ggml_backend_opencl_context *)backend->context;
+
+    const int n_q = q->ne[1];
+    const int n_kv = k->ne[1];
+    const int d_head_q = q->ne[0];
+    const int d_head_v = v->ne[0];
+    const int n_head = q->ne[2];
+    const int n_head_kv = k->ne[2];
+    const int n_batch = q->ne[3];
+
+    cl_kernel kernel = NULL;
+
+    const bool is_f16 = q->type == GGML_TYPE_F16;
+    const bool is_mixed = q->type == GGML_TYPE_F32 && k->type == GGML_TYPE_F16;
+    const std::pair<int, int> dk_dv = {d_head_q, d_head_v};
+
+    if (n_q == 1) {
+        if (is_mixed) {
+            kernel = backend_ctx->kernels_flash_attn_f32_f16_q1.at(dk_dv);
+        } else if (is_f16) {
+            kernel = backend_ctx->kernels_flash_attn_f16_q1.at(dk_dv);
+        } else {
+            kernel = backend_ctx->kernels_flash_attn_f32_q1.at(dk_dv);
+        }
+    } else {
+        if (is_mixed) {
+            kernel = backend_ctx->kernels_flash_attn_f32_f16.at(dk_dv);
+        } else if (is_f16) {
+            kernel = backend_ctx->kernels_flash_attn_f16.at(dk_dv);
+        } else {
+            kernel = backend_ctx->kernels_flash_attn_f32.at(dk_dv);
+        }
+    }
+    GGML_ASSERT(kernel != NULL);
+
+    ggml_tensor_extra_cl * extra_q = (ggml_tensor_extra_cl *)q->extra;
+    ggml_tensor_extra_cl * extra_k = (ggml_tensor_extra_cl *)k->extra;
+    ggml_tensor_extra_cl * extra_v = (ggml_tensor_extra_cl *)v->extra;
+    ggml_tensor_extra_cl * extra_o = (ggml_tensor_extra_cl *)dst->extra;
+    ggml_tensor_extra_cl * extra_mask = mask ? (ggml_tensor_extra_cl *)mask->extra : NULL;
+    ggml_tensor_extra_cl * extra_sinks = sinks ? (ggml_tensor_extra_cl *)sinks->extra : NULL;
+
+    cl_ulong offset_q = extra_q->offset + q->view_offs;
+    cl_ulong offset_k = extra_k->offset + k->view_offs;
+    cl_ulong offset_v = extra_v->offset + v->view_offs;
+    cl_ulong offset_o = extra_o->offset + dst->view_offs;
+    cl_mem   mask_buffer = extra_mask ? extra_mask->data_device : NULL;
+    cl_ulong offset_mask = extra_mask ? extra_mask->offset + mask->view_offs : 0;
+    cl_mem   sinks_buffer = extra_sinks ? extra_sinks->data_device : NULL;
+    cl_ulong offset_sinks = extra_sinks ? extra_sinks->offset + sinks->view_offs : 0;
+
+    const cl_ulong q_nb1 = q->nb[1], q_nb2 = q->nb[2], q_nb3 = q->nb[3];
+    const cl_ulong k_nb1 = k->nb[1], k_nb2 = k->nb[2], k_nb3 = k->nb[3];
+    const cl_ulong v_nb1 = v->nb[1], v_nb2 = v->nb[2], v_nb3 = v->nb[3];
+    const cl_ulong o_nb1 = dst->nb[1], o_nb2 = dst->nb[2], o_nb3 = dst->nb[3];
+    const cl_ulong mask_nb1 = mask ? mask->nb[1] : 0;
+    const cl_ulong mask_nb2 = mask ? mask->nb[2] : 0;
+    const cl_ulong mask_nb3 = mask ? mask->nb[3] : 0;
+    const int mask_ne2 = mask ? mask->ne[2] : 0;
+    const int mask_ne3 = mask ? mask->ne[3] : 0;
+
+    float scale, max_bias, logit_softcap;
+    const float * params = (const float *)dst->op_params;
+    scale         = params[0];
+    max_bias      = params[1];
+    logit_softcap = params[2];
+
+    const int is_causal = (mask == NULL && n_q > 1 && n_q == n_kv);
+
+    const int n_head_log2_val = n_head > 0 ? 1u << (int)floorf(log2f((float)n_head)) : 0;
+    const float n_head_log2_f = n_head_log2_val > 0 ? (float)n_head_log2_val : 1.0f;
+    const float m0 = powf(2.0f, -(max_bias) / n_head_log2_f);
+    const float m1 = powf(2.0f, -(max_bias / 2.0f) / n_head_log2_f);
+
+    CL_CHECK(clSetKernelArg(kernel, 0, sizeof(cl_mem),   &extra_q->data_device));
+    CL_CHECK(clSetKernelArg(kernel, 1, sizeof(cl_ulong), &offset_q));
+    CL_CHECK(clSetKernelArg(kernel, 2, sizeof(cl_mem),   &extra_k->data_device));
+    CL_CHECK(clSetKernelArg(kernel, 3, sizeof(cl_ulong), &offset_k));
+    CL_CHECK(clSetKernelArg(kernel, 4, sizeof(cl_mem),   &extra_v->data_device));
+    CL_CHECK(clSetKernelArg(kernel, 5, sizeof(cl_ulong), &offset_v));
+    CL_CHECK(clSetKernelArg(kernel, 6, sizeof(cl_mem),   &extra_o->data_device));
+    CL_CHECK(clSetKernelArg(kernel, 7, sizeof(cl_ulong), &offset_o));
+    CL_CHECK(clSetKernelArg(kernel, 8, sizeof(float),    &scale));
+    CL_CHECK(clSetKernelArg(kernel, 9, sizeof(int),      &n_q));
+    CL_CHECK(clSetKernelArg(kernel, 10, sizeof(int),     &n_kv));
+    CL_CHECK(clSetKernelArg(kernel, 11, sizeof(int),     &is_causal));
+    CL_CHECK(clSetKernelArg(kernel, 12, sizeof(int),     &n_head));
+    CL_CHECK(clSetKernelArg(kernel, 13, sizeof(cl_ulong), &q_nb1)); CL_CHECK(clSetKernelArg(kernel, 14, sizeof(cl_ulong), &q_nb2)); CL_CHECK(clSetKernelArg(kernel, 15, sizeof(cl_ulong), &q_nb3));
+    CL_CHECK(clSetKernelArg(kernel, 16, sizeof(cl_ulong), &k_nb1)); CL_CHECK(clSetKernelArg(kernel, 17, sizeof(cl_ulong), &k_nb2)); CL_CHECK(clSetKernelArg(kernel, 18, sizeof(cl_ulong), &k_nb3));
+    CL_CHECK(clSetKernelArg(kernel, 19, sizeof(cl_ulong), &v_nb1)); CL_CHECK(clSetKernelArg(kernel, 20, sizeof(cl_ulong), &v_nb2)); CL_CHECK(clSetKernelArg(kernel, 21, sizeof(cl_ulong), &v_nb3));
+    CL_CHECK(clSetKernelArg(kernel, 22, sizeof(cl_ulong), &o_nb1)); CL_CHECK(clSetKernelArg(kernel, 23, sizeof(cl_ulong), &o_nb2)); CL_CHECK(clSetKernelArg(kernel, 24, sizeof(cl_ulong), &o_nb3));
+    CL_CHECK(clSetKernelArg(kernel, 25, sizeof(float),    &max_bias));
+    CL_CHECK(clSetKernelArg(kernel, 26, sizeof(float),    &m0));
+    CL_CHECK(clSetKernelArg(kernel, 27, sizeof(float),    &m1));
+    CL_CHECK(clSetKernelArg(kernel, 28, sizeof(int),      &n_head_log2_val));
+    CL_CHECK(clSetKernelArg(kernel, 29, sizeof(float),    &logit_softcap));
+    CL_CHECK(clSetKernelArg(kernel, 30, sizeof(int),      &n_head_kv));
+    CL_CHECK(clSetKernelArg(kernel, 31, sizeof(cl_mem),   &mask_buffer));
+    CL_CHECK(clSetKernelArg(kernel, 32, sizeof(cl_ulong), &offset_mask));
+    CL_CHECK(clSetKernelArg(kernel, 33, sizeof(cl_ulong), &mask_nb1));
+    CL_CHECK(clSetKernelArg(kernel, 34, sizeof(cl_ulong), &mask_nb2));
+    CL_CHECK(clSetKernelArg(kernel, 35, sizeof(cl_ulong), &mask_nb3));
+    CL_CHECK(clSetKernelArg(kernel, 36, sizeof(int),      &mask_ne2));
+    CL_CHECK(clSetKernelArg(kernel, 37, sizeof(int),      &mask_ne3));
+    CL_CHECK(clSetKernelArg(kernel, 38, sizeof(cl_mem),   &sinks_buffer));
+    CL_CHECK(clSetKernelArg(kernel, 39, sizeof(cl_ulong), &offset_sinks));
+
+    if (n_q == 1) {
+        const size_t wg_size = 64;
+        size_t local_work_size[] = { wg_size, 1 };
+        size_t global_work_size[] = { wg_size, (size_t)(n_head * n_batch) };
+        backend_ctx->enqueue_ndrange_kernel(kernel, 2, global_work_size, local_work_size, dst);
+    } else {
+        const int block_m = backend_ctx->kernels_flash_attn_bm.at(dk_dv);
+        const size_t wg_size = block_m;
+        size_t local_work_size[] = { wg_size, 1 };
+        size_t global_work_size[] = { (size_t)((n_q + block_m - 1) / block_m) * wg_size, (size_t)(n_head * n_batch) };
+        backend_ctx->enqueue_ndrange_kernel(kernel, 2, global_work_size, local_work_size, dst);
+    }
+}
+
+static void ggml_cl_mul_mat_f16_f32_tiled(ggml_backend_t backend, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
+    ggml_backend_opencl_context *backend_ctx = (ggml_backend_opencl_context *)backend->context;
+
+    ggml_tensor_extra_cl * extra0 = (ggml_tensor_extra_cl *)src0->extra;
+    ggml_tensor_extra_cl * extra1 = (ggml_tensor_extra_cl *)src1->extra;
+    ggml_tensor_extra_cl * extrad = (ggml_tensor_extra_cl *)dst->extra;
+
+    cl_ulong offset0 = extra0->offset + src0->view_offs;
+    cl_ulong offset1 = extra1->offset + src1->view_offs;
+    cl_ulong offsetd = extrad->offset + dst->view_offs;
