ggml: fix loongarch quantize_row_q8_1 error (#14827)

* weight format to nz for 310p

* remove quant weight format to nz

* clean code

* fix

* make the conditions for converting weights to NZ format consistent

* clean code

README.md

@@ -270,7 +270,6 @@ Instructions for adding support for new models: [HOWTO-add-model.md](docs/develo
 | [CANN](docs/build.md#cann) | Ascend NPU |
 | [OpenCL](docs/backend/OPENCL.md) | Adreno GPU |
 | [WebGPU [In Progress]](docs/build.md#webgpu) | All |
-
 | [RPC](https://github.com/ggml-org/llama.cpp/tree/master/tools/rpc) | All |
 
 ## Obtaining and quantizing models

common/arg.cpp

@@ -1612,7 +1612,7 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
         [](common_params & params, const std::string & value) {
             params.antiprompt.emplace_back(value);
         }
-    ).set_examples({LLAMA_EXAMPLE_MAIN}));
+    ).set_examples({LLAMA_EXAMPLE_MAIN, LLAMA_EXAMPLE_SERVER}));
     add_opt(common_arg(
         {"-sp", "--special"},
         string_format("special tokens output enabled (default: %s)", params.special ? "true" : "false"),
@@ -2655,6 +2655,13 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
             params.i_chunk = value;
         }
     ).set_examples({LLAMA_EXAMPLE_IMATRIX}));
+    add_opt(common_arg(
+        {"--show-statistics"},
+        string_format("show imatrix statistics and then exit (default: %s)", params.show_statistics ? "true" : "false"),
+        [](common_params & params) {
+            params.show_statistics = true;
+        }
+    ).set_examples({LLAMA_EXAMPLE_IMATRIX}));
     add_opt(common_arg(
         {"--parse-special"},
         string_format("prase special tokens (chat, tool, etc) (default: %s)", params.parse_special ? "true" : "false"),

common/common.h

@@ -432,9 +432,10 @@ struct common_params {
     int32_t n_save_freq =  0; // save the imatrix every n_save_freq iterations
     int32_t i_chunk     =  0; // start processing from this chunk
 
-    bool process_output = false; // collect data for the output tensor
-    bool compute_ppl    = true;  // whether to compute perplexity
-    bool parse_special  = false; // whether to parse special tokens during imatrix tokenization
+    bool process_output  = false; // collect data for the output tensor
+    bool compute_ppl     = true;  // whether to compute perplexity
+    bool show_statistics = false; // show imatrix statistics per tensor
+    bool parse_special   = false; // whether to parse special tokens during imatrix tokenization
 
     // cvector-generator params
     int n_pca_batch = 100;

ggml/src/ggml-cann/aclnn_ops.cpp

@@ -1785,8 +1785,27 @@ static void ggml_cann_mat_mul_fp(ggml_backend_cann_context& ctx,
     size_t transpose_nb[] = {bcast_weight_nb[1], bcast_weight_nb[0],
                              bcast_weight_nb[2], bcast_weight_nb[3],
                              bcast_weight_nb[4], bcast_weight_nb[5]};
-    aclTensor* acl_weight_tensor =
-        ggml_cann_create_tensor(weight, transpose_ne, transpose_nb, n_dims);
+    aclTensor* acl_weight_tensor;
+
+    bool weightToNZ = false;
+#ifdef ASCEND_310P
+    weightToNZ = (getenv("GGML_CANN_WEIGHT_NZ") != nullptr);
+#endif
+    if (weightToNZ && is_matmul_weight(weight)) {
+        int64_t acl_stride[2] = {1, transpose_ne[1]};
+
+        // Reverse ne.
+        std::reverse(transpose_ne, transpose_ne + n_dims);
+
+        std::vector<int64_t> storageDims = {transpose_ne[0], transpose_ne[1]};
+
+        acl_weight_tensor = aclCreateTensor(
+            transpose_ne, n_dims, ggml_cann_type_mapping(weight->type), acl_stride,
+            0, ACL_FORMAT_FRACTAL_NZ, storageDims.data(), 2, weight->data);
+    } else {
+        acl_weight_tensor =
+            ggml_cann_create_tensor(weight, transpose_ne, transpose_nb, n_dims, ACL_FORMAT_ND);
+    }
     aclTensor* acl_dst =
         ggml_cann_create_tensor(dst, bcast_dst_ne, bcast_dst_nb, n_dims);
 