+
+    const int M = src0->ne[1];
+    const int N = src1->ne[1];
+    const int K = src0->ne[0];
+
+    cl_kernel kernel = backend_ctx->kernel_mul_mat_f16_f32_tiled;
+
+    CL_CHECK(clSetKernelArg(kernel, 0, sizeof(int),      &M));
+    CL_CHECK(clSetKernelArg(kernel, 1, sizeof(int),      &N));
+    CL_CHECK(clSetKernelArg(kernel, 2, sizeof(int),      &K));
+    CL_CHECK(clSetKernelArg(kernel, 3, sizeof(cl_mem),   &extra0->data_device));
+    CL_CHECK(clSetKernelArg(kernel, 4, sizeof(cl_ulong), &offset0));
+    CL_CHECK(clSetKernelArg(kernel, 5, sizeof(cl_mem),   &extra1->data_device));
+    CL_CHECK(clSetKernelArg(kernel, 6, sizeof(cl_ulong), &offset1));
+    CL_CHECK(clSetKernelArg(kernel, 7, sizeof(cl_mem),   &extrad->data_device));
+    CL_CHECK(clSetKernelArg(kernel, 8, sizeof(cl_ulong), &offsetd));
+
+    // Tiling parameters. These need to be tuned for optimal performance.
+    // They must match the #defines in the kernel mul_mat_f16_f32.cl.
+    //
+    // OPWM / OPWN: Output tile size per Work-Group. A work-group computes a tile of size OPWM x OPWN.
+    // TPWM / TPWN: Threads per Work-group. This is the work-group size.
+    // OPTM / OPTN: Output elements per Thread. Each thread computes OPTM x OPTN elements.
+    //
+    // The following relationships must hold:
+    //   OPWM = TPWM * OPTM
+    //   OPWN = TPWN * OPTN
+    //
+    const int OPWM = 64;
+    const int OPWN = 64;
+    const int TPWM = 16;
+    const int TPWN = 8;
+
+    size_t local_work_size[2] = { TPWM, TPWN };
+    size_t global_work_size[2] = {
+        (size_t) ((M + OPWM - 1) / OPWM) * TPWM,
+        (size_t) ((N + OPWN - 1) / OPWN) * TPWN,
+    };
+
+    backend_ctx->enqueue_ndrange_kernel(kernel, 2, global_work_size, local_work_size, dst);
+}
+
+static void ggml_cl_conv_2d(ggml_backend_t backend, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
+    GGML_TENSOR_BINARY_OP_LOCALS;
+    ggml_backend_opencl_context *backend_ctx = (ggml_backend_opencl_context *)backend->context;
+
+    ggml_tensor_extra_cl * extra0 = (ggml_tensor_extra_cl *)src0->extra;
+    ggml_tensor_extra_cl * extra1 = (ggml_tensor_extra_cl *)src1->extra;
+    ggml_tensor_extra_cl * extrad = (ggml_tensor_extra_cl *)dst->extra;
+
+    cl_ulong offset0 = extra0->offset + src0->view_offs;
+    cl_ulong offset1 = extra1->offset + src1->view_offs;
+    cl_ulong offsetd = extrad->offset + dst->view_offs;
+
+    const cl_uint Cout = ne03; const cl_uint Cin = ne02; const cl_uint N = ne13;
+    const cl_uint KW = ne00; const cl_uint KH = ne01; const cl_uint W = ne10; const cl_uint H = ne11; const cl_uint OW = ne0; const cl_uint OH = ne1;
+
+    const cl_uint s0 = dst->op_params[0]; const cl_uint s1 = dst->op_params[1];
+    const cl_uint p0 = dst->op_params[2]; const cl_uint p1 = dst->op_params[3];
+    const cl_uint d0 = dst->op_params[4]; const cl_uint d1 = dst->op_params[5];
+
+    const cl_uint cl_nb01 = nb01/ggml_type_size(src0->type); const cl_uint cl_nb02 = nb02/ggml_type_size(src0->type); const cl_uint cl_nb03 = nb03/ggml_type_size(src0->type);
+    const cl_uint cl_nb11 = nb11/ggml_type_size(src1->type); const cl_uint cl_nb12 = nb12/ggml_type_size(src1->type); const cl_uint cl_nb13 = nb13/ggml_type_size(src1->type);
+    const cl_uint cl_nb1 = nb1/ggml_type_size(dst->type); const cl_uint cl_nb2 = nb2/ggml_type_size(dst->type); const cl_uint cl_nb3 = nb3/ggml_type_size(dst->type);
+
+    const int64_t NPQ = (int64_t)N * OW * OH;
+
+    const uint32_t BS_K = 64;
+    const uint32_t BS_NPQ = 64;
+    const uint32_t BS_CRS = 16;
+    const uint32_t VEC_SIZE = 4;
+
+    const uint32_t TS_K = 4;
+    const uint32_t TS_NPQ = 8;
+
+    const uint32_t WG_K = BS_K / TS_K;
+    const uint32_t WG_NPQ = BS_NPQ / TS_NPQ;
+
+    auto splitWork = [](uint32_t work_size, uint32_t block_size) { return (block_size + work_size - 1) / block_size; };
+    const uint32_t NB_K = splitWork(Cout, BS_K);
+    const uint32_t NB_NPQ = splitWork(NPQ, BS_NPQ);
+
+    cl_kernel kernel;
+    size_t shmem_size;
+
+    if (src0->type == GGML_TYPE_F16 && src1->type == GGML_TYPE_F16) {
+        kernel = backend_ctx->kernel_conv_2d_f16;
+        shmem_size = (size_t)(BS_K * BS_CRS * sizeof(cl_half) + BS_CRS * (BS_NPQ / VEC_SIZE) * sizeof(cl_half4));
+    } else if (src0->type == GGML_TYPE_F32 && src1->type == GGML_TYPE_F32) {
+        kernel = backend_ctx->kernel_conv_2d_f32;
+        shmem_size = (size_t)(BS_K * BS_CRS * sizeof(cl_float) + BS_CRS * (BS_NPQ / VEC_SIZE) * sizeof(cl_float4));
+    } else if (src0->type == GGML_TYPE_F16 && src1->type == GGML_TYPE_F32) {
+        kernel = backend_ctx->kernel_conv_2d_f16_f32;
+        shmem_size = (size_t)(BS_K * BS_CRS * sizeof(cl_half) + BS_CRS * (BS_NPQ / VEC_SIZE) * sizeof(cl_float4));
+    } else {
+        GGML_ASSERT(false && "Unsupported data type combination for conv2d");
+    }
+
+    cl_uint idx = 0;
+    CL_CHECK(clSetKernelArg(kernel, idx++, sizeof(cl_mem), &extra0->data_device)); CL_CHECK(clSetKernelArg(kernel, idx++, sizeof(cl_ulong), &offset0));
+    CL_CHECK(clSetKernelArg(kernel, idx++, sizeof(cl_mem), &extra1->data_device)); CL_CHECK(clSetKernelArg(kernel, idx++, sizeof(cl_ulong), &offset1));
+    CL_CHECK(clSetKernelArg(kernel, idx++, sizeof(cl_mem), &extrad->data_device)); CL_CHECK(clSetKernelArg(kernel, idx++, sizeof(cl_ulong), &offsetd));
+    CL_CHECK(clSetKernelArg(kernel, idx++, shmem_size, NULL));
+    CL_CHECK(clSetKernelArg(kernel, idx++, sizeof(cl_uint), &Cout)); CL_CHECK(clSetKernelArg(kernel, idx++, sizeof(cl_uint), &Cin)); CL_CHECK(clSetKernelArg(kernel, idx++, sizeof(cl_uint), &N));
+    CL_CHECK(clSetKernelArg(kernel, idx++, sizeof(cl_uint), &KW)); CL_CHECK(clSetKernelArg(kernel, idx++, sizeof(cl_uint), &KH)); CL_CHECK(clSetKernelArg(kernel, idx++, sizeof(cl_uint), &W)); CL_CHECK(clSetKernelArg(kernel, idx++, sizeof(cl_uint), &H));
+    CL_CHECK(clSetKernelArg(kernel, idx++, sizeof(cl_uint), &OW)); CL_CHECK(clSetKernelArg(kernel, idx++, sizeof(cl_uint), &OH));
+    CL_CHECK(clSetKernelArg(kernel, idx++, sizeof(cl_uint), &s0)); CL_CHECK(clSetKernelArg(kernel, idx++, sizeof(cl_uint), &s1)); CL_CHECK(clSetKernelArg(kernel, idx++, sizeof(cl_uint), &p0)); CL_CHECK(clSetKernelArg(kernel, idx++, sizeof(cl_uint), &p1));
+    CL_CHECK(clSetKernelArg(kernel, idx++, sizeof(cl_uint), &d0)); CL_CHECK(clSetKernelArg(kernel, idx++, sizeof(cl_uint), &d1));
+    CL_CHECK(clSetKernelArg(kernel, idx++, sizeof(cl_uint), &cl_nb01)); CL_CHECK(clSetKernelArg(kernel, idx++, sizeof(cl_uint), &cl_nb02)); CL_CHECK(clSetKernelArg(kernel, idx++, sizeof(cl_uint), &cl_nb03));
+    CL_CHECK(clSetKernelArg(kernel, idx++, sizeof(cl_uint), &cl_nb11)); CL_CHECK(clSetKernelArg(kernel, idx++, sizeof(cl_uint), &cl_nb12)); CL_CHECK(clSetKernelArg(kernel, idx++, sizeof(cl_uint), &cl_nb13));
+    CL_CHECK(clSetKernelArg(kernel, idx++, sizeof(cl_uint), &cl_nb1)); CL_CHECK(clSetKernelArg(kernel, idx++, sizeof(cl_uint), &cl_nb2)); CL_CHECK(clSetKernelArg(kernel, idx++, sizeof(cl_uint), &cl_nb3));
+
+    size_t global_work_size[] = { (size_t)NB_K * WG_K, (size_t)NB_NPQ * WG_NPQ, 1 };
+    size_t local_work_size[] = { (size_t)WG_K, (size_t)WG_NPQ, 1 };
+
+    backend_ctx->enqueue_ndrange_kernel(kernel, 2, global_work_size, local_work_size, dst);
+}
+
 static void ggml_cl_mul_mat(ggml_backend_t backend, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
     GGML_ASSERT(src0);
     GGML_ASSERT(src0->extra);
@@ -4741,6 +6498,8 @@ static void ggml_cl_mul_mat(ggml_backend_t backend, const ggml_tensor * src0, co
 
 #ifdef GGML_OPENCL_SOA_Q
     ggml_tensor_extra_cl_q4_0 * extra0_q4_0 = (ggml_tensor_extra_cl_q4_0 *)src0->extra;
+    ggml_tensor_extra_cl_mxfp4 * extra0_mxfp4 = (ggml_tensor_extra_cl_mxfp4 *)src0->extra;
+    ggml_tensor_extra_cl_q8_0 * extra0_q8_0 = (ggml_tensor_extra_cl_q8_0 *)src0->extra;
 #endif
 
     const int  ne00 = src0 ? src0->ne[0] : 0;
@@ -5070,12 +6829,107 @@ static void ggml_cl_mul_mat(ggml_backend_t backend, const ggml_tensor * src0, co
             CL_CHECK(clReleaseMemObject(B_d_input_image));
             CL_CHECK(clReleaseMemObject(C_d));
         }
-        // <--------------------------------------------> //
+        // <--------------------------------------------> //
+
+        return;
+    }
+    } // if (ne01 && ne1)
+#endif // GGML_OPENCL_USE_ADRENO_KERNELS
+
+    // GEMM using local memory
+    // Current BK = 16, so ne00 % 16 == 0
+    if (ggml_is_contiguous(src0) &&
+        ggml_is_contiguous(src1) &&
+        src1t == GGML_TYPE_F32 &&
+        ne00 % 16 == 0 &&
+        ne11 > 1) {
+        switch(src0t) {
+            case GGML_TYPE_F32: {
+                kernel = backend_ctx->kernel_mul_mm_f32_f32_l4_lm;
+                nth0 = 128; // calculated as (BM*BN)/(TM*TN)
+
+                int batch_stride_a = ne00*ne01;
+                int batch_stride_b = ne10*ne11;
+                int batch_stride_d = ne0*ne1;
+
+                CL_CHECK(clSetKernelArg(kernel,  0, sizeof(cl_mem),   &extra0->data_device));
+                CL_CHECK(clSetKernelArg(kernel,  1, sizeof(cl_ulong), &offset0));
+                CL_CHECK(clSetKernelArg(kernel,  2, sizeof(cl_mem),   &extra1->data_device));
+                CL_CHECK(clSetKernelArg(kernel,  3, sizeof(cl_ulong), &offset1));
+                CL_CHECK(clSetKernelArg(kernel,  4, sizeof(cl_mem),   &extrad->data_device));
+                CL_CHECK(clSetKernelArg(kernel,  5, sizeof(cl_ulong), &offsetd));
+                CL_CHECK(clSetKernelArg(kernel,  6, sizeof(int),      &ne00));
+                CL_CHECK(clSetKernelArg(kernel,  7, sizeof(int),      &ne01));
+                CL_CHECK(clSetKernelArg(kernel,  8, sizeof(int),      &ne02));
+                CL_CHECK(clSetKernelArg(kernel,  9, sizeof(int),      &ne11));
+                CL_CHECK(clSetKernelArg(kernel, 10, sizeof(int),      &ne12));
+                CL_CHECK(clSetKernelArg(kernel, 11, sizeof(int),      &ne10)); // stride_a
+                CL_CHECK(clSetKernelArg(kernel, 12, sizeof(int),      &ne10)); // stride_b
+                CL_CHECK(clSetKernelArg(kernel, 13, sizeof(int),      &ne01)); // stride_d
+                CL_CHECK(clSetKernelArg(kernel, 14, sizeof(int),      &batch_stride_a));
+                CL_CHECK(clSetKernelArg(kernel, 15, sizeof(int),      &batch_stride_b));
+                CL_CHECK(clSetKernelArg(kernel, 16, sizeof(int),      &batch_stride_d));
+                CL_CHECK(clSetKernelArg(kernel, 17, sizeof(int),      &r2));
+                CL_CHECK(clSetKernelArg(kernel, 18, sizeof(int),      &r3));
+
+                // 64 is block tile size BM and BN - change here when BM and BN in the kernel are changed.
+                size_t global_work_size[] = {(size_t)(CEIL_DIV(ne01, 64)*nth0), (size_t)(CEIL_DIV(ne11, 64)), (size_t)ne12*ne13};
+                size_t local_work_size[] = {(size_t)nth0, 1, 1};
+
+                backend_ctx->enqueue_ndrange_kernel(kernel, 3, global_work_size, local_work_size, dst);
+                return;
+            }
+            case GGML_TYPE_F16: {
+                kernel = backend_ctx->kernel_mul_mm_f16_f32_l4_lm;
+                nth0 = 128; // calculated as (BM*BN)/(TM*TN)
+
+                int batch_stride_a = ne00*ne01;
+                int batch_stride_b = ne10*ne11;
+                int batch_stride_d = ne0*ne1;
+
+                CL_CHECK(clSetKernelArg(kernel,  0, sizeof(cl_mem),   &extra0->data_device));
+                CL_CHECK(clSetKernelArg(kernel,  1, sizeof(cl_ulong), &offset0));
+                CL_CHECK(clSetKernelArg(kernel,  2, sizeof(cl_mem),   &extra1->data_device));
+                CL_CHECK(clSetKernelArg(kernel,  3, sizeof(cl_ulong), &offset1));
+                CL_CHECK(clSetKernelArg(kernel,  4, sizeof(cl_mem),   &extrad->data_device));
+                CL_CHECK(clSetKernelArg(kernel,  5, sizeof(cl_ulong), &offsetd));
+                CL_CHECK(clSetKernelArg(kernel,  6, sizeof(int),      &ne00));
+                CL_CHECK(clSetKernelArg(kernel,  7, sizeof(int),      &ne01));
+                CL_CHECK(clSetKernelArg(kernel,  8, sizeof(int),      &ne02));
+                CL_CHECK(clSetKernelArg(kernel,  9, sizeof(int),      &ne11));
+                CL_CHECK(clSetKernelArg(kernel, 10, sizeof(int),      &ne12));
+                CL_CHECK(clSetKernelArg(kernel, 11, sizeof(int),      &ne10)); // stride_a
+                CL_CHECK(clSetKernelArg(kernel, 12, sizeof(int),      &ne10)); // stride_b
+                CL_CHECK(clSetKernelArg(kernel, 13, sizeof(int),      &ne01)); // stride_d
+                CL_CHECK(clSetKernelArg(kernel, 14, sizeof(int),      &batch_stride_a));
+                CL_CHECK(clSetKernelArg(kernel, 15, sizeof(int),      &batch_stride_b));
+                CL_CHECK(clSetKernelArg(kernel, 16, sizeof(int),      &batch_stride_d));
+                CL_CHECK(clSetKernelArg(kernel, 17, sizeof(int),      &r2));
+                CL_CHECK(clSetKernelArg(kernel, 18, sizeof(int),      &r3));
+
+                // 64 is block tile size BM and BN - change here when BM and BN in the kernel are changed.
+                size_t global_work_size[] = {(size_t)(CEIL_DIV(ne01, 64)*nth0), (size_t)(CEIL_DIV(ne11, 64)), (size_t)ne12*ne13};
+                size_t local_work_size[] = {(size_t)nth0, 1, 1};
+
+                backend_ctx->enqueue_ndrange_kernel(kernel, 3, global_work_size, local_work_size, dst);
+                return;
+            }
+            default:
+                break;
+        }
+    }
 