ggml/src/ggml-cann/aclnn_ops.h

@@ -23,6 +23,7 @@
 #ifndef CANN_ACLNN_OPS
 #define CANN_ACLNN_OPS
 
+#include <unordered_set>
 #include <functional>
 #include <aclnnop/aclnn_abs.h>
 #include <aclnnop/aclnn_neg.h>
@@ -1020,6 +1021,37 @@ inline void ggml_cann_async_memset(ggml_backend_cann_context & ctx, void * buffe
  */
 void ggml_cann_mul_mat_id(ggml_backend_cann_context& ctx, ggml_tensor* dst);
 
+/**
+ * @brief   Check whether a tensor is a weight tensor for matrix multiplication.
+ *
+ * @details Checks whether the given tensor serves as weight parameters in matrix multiplication operations,
+ *          typically within neural network layers. The function maintains a static set of canonical weight
+ *          naming suffixes from Transformer-based architectures. Uses substring matching to identify weight
+ *          tensors even with hierarchical naming patterns.
+ *
+ * @param tensor Pointer to the target ggml_tensor object (const-qualified).
+ */
+static bool is_matmul_weight(const ggml_tensor* tensor) {
+    std::string name = ggml_get_name(tensor);
+    static const std::unordered_set<std::string> weight_suffixes{
+        "output.weight",
+        "attn_q.weight",
+        "attn_k.weight",
+        "attn_v.weight",
+        "attn_output.weight",
+        "ffn_gate.weight",
+        "ffn_up.weight",
+        "ffn_down.weight"
+    };
+
+    for (const auto& suffix : weight_suffixes) {
+        if (name.find(suffix) != std::string::npos) {
+            return true;
+        }
+    }
+    return false;
+}
+
 /**
  * @brief Applies a element-wise operation to two input tensors using the CANN
  * backend.

ggml/src/ggml-cann/ggml-cann.cpp

@@ -24,6 +24,7 @@
 
 #include <acl/acl.h>
 #include <stdarg.h>
+#include <aclnnop/aclnn_trans_matmul_weight.h>
 
 #include <cmath>
 #include <cstdio>
@@ -1115,6 +1116,63 @@ static enum ggml_status ggml_backend_cann_buffer_init_tensor(
     return GGML_STATUS_SUCCESS;
 }
 