+    if (src0t == GGML_TYPE_F16 && src1t == GGML_TYPE_F32 &&
+        src0->ne[1] > 32 &&   // M > 32
+        src1->ne[1] > 32 &&   // N > 32
+        src0->ne[0] > 32 &&   // K > 32
+        src0->ne[2] == 1 && src0->ne[3] == 1 &&
+        src1->ne[2] == 1 && src1->ne[3] == 1 &&
+        ggml_is_contiguous(src0) && ggml_is_contiguous(src1) &&
+        backend_ctx->kernel_mul_mat_f16_f32_tiled != NULL) {
+        ggml_cl_mul_mat_f16_f32_tiled(backend, src0, src1, dst);
         return;
     }
-    } // if (ne01 && ne1)
-#endif // GGML_OPENCL_USE_ADRENO_KERNELS
 
     if (!ggml_is_transposed(src0) &&
         !ggml_is_transposed(src1) &&
@@ -5315,7 +7169,84 @@ static void ggml_cl_mul_mat(ggml_backend_t backend, const ggml_tensor * src0, co
 #endif // GGML_OPENCL_SOA_Q
             break;
         case GGML_TYPE_Q4_1:
-        case GGML_TYPE_Q8_0:
+        case GGML_TYPE_Q8_0: {
+#ifdef GGML_OPENCL_SOA_Q
+            kernel = backend_ctx->kernel_mul_mv_q8_0_f32_flat;
+
+            // nth0 - subgroup size
+            // nth1 - number of subgroups per workgroup
+            // ndst - number of output values per workgroup = output per subgroup * number of subgroups
+            if (backend_ctx->gpu_family == INTEL) {
+                nth0 = 16;
+                nth1 = 2;
+                ndst = nth1*4;
+            } else if (backend_ctx->gpu_family == ADRENO) {
+                nth0 = 64;
+                nth1 = 2;
+                ndst = nth1*4;
+            } else {
+                GGML_ASSERT(false && "TODO: Unknown GPU");
+            }
+
+            CL_CHECK(clSetKernelArg(kernel,  0, sizeof(cl_mem),   &extra0_q8_0->q));
+            CL_CHECK(clSetKernelArg(kernel,  1, sizeof(cl_mem),   &extra0_q8_0->d));
+            CL_CHECK(clSetKernelArg(kernel,  2, sizeof(cl_mem),   &extra1->data_device));
+            CL_CHECK(clSetKernelArg(kernel,  3, sizeof(cl_ulong), &offset1));
+            CL_CHECK(clSetKernelArg(kernel,  4, sizeof(cl_mem),   &extrad->data_device));
+            CL_CHECK(clSetKernelArg(kernel,  5, sizeof(cl_ulong), &offsetd));
+            CL_CHECK(clSetKernelArg(kernel,  6, sizeof(int),      &ne00));
+            CL_CHECK(clSetKernelArg(kernel,  7, sizeof(int),      &ne01));
+            CL_CHECK(clSetKernelArg(kernel,  8, sizeof(cl_ulong), &nb01));
+            CL_CHECK(clSetKernelArg(kernel,  9, sizeof(cl_ulong), &nb02));
+            CL_CHECK(clSetKernelArg(kernel, 10, sizeof(cl_ulong), &nb03));
+            CL_CHECK(clSetKernelArg(kernel, 11, sizeof(int),      &ne12));
+            CL_CHECK(clSetKernelArg(kernel, 12, sizeof(cl_ulong), &nb11));
+            CL_CHECK(clSetKernelArg(kernel, 13, sizeof(cl_ulong), &nb12));
+            CL_CHECK(clSetKernelArg(kernel, 14, sizeof(cl_ulong), &nb13));
+            CL_CHECK(clSetKernelArg(kernel, 15, sizeof(int),      &ne0));
+            CL_CHECK(clSetKernelArg(kernel, 16, sizeof(int),      &ne1));
+            CL_CHECK(clSetKernelArg(kernel, 17, sizeof(int),      &r2));
+            CL_CHECK(clSetKernelArg(kernel, 18, sizeof(int),      &r3));
+#else
+            kernel = backend_ctx->kernel_mul_mv_q8_0_f32;
+
+            // nth0 - subgroup size
+            // nth1 - number of subgroups per workgroup
+            // ndst - number of output values per workgroup = output per subgroup * number of subgroups
+            if (backend_ctx->gpu_family == INTEL) {
+                nth0 = 16;
+                nth1 = 2;
+                ndst = nth1*4;
+            } else if (backend_ctx->gpu_family == ADRENO) {
+                nth0 = 64;
+                nth1 = 2;
+                ndst = nth1*4;
+            } else {
+                GGML_ASSERT(false && "TODO: Unknown GPU");
+            }
+
+            CL_CHECK(clSetKernelArg(kernel,  0, sizeof(cl_mem),   &extra0->data_device));
+            CL_CHECK(clSetKernelArg(kernel,  1, sizeof(cl_ulong), &offset0));
+            CL_CHECK(clSetKernelArg(kernel,  2, sizeof(cl_mem),   &extra1->data_device));
+            CL_CHECK(clSetKernelArg(kernel,  3, sizeof(cl_ulong), &offset1));
+            CL_CHECK(clSetKernelArg(kernel,  4, sizeof(cl_mem),   &extrad->data_device));
+            CL_CHECK(clSetKernelArg(kernel,  5, sizeof(cl_ulong), &offsetd));
+            CL_CHECK(clSetKernelArg(kernel,  6, sizeof(int),      &ne00));
+            CL_CHECK(clSetKernelArg(kernel,  7, sizeof(int),      &ne01));
+            CL_CHECK(clSetKernelArg(kernel,  8, sizeof(cl_ulong), &nb01));
+            CL_CHECK(clSetKernelArg(kernel,  9, sizeof(cl_ulong), &nb02));
+            CL_CHECK(clSetKernelArg(kernel, 10, sizeof(cl_ulong), &nb03));
+            CL_CHECK(clSetKernelArg(kernel, 11, sizeof(int),      &ne12));
+            CL_CHECK(clSetKernelArg(kernel, 12, sizeof(cl_ulong), &nb11));
+            CL_CHECK(clSetKernelArg(kernel, 13, sizeof(cl_ulong), &nb12));
+            CL_CHECK(clSetKernelArg(kernel, 14, sizeof(cl_ulong), &nb13));
+            CL_CHECK(clSetKernelArg(kernel, 15, sizeof(int),      &ne0));
+            CL_CHECK(clSetKernelArg(kernel, 16, sizeof(int),      &ne1));
+            CL_CHECK(clSetKernelArg(kernel, 17, sizeof(int),      &r2));
+            CL_CHECK(clSetKernelArg(kernel, 18, sizeof(int),      &r3));
+#endif // GGML_OPENCL_SOA_Q
+            break;
+        }
         case GGML_TYPE_Q2_K:
         case GGML_TYPE_Q3_K:
         case GGML_TYPE_Q4_K:
@@ -5349,11 +7280,87 @@ static void ggml_cl_mul_mat(ggml_backend_t backend, const ggml_tensor * src0, co
             CL_CHECK(clSetKernelArg(kernel, 13, sizeof(int),      &r2));
             CL_CHECK(clSetKernelArg(kernel, 14, sizeof(int),      &r3));
             break;
+        case GGML_TYPE_MXFP4: {
+#ifdef GGML_OPENCL_SOA_Q
+            kernel = backend_ctx->kernel_mul_mv_mxfp4_f32_flat;
+
+            cl_mem q;
+            if (backend_ctx->gpu_family == INTEL) {
+                nth0 = 16;
+                nth1 = 2;
+                ndst = nth1*2;
+
+                q = extra0_mxfp4->q;
+            } else if (backend_ctx->gpu_family == ADRENO) {
+                nth0 = 64;
+                nth1 = 2;
+                ndst = nth1*2;
+
+                q = extra0_mxfp4->q_img;
+            } else {
+                GGML_ASSERT(false && "TODO: Unknown GPU");
+            }
+
+            CL_CHECK(clSetKernelArg(kernel,  0, sizeof(cl_mem),   &q));
+            CL_CHECK(clSetKernelArg(kernel,  1, sizeof(cl_mem),   &extra0_mxfp4->e));
+            CL_CHECK(clSetKernelArg(kernel,  2, sizeof(cl_mem),   &extra1->data_device));
+            CL_CHECK(clSetKernelArg(kernel,  3, sizeof(cl_ulong), &offset1));
+            CL_CHECK(clSetKernelArg(kernel,  4, sizeof(cl_mem),   &extrad->data_device));
+            CL_CHECK(clSetKernelArg(kernel,  5, sizeof(cl_ulong), &offsetd));
+            CL_CHECK(clSetKernelArg(kernel,  6, sizeof(int),      &ne00));
+            CL_CHECK(clSetKernelArg(kernel,  7, sizeof(cl_ulong), &nb01));
+            CL_CHECK(clSetKernelArg(kernel,  8, sizeof(cl_ulong), &nb02));
+            CL_CHECK(clSetKernelArg(kernel,  9, sizeof(cl_ulong), &nb03));
+            CL_CHECK(clSetKernelArg(kernel, 10, sizeof(int),      &ne12));
+            CL_CHECK(clSetKernelArg(kernel, 11, sizeof(cl_ulong), &nb11));
+            CL_CHECK(clSetKernelArg(kernel, 12, sizeof(cl_ulong), &nb12));
+            CL_CHECK(clSetKernelArg(kernel, 13, sizeof(cl_ulong), &nb13));
+            CL_CHECK(clSetKernelArg(kernel, 14, sizeof(int),      &ne0));
+            CL_CHECK(clSetKernelArg(kernel, 15, sizeof(int),      &ne1));
+            CL_CHECK(clSetKernelArg(kernel, 16, sizeof(int),      &r2));
+            CL_CHECK(clSetKernelArg(kernel, 17, sizeof(int),      &r3));
+#else
+            kernel = backend_ctx->kernel_mul_mv_mxfp4_f32;
+
+            if (backend_ctx->gpu_family == INTEL) {
+                nth0 = 16;
+                nth1 = 2;
+                ndst = nth1*2;
+            } else if (backend_ctx->gpu_family == ADRENO) {
+                nth0 = 64;
+                nth1 = 2;
+                ndst = nth1*2;
+            } else {
+                GGML_ASSERT(false && "TODO: Unknown GPU");
+            }
+
+            CL_CHECK(clSetKernelArg(kernel,  0, sizeof(cl_mem),   &extra0->data_device));
+            CL_CHECK(clSetKernelArg(kernel,  1, sizeof(cl_ulong), &offset0));
+            CL_CHECK(clSetKernelArg(kernel,  2, sizeof(cl_mem),   &extra1->data_device));
+            CL_CHECK(clSetKernelArg(kernel,  3, sizeof(cl_ulong), &offset1));
+            CL_CHECK(clSetKernelArg(kernel,  4, sizeof(cl_mem),   &extrad->data_device));
+            CL_CHECK(clSetKernelArg(kernel,  5, sizeof(cl_ulong), &offsetd));
+            CL_CHECK(clSetKernelArg(kernel,  6, sizeof(int),      &ne00));
+            CL_CHECK(clSetKernelArg(kernel,  7, sizeof(cl_ulong), &nb01));
+            CL_CHECK(clSetKernelArg(kernel,  8, sizeof(cl_ulong), &nb02));
+            CL_CHECK(clSetKernelArg(kernel,  9, sizeof(cl_ulong), &nb03));
+            CL_CHECK(clSetKernelArg(kernel, 10, sizeof(int),      &ne12));
+            CL_CHECK(clSetKernelArg(kernel, 11, sizeof(cl_ulong), &nb11));
+            CL_CHECK(clSetKernelArg(kernel, 12, sizeof(cl_ulong), &nb12));
+            CL_CHECK(clSetKernelArg(kernel, 13, sizeof(cl_ulong), &nb13));
+            CL_CHECK(clSetKernelArg(kernel, 14, sizeof(int),      &ne0));
+            CL_CHECK(clSetKernelArg(kernel, 15, sizeof(int),      &ne1));
+            CL_CHECK(clSetKernelArg(kernel, 16, sizeof(int),      &r2));
+            CL_CHECK(clSetKernelArg(kernel, 17, sizeof(int),      &r3));
+            CL_CHECK(clSetKernelArg(kernel, 18, sizeof(float)*nth0,nullptr));
+#endif
+            break;
+        }
         default:
             GGML_ASSERT(false && "not implemented");
     }
 
-    if (src0t == GGML_TYPE_Q4_0 ||
+    if (src0t == GGML_TYPE_Q4_0 || src0t == GGML_TYPE_MXFP4 ||
         src0t == GGML_TYPE_Q4_1 ||
         src0t == GGML_TYPE_Q8_0 ||
         src0t == GGML_TYPE_Q2_K) {
@@ -5402,16 +7409,22 @@ static void ggml_cl_mul_mat_id(ggml_backend_t backend, const ggml_tensor * src0,
 
     ggml_backend_opencl_context *backend_ctx = (ggml_backend_opencl_context *)backend->context;
 
+    ggml_tensor_extra_cl * extra0 = (ggml_tensor_extra_cl *)src0->extra;
     ggml_tensor_extra_cl * extra1 = (ggml_tensor_extra_cl *)src1->extra;
     ggml_tensor_extra_cl * extra2 = (ggml_tensor_extra_cl *)src2->extra;
     ggml_tensor_extra_cl * extrad = (ggml_tensor_extra_cl *)dst->extra;
 
+    cl_ulong offset0 = extra0->offset + src0->view_offs;
     cl_ulong offset1 = extra1->offset + src1->view_offs;
     cl_ulong offset2 = extra2->offset + src2->view_offs;
     cl_ulong offsetd = extrad->offset + dst->view_offs;
 
+    GGML_UNUSED(offset0);
+
 #ifdef GGML_OPENCL_SOA_Q
     ggml_tensor_extra_cl_q4_0 * extra0_q4_0 = (ggml_tensor_extra_cl_q4_0 *)src0->extra;
+    ggml_tensor_extra_cl_mxfp4 * extra0_mxfp4 = (ggml_tensor_extra_cl_mxfp4 *)src0->extra;
+    ggml_tensor_extra_cl_q8_0 * extra0_q8_0 = (ggml_tensor_extra_cl_q8_0 *)src0->extra;
 #endif
 
     const int ne00 = src0->ne[0];
@@ -5420,7 +7433,9 @@ static void ggml_cl_mul_mat_id(ggml_backend_t backend, const ggml_tensor * src0,
     const int ne03 = src0->ne[3];
 
     const cl_ulong nb00 = src0->nb[0];
+    const cl_ulong nb01 = src0->nb[1];
     const cl_ulong nb02 = src0->nb[2];
+    const cl_ulong nb03 = src0->nb[3];
 
     const int ne10 = src1->ne[0];
     const int ne11 = src1->ne[1];
@@ -5429,6 +7444,7 @@ static void ggml_cl_mul_mat_id(ggml_backend_t backend, const ggml_tensor * src0,
 
     const cl_ulong nb11 = src1->nb[1];
     const cl_ulong nb12 = src1->nb[2];
+    const cl_ulong nb13 = src1->nb[3];
 
     const int ne20 = src2->ne[0];
     const int ne21 = src2->ne[1];
@@ -5496,6 +7512,170 @@ static void ggml_cl_mul_mat_id(ggml_backend_t backend, const ggml_tensor * src0,
 
             break;
         }
+        case GGML_TYPE_Q8_0: {
+#ifdef GGML_OPENCL_SOA_Q
+            kernel = backend_ctx->kernel_mul_mv_id_q8_0_f32_flat;
+
+            if (backend_ctx->gpu_family == INTEL) {
+                sgs  = 16;
+                nsg  = 2;
+                ndst = 4;
+            } else if (backend_ctx->gpu_family == ADRENO) {
+                sgs  = 64;
+                nsg  = 2;
+                ndst = 4;
+            } else {
+                GGML_ASSERT(false && "TODO: Unknown GPU");
+            }
+
+            CL_CHECK(clSetKernelArg(kernel,  0, sizeof(cl_mem),   &extra0_q8_0->q));
+            CL_CHECK(clSetKernelArg(kernel,  1, sizeof(cl_mem),   &extra0_q8_0->d));
+            CL_CHECK(clSetKernelArg(kernel,  2, sizeof(cl_mem),   &extra1->data_device));
+            CL_CHECK(clSetKernelArg(kernel,  3, sizeof(cl_ulong), &offset1));
+            CL_CHECK(clSetKernelArg(kernel,  4, sizeof(cl_mem),   &extra2->data_device));
+            CL_CHECK(clSetKernelArg(kernel,  5, sizeof(cl_ulong), &offset2));
+            CL_CHECK(clSetKernelArg(kernel,  6, sizeof(cl_mem),   &extrad->data_device));
+            CL_CHECK(clSetKernelArg(kernel,  7, sizeof(cl_ulong), &offsetd));
+            CL_CHECK(clSetKernelArg(kernel,  8, sizeof(int),      &ne00));
+            CL_CHECK(clSetKernelArg(kernel,  9, sizeof(int),      &ne01));
+            CL_CHECK(clSetKernelArg(kernel, 10, sizeof(cl_ulong), &nb01));
+            CL_CHECK(clSetKernelArg(kernel, 11, sizeof(cl_ulong), &nb02));
+            CL_CHECK(clSetKernelArg(kernel, 12, sizeof(int),      &ne11));
+            CL_CHECK(clSetKernelArg(kernel, 13, sizeof(int),      &ne12));
+            CL_CHECK(clSetKernelArg(kernel, 14, sizeof(cl_ulong), &nb11));
+            CL_CHECK(clSetKernelArg(kernel, 15, sizeof(cl_ulong), &nb12));
+            CL_CHECK(clSetKernelArg(kernel, 16, sizeof(int),      &ne20));
+            CL_CHECK(clSetKernelArg(kernel, 17, sizeof(int),      &ne21));
+            CL_CHECK(clSetKernelArg(kernel, 18, sizeof(cl_ulong), &nb21));
+            CL_CHECK(clSetKernelArg(kernel, 19, sizeof(int),      &ne0));
+            CL_CHECK(clSetKernelArg(kernel, 20, sizeof(int),      &ne1));
+#else
+            kernel = backend_ctx->kernel_mul_mv_id_q8_0_f32;
+
+            if (backend_ctx->gpu_family == INTEL) {
+                sgs  = 16;
+                nsg  = 2;
+                ndst = 4;
+            } else if (backend_ctx->gpu_family == ADRENO) {
+                sgs  = 64;
+                nsg  = 2;
+                ndst = 4;
+            } else {
+                GGML_ASSERT(false && "TODO: Unknown GPU");
+            }
+
+            CL_CHECK(clSetKernelArg(kernel,  0, sizeof(cl_mem),   &extra0->data_device));
+            CL_CHECK(clSetKernelArg(kernel,  1, sizeof(cl_ulong), &offset0));
+            CL_CHECK(clSetKernelArg(kernel,  2, sizeof(cl_mem),   &extra1->data_device));
+            CL_CHECK(clSetKernelArg(kernel,  3, sizeof(cl_ulong), &offset1));
+            CL_CHECK(clSetKernelArg(kernel,  4, sizeof(cl_mem),   &extra2->data_device));
+            CL_CHECK(clSetKernelArg(kernel,  5, sizeof(cl_ulong), &offset2));
+            CL_CHECK(clSetKernelArg(kernel,  6, sizeof(cl_mem),   &extrad->data_device));
+            CL_CHECK(clSetKernelArg(kernel,  7, sizeof(cl_ulong), &offsetd));
+            CL_CHECK(clSetKernelArg(kernel,  8, sizeof(int),      &ne00));
+            CL_CHECK(clSetKernelArg(kernel,  9, sizeof(int),      &ne01));
+            CL_CHECK(clSetKernelArg(kernel, 10, sizeof(cl_ulong), &nb01));
+            CL_CHECK(clSetKernelArg(kernel, 11, sizeof(cl_ulong), &nb02));
+            CL_CHECK(clSetKernelArg(kernel, 12, sizeof(int),      &ne11));
+            CL_CHECK(clSetKernelArg(kernel, 13, sizeof(int),      &ne12));
+            CL_CHECK(clSetKernelArg(kernel, 14, sizeof(cl_ulong), &nb11));
+            CL_CHECK(clSetKernelArg(kernel, 15, sizeof(cl_ulong), &nb12));
+            CL_CHECK(clSetKernelArg(kernel, 16, sizeof(int),      &ne20));
+            CL_CHECK(clSetKernelArg(kernel, 17, sizeof(int),      &ne21));
+            CL_CHECK(clSetKernelArg(kernel, 18, sizeof(cl_ulong), &nb21));
+            CL_CHECK(clSetKernelArg(kernel, 19, sizeof(int),      &ne0));
+            CL_CHECK(clSetKernelArg(kernel, 20, sizeof(int),      &ne1));
+#endif // GGML_OPENCL_SOA_Q
+            break;
+        }
+        case GGML_TYPE_MXFP4: {
+#ifdef GGML_OPENCL_SOA_Q
+            kernel = backend_ctx->kernel_mul_mv_id_mxfp4_f32_flat;
+
+            cl_mem q;
+            if (backend_ctx->gpu_family == INTEL) {
+                sgs  = 16;
+                nsg  = 2;
+                ndst = 2;
+
+                q = extra0_mxfp4->q;
+            } else if (backend_ctx->gpu_family == ADRENO) {
+                sgs  = 64;
+                nsg  = 1;
+                ndst = 4;
+
+                q = extra0_mxfp4->q_img;
+            } else {
+                GGML_ASSERT(false && "TODO: Unknown GPU");
+            }
+
+            CL_CHECK(clSetKernelArg(kernel,  0, sizeof(cl_mem),   &q));
+            CL_CHECK(clSetKernelArg(kernel,  1, sizeof(cl_mem),   &extra0_mxfp4->e));
+            CL_CHECK(clSetKernelArg(kernel,  2, sizeof(cl_mem),   &extra1->data_device));
+            CL_CHECK(clSetKernelArg(kernel,  3, sizeof(cl_ulong), &offset1));
+            CL_CHECK(clSetKernelArg(kernel,  4, sizeof(cl_mem),   &extra2->data_device));
+            CL_CHECK(clSetKernelArg(kernel,  5, sizeof(cl_ulong), &offset2));
+            CL_CHECK(clSetKernelArg(kernel,  6, sizeof(cl_mem),   &extrad->data_device));
+            CL_CHECK(clSetKernelArg(kernel,  7, sizeof(cl_ulong), &offsetd));
+            CL_CHECK(clSetKernelArg(kernel,  8, sizeof(int),      &ne00));
+            CL_CHECK(clSetKernelArg(kernel,  9, sizeof(cl_ulong), &nb01));
+            CL_CHECK(clSetKernelArg(kernel, 10, sizeof(cl_ulong), &nb02));
+            CL_CHECK(clSetKernelArg(kernel, 11, sizeof(cl_ulong), &nb03));
+            CL_CHECK(clSetKernelArg(kernel, 12, sizeof(int),      &ne11));
+            CL_CHECK(clSetKernelArg(kernel, 13, sizeof(int),      &ne12));
+            CL_CHECK(clSetKernelArg(kernel, 14, sizeof(cl_ulong), &nb11));
+            CL_CHECK(clSetKernelArg(kernel, 15, sizeof(cl_ulong), &nb12));
+            CL_CHECK(clSetKernelArg(kernel, 16, sizeof(cl_ulong), &nb13));
+            CL_CHECK(clSetKernelArg(kernel, 17, sizeof(int),      &ne20));
+            CL_CHECK(clSetKernelArg(kernel, 18, sizeof(int),      &ne21));
+            CL_CHECK(clSetKernelArg(kernel, 19, sizeof(cl_ulong), &nb21));
+            CL_CHECK(clSetKernelArg(kernel, 20, sizeof(int),      &ne0));
+            CL_CHECK(clSetKernelArg(kernel, 21, sizeof(int),      &ne1));
+            CL_CHECK(clSetKernelArg(kernel, 22, sizeof(int),      &r2));
+            CL_CHECK(clSetKernelArg(kernel, 23, sizeof(int),      &r3));
+#else // GGML_OPENCL_SOA_Q
+            kernel = backend_ctx->kernel_mul_mv_id_mxfp4_f32;
+
+            if (backend_ctx->gpu_family == INTEL) {
+                sgs  = 16;
+                nsg  = 2;
+                ndst = 2;
+            } else if (backend_ctx->gpu_family == ADRENO) {
+                sgs  = 64;
+                nsg  = 2;
+                ndst = 2;
+            } else {
+                GGML_ASSERT(false && "TODO: Unknown GPU");
+            }
+
+            CL_CHECK(clSetKernelArg(kernel,  0, sizeof(cl_mem),   &extra0->data_device));
+            CL_CHECK(clSetKernelArg(kernel,  1, sizeof(cl_ulong), &offset0));
+            CL_CHECK(clSetKernelArg(kernel,  2, sizeof(cl_mem),   &extra1->data_device));
+            CL_CHECK(clSetKernelArg(kernel,  3, sizeof(cl_ulong), &offset1));
+            CL_CHECK(clSetKernelArg(kernel,  4, sizeof(cl_mem),   &extra2->data_device));
+            CL_CHECK(clSetKernelArg(kernel,  5, sizeof(cl_ulong), &offset2));
+            CL_CHECK(clSetKernelArg(kernel,  6, sizeof(cl_mem),   &extrad->data_device));
+            CL_CHECK(clSetKernelArg(kernel,  7, sizeof(cl_ulong), &offsetd));
+            CL_CHECK(clSetKernelArg(kernel,  8, sizeof(int),      &ne00));
+            CL_CHECK(clSetKernelArg(kernel,  9, sizeof(cl_ulong), &nb01));
+            CL_CHECK(clSetKernelArg(kernel, 10, sizeof(cl_ulong), &nb02));
+            CL_CHECK(clSetKernelArg(kernel, 11, sizeof(cl_ulong), &nb03));
+            CL_CHECK(clSetKernelArg(kernel, 12, sizeof(int),      &ne11));
+            CL_CHECK(clSetKernelArg(kernel, 13, sizeof(int),      &ne12));
+            CL_CHECK(clSetKernelArg(kernel, 14, sizeof(cl_ulong), &nb11));
+            CL_CHECK(clSetKernelArg(kernel, 15, sizeof(cl_ulong), &nb12));
+            CL_CHECK(clSetKernelArg(kernel, 16, sizeof(cl_ulong), &nb13));
+            CL_CHECK(clSetKernelArg(kernel, 17, sizeof(int),      &ne20));
+            CL_CHECK(clSetKernelArg(kernel, 18, sizeof(int),      &ne21));
+            CL_CHECK(clSetKernelArg(kernel, 19, sizeof(cl_ulong), &nb21));
+            CL_CHECK(clSetKernelArg(kernel, 20, sizeof(int),      &ne0));
+            CL_CHECK(clSetKernelArg(kernel, 21, sizeof(int),      &ne1));
+            CL_CHECK(clSetKernelArg(kernel, 22, sizeof(int),      &r2));
+            CL_CHECK(clSetKernelArg(kernel, 23, sizeof(int),      &r3));
+            CL_CHECK(clSetKernelArg(kernel, 24, sizeof(float)*sgs,nullptr));
+#endif // GGML_OPENCL_SOA_Q
+            break;
+        }
         default:
             GGML_ASSERT(false && "not implemented");;
     }
@@ -5521,7 +7701,9 @@ static void ggml_cl_scale(ggml_backend_t backend, const ggml_tensor * src0, cons
     ggml_backend_opencl_context *backend_ctx = (ggml_backend_opencl_context *)backend->context;
 
     float scale;
-    memcpy(&scale, dst->op_params, sizeof(scale));
+    float bias;
+    memcpy(&scale, ((int32_t *) dst->op_params) + 0, sizeof(float));
+    memcpy(&bias,  ((int32_t *) dst->op_params) + 1, sizeof(float));
 
     ggml_tensor_extra_cl * extra0 = (ggml_tensor_extra_cl *)src0->extra;
     ggml_tensor_extra_cl * extrad = (ggml_tensor_extra_cl *)dst->extra;
@@ -5536,6 +7718,7 @@ static void ggml_cl_scale(ggml_backend_t backend, const ggml_tensor * src0, cons
     CL_CHECK(clSetKernelArg(kernel, 2, sizeof(cl_mem),   &extrad->data_device));
     CL_CHECK(clSetKernelArg(kernel, 3, sizeof(cl_ulong), &offsetd));
     CL_CHECK(clSetKernelArg(kernel, 4, sizeof(float),    &scale));
+    CL_CHECK(clSetKernelArg(kernel, 5, sizeof(float),    &bias));
 
     int n = ggml_nelements(dst)/4;
 
@@ -5733,31 +7916,50 @@ static void ggml_cl_soft_max(ggml_backend_t backend, const ggml_tensor * src0, c
         GGML_ASSERT(src1->extra);
     }
 
+    const ggml_tensor * src2 = dst->src[2];
+    if (src2) {
+        GGML_ASSERT(src2->extra);
+    }
+
     ggml_backend_opencl_context *backend_ctx = (ggml_backend_opencl_context *)backend->context;
 
     ggml_tensor_extra_cl * extra0 = (ggml_tensor_extra_cl *)src0->extra;
     ggml_tensor_extra_cl * extrad = (ggml_tensor_extra_cl *)dst->extra;
 
     ggml_tensor_extra_cl * extra1 = src1 ? (ggml_tensor_extra_cl *)src1->extra : nullptr;
+    ggml_tensor_extra_cl * extra2 = src2 ? (ggml_tensor_extra_cl *)src2->extra : nullptr;
 
     cl_ulong offset0 = extra0->offset + src0->view_offs;
     cl_ulong offsetd = extrad->offset + dst->view_offs;
 
     cl_ulong offset1 = extra1 ? extra1->offset + src1->view_offs : offset0;
+    cl_ulong offset2 = extra2 ? extra2->offset + src2->view_offs : offset0;
 
-    const int  ne00 = src0 ? src0->ne[0] : 0;
-    const int  ne01 = src0 ? src0->ne[1] : 0;
-    const int  ne02 = src0 ? src0->ne[2] : 0;
-    const int  ne03 = src0 ? src0->ne[3] : 0;
+    const int ne00 = src0->ne[0];
+    const int ne01 = src0->ne[1];
+    const int ne02 = src0->ne[2];
+    const int ne03 = src0->ne[3];
+
+    const cl_long nb01 = src0->nb[1];
+    const cl_long nb02 = src0->nb[2];
+    const cl_long nb03 = src0->nb[3];
+
+    const int ne12 = src1 ? src1->ne[2] : 0;
+    const int ne13 = src1 ? src1->ne[3] : 0;
+
+    const cl_long nb11 = src1 ? src1->nb[1] : 0;
+    const cl_long nb12 = src1 ? src1->nb[2] : 0;
+    const cl_long nb13 = src1 ? src1->nb[3] : 0;
+
+    const cl_long nb1 = dst->nb[1];
+    const cl_long nb2 = dst->nb[2];
+    const cl_long nb3 = dst->nb[3];
 
     float scale, max_bias;
     memcpy(&scale,    dst->op_params + 0, sizeof(float));
     memcpy(&max_bias, dst->op_params + 1, sizeof(float));
 
-    const int nrows_x = ggml_nrows(src0);
-    const int nrows_y = src0->ne[1];
-
-    const int n_head      = nrows_x/nrows_y;
+    const int n_head      = src0->ne[2];
     const int n_head_log2 = 1u << (uint32_t) floorf(log2f((float) n_head));
 
     const float m0 = powf(2.0f, -(max_bias       ) / n_head_log2);
@@ -5799,16 +8001,27 @@ static void ggml_cl_soft_max(ggml_backend_t backend, const ggml_tensor * src0, c
     CL_CHECK(clSetKernelArg(kernel,  1, sizeof(cl_ulong), &offset0));
     CL_CHECK(clSetKernelArg(kernel,  2, sizeof(cl_mem),   extra1 ? &extra1->data_device : &extra0->data_device));
     CL_CHECK(clSetKernelArg(kernel,  3, sizeof(cl_ulong), &offset1));
-    CL_CHECK(clSetKernelArg(kernel,  4, sizeof(cl_mem),   &extrad->data_device));
-    CL_CHECK(clSetKernelArg(kernel,  5, sizeof(cl_ulong), &offsetd));
-    CL_CHECK(clSetKernelArg(kernel,  6, sizeof(int),      &ne00));
-    CL_CHECK(clSetKernelArg(kernel,  7, sizeof(int),      &ne01));
-    CL_CHECK(clSetKernelArg(kernel,  8, sizeof(int),      &ne02));
-    CL_CHECK(clSetKernelArg(kernel,  9, sizeof(float),    &scale));
-    CL_CHECK(clSetKernelArg(kernel, 10, sizeof(float),    &max_bias));
-    CL_CHECK(clSetKernelArg(kernel, 11, sizeof(float),    &m0));
-    CL_CHECK(clSetKernelArg(kernel, 12, sizeof(float),    &m1));
-    CL_CHECK(clSetKernelArg(kernel, 13, sizeof(int),      &n_head_log2));
+    CL_CHECK(clSetKernelArg(kernel,  4, sizeof(cl_mem),   extra2 ? &extra2->data_device : &extra0->data_device));
+    CL_CHECK(clSetKernelArg(kernel,  5, sizeof(cl_ulong), &offset2));
+    CL_CHECK(clSetKernelArg(kernel,  6, sizeof(cl_mem),   &extrad->data_device));
+    CL_CHECK(clSetKernelArg(kernel,  7, sizeof(cl_ulong), &offsetd));
+    CL_CHECK(clSetKernelArg(kernel,  8, sizeof(int),      &ne00));
+    CL_CHECK(clSetKernelArg(kernel,  9, sizeof(cl_ulong), &nb01));
+    CL_CHECK(clSetKernelArg(kernel, 10, sizeof(cl_ulong), &nb02));
+    CL_CHECK(clSetKernelArg(kernel, 11, sizeof(cl_ulong), &nb03));
+    CL_CHECK(clSetKernelArg(kernel, 12, sizeof(int),      &ne12));
+    CL_CHECK(clSetKernelArg(kernel, 13, sizeof(int),      &ne13));
+    CL_CHECK(clSetKernelArg(kernel, 14, sizeof(cl_ulong), &nb11));
+    CL_CHECK(clSetKernelArg(kernel, 15, sizeof(cl_ulong), &nb12));
+    CL_CHECK(clSetKernelArg(kernel, 16, sizeof(cl_ulong), &nb13));
+    CL_CHECK(clSetKernelArg(kernel, 17, sizeof(cl_ulong), &nb1));
+    CL_CHECK(clSetKernelArg(kernel, 18, sizeof(cl_ulong), &nb2));
+    CL_CHECK(clSetKernelArg(kernel, 19, sizeof(cl_ulong), &nb3));
+    CL_CHECK(clSetKernelArg(kernel, 20, sizeof(float),    &scale));
+    CL_CHECK(clSetKernelArg(kernel, 21, sizeof(float),    &max_bias));
+    CL_CHECK(clSetKernelArg(kernel, 22, sizeof(float),    &m0));
+    CL_CHECK(clSetKernelArg(kernel, 23, sizeof(float),    &m1));
+    CL_CHECK(clSetKernelArg(kernel, 24, sizeof(int),      &n_head_log2));
 
     size_t global_work_size[] = {(size_t)ne01*nth, (size_t)ne02, (size_t)ne03};
     size_t local_work_size[] = {(size_t)nth, 1, 1};
@@ -6215,6 +8428,23 @@ static void ggml_cl_glu(ggml_backend_t backend, const ggml_tensor * src0, const
                 kernel = backend_ctx->kernel_swiglu_f16;
             }
             break;
+        case GGML_GLU_OP_SWIGLU_OAI:
+            kernel = backend_ctx->kernel_swiglu_oai;
+            break;
+        case GGML_GLU_OP_GEGLU_ERF:
+            if (dst->type == GGML_TYPE_F32) {
+                kernel = backend_ctx->kernel_geglu_erf;
+            } else {
+                kernel = backend_ctx->kernel_geglu_erf_f16;
+            }
+            break;
+        case GGML_GLU_OP_GEGLU_QUICK:
+            if (dst->type == GGML_TYPE_F32) {
+                kernel = backend_ctx->kernel_geglu_quick;
+            } else {
+                kernel = backend_ctx->kernel_geglu_quick_f16;
+            }
+            break;
         default:
             GGML_ABORT("Unsupported glu op");
     }
@@ -6236,7 +8466,10 @@ static void ggml_cl_glu(ggml_backend_t backend, const ggml_tensor * src0, const
 
     const cl_ulong nb1  = dst->nb[1];
 
-    const int swp = ((const int32_t *) dst->op_params)[1];
+    const int   swp   = ggml_get_op_params_i32(dst, 1);
+    const float alpha = ggml_get_op_params_f32(dst, 2);
+    const float limit = ggml_get_op_params_f32(dst, 3);
+
     const int ne00_off = src1 ? 0 : (swp ? ne0 : 0);
     const int ne10_off = src1 ? 0 : (swp ? 0 : ne0);
 
@@ -6253,6 +8486,11 @@ static void ggml_cl_glu(ggml_backend_t backend, const ggml_tensor * src0, const
     CL_CHECK(clSetKernelArg(kernel, 10, sizeof(int),      &ne00_off));
     CL_CHECK(clSetKernelArg(kernel, 11, sizeof(int),      &ne10_off));
 
+    if (ggml_get_glu_op(dst) == GGML_GLU_OP_SWIGLU_OAI) {
+        CL_CHECK(clSetKernelArg(kernel, 12, sizeof(float), &limit));
+        CL_CHECK(clSetKernelArg(kernel, 13, sizeof(float), &alpha));
+    }
+
     const size_t nrows = ggml_nrows(src0);
     size_t nth = 512;
     size_t global_work_size[] = {nrows*nth, 1, 1};
@@ -6284,6 +8522,12 @@ bool ggml_cl_compute_forward(ggml_backend_t backend, struct ggml_tensor * tensor
             }
             func = ggml_cl_get_rows;
             break;
+        case GGML_OP_SET_ROWS:
+            if (!any_on_device) {
+                return false;
+            }
+            func = ggml_cl_set_rows;
+            break;
         case GGML_OP_CPY:
             if (!any_on_device) {
                 return false;
@@ -6303,6 +8547,12 @@ bool ggml_cl_compute_forward(ggml_backend_t backend, struct ggml_tensor * tensor
             }
             func = ggml_cl_add;
             break;
+        case GGML_OP_ADD_ID:
+            if (!any_on_device) {
+                return false;
+            }
+            func = ggml_cl_add_id;
+            break;
         case GGML_OP_MUL:
             if (!any_on_device) {
                 return false;
@@ -6329,6 +8579,12 @@ bool ggml_cl_compute_forward(ggml_backend_t backend, struct ggml_tensor * tensor
                     }
                     func = ggml_cl_gelu;
                     break;
+                case GGML_UNARY_OP_GELU_ERF:
+                    if (!any_on_device) {
+                        return false;
+                    }
+                    func = ggml_cl_gelu_erf;
+                    break;
                 case GGML_UNARY_OP_GELU_QUICK:
                     if (!any_on_device) {
                         return false;
@@ -6410,6 +8666,12 @@ bool ggml_cl_compute_forward(ggml_backend_t backend, struct ggml_tensor * tensor
             }
             ggml_cl_upscale(backend, tensor->src[0], tensor);
             return true;
+        case GGML_OP_CONV_2D:
+            if (!any_on_device) {
+                return false;
+            }
+            func = ggml_cl_conv_2d;
+            break;
         case GGML_OP_CONCAT:
             if (!any_on_device) {
                 return false;
@@ -6485,6 +8747,12 @@ bool ggml_cl_compute_forward(ggml_backend_t backend, struct ggml_tensor * tensor
             }
             func = ggml_cl_sum_rows;
             break;
+        case GGML_OP_FLASH_ATTN_EXT:
+            if (!any_on_device) {
+                return false;
+            }
+            ggml_cl_flash_attn(backend, tensor->src[0], tensor->src[1], tensor);
+            return true;
         default:
             return false;
     }