+static int CreateAclTensorWeight(const void *hostData, const std::vector<int64_t> &shape, void **deviceAddr,
+                      aclDataType dataType, aclTensor **tensor)
+{
+    uint64_t size = 1;
+    for (auto i : shape) {
+        size *= i;
+    }
+
+    const aclIntArray *mat2Size = aclCreateIntArray(shape.data(), shape.size());
+    ACL_CHECK(aclnnCalculateMatmulWeightSizeV2(mat2Size, dataType, &size));
+
+    size *= sizeof(int16_t);
+
+    ACL_CHECK(aclrtMalloc(deviceAddr, size, ACL_MEM_MALLOC_HUGE_FIRST));
+    aclrtMemcpy(*deviceAddr, size, hostData, size, ACL_MEMCPY_HOST_TO_DEVICE);
+
+    std::vector<int64_t> strides(shape.size(), 1);
+    for (int64_t i = shape.size() - 2; i >= 0; i--) {
+        strides[i] = shape[i + 1] * strides[i + 1];
+    }
+
+    *tensor = aclCreateTensor(shape.data(), shape.size(), dataType, strides.data(), 0, aclFormat::ACL_FORMAT_ND,
+                              shape.data(), shape.size(), *deviceAddr);
+    return 0;
+}
+
+static void weight_format_to_nz(ggml_tensor *tensor, const void *data, size_t offset) {
+    aclrtStream stream;
+    ACL_CHECK(aclrtCreateStream(&stream));
+
+    std::vector<int64_t> weightTransposedShape = {tensor->ne[1], tensor->ne[0]};
+    void *weightTransposedDeviceAddr = nullptr;
+    aclTensor *weightTransposed = nullptr;
+    CreateAclTensorWeight(data, weightTransposedShape, &weightTransposedDeviceAddr,
+                          ggml_cann_type_mapping(tensor->type), &weightTransposed);
+
+    uint64_t workspaceSize = 0;
+    aclOpExecutor *executor;
+    void *workspaceAddr = nullptr;
+
+    // TransMatmulWeight
+    ACL_CHECK(aclnnTransMatmulWeightGetWorkspaceSize(weightTransposed, &workspaceSize, &executor));
+    std::unique_ptr<void, aclError (*)(void *)> workspaceAddrPtrTrans(nullptr, aclrtFree);
+    if (workspaceSize > 0) {
+        ACL_CHECK(aclrtMalloc(&workspaceAddr, workspaceSize, ACL_MEM_MALLOC_HUGE_FIRST));
+        workspaceAddrPtrTrans.reset(workspaceAddr);
+    }
+    ACL_CHECK(aclnnTransMatmulWeight(workspaceAddr, workspaceSize, executor, stream));
+
+    size_t size = ggml_nelements(tensor) * ggml_element_size(tensor);
+
+    aclrtMemcpy((char *)tensor->data + offset, size,
+                weightTransposedDeviceAddr, size, ACL_MEMCPY_HOST_TO_DEVICE);
+    ACL_CHECK(aclDestroyTensor(weightTransposed));
+    aclrtFree(weightTransposedDeviceAddr);
+}
+
 // TODO: need handle tensor which has paddings.
 /**
  * @brief Set tensor data in a CANN buffer.
@@ -1139,9 +1197,16 @@ static void ggml_backend_cann_buffer_set_tensor(
     // For acl, synchronous functions use this default stream.
     // Why aclrtSynchronizeDevice?
 
+    bool weightToNZ = false;
+#ifdef ASCEND_310P
+    weightToNZ = (getenv("GGML_CANN_WEIGHT_NZ") != nullptr);
+#endif
     if (!need_transform(tensor->type)) {
         ACL_CHECK(aclrtMemcpy((char *)tensor->data + offset, size, data, size,
                               ACL_MEMCPY_HOST_TO_DEVICE));
+        if (weightToNZ && is_matmul_weight((const ggml_tensor*)tensor)) {
+            weight_format_to_nz(tensor, data, offset);
+        }
     } else {
         void *transform_buffer = malloc(size);
         ggml_backend_cann_transform(tensor, data, transform_buffer);

ggml/src/ggml-cpu/CMakeLists.txt

@@ -494,9 +494,9 @@ function(ggml_add_cpu_backend_variant_impl tag_name)
 
         # Fetch KleidiAI sources:
         include(FetchContent)
-        set(KLEIDIAI_COMMIT_TAG "v1.9.0")
+        set(KLEIDIAI_COMMIT_TAG "v1.11.0")
         set(KLEIDIAI_DOWNLOAD_URL "https://github.com/ARM-software/kleidiai/archive/refs/tags/${KLEIDIAI_COMMIT_TAG}.tar.gz")
-        set(KLEIDIAI_ARCHIVE_MD5  "2a8e1bb55d201557553545536489a017")
+        set(KLEIDIAI_ARCHIVE_MD5  "3fe9e5ab964c375c53839296eb71eaa2")
 
         if (POLICY CMP0135)
             cmake_policy(SET CMP0135 NEW)

ggml/src/ggml-cpu/arch/loongarch/quants.c

@@ -544,7 +544,7 @@ void quantize_row_q8_1(const float * GGML_RESTRICT x, void * GGML_RESTRICT vy, i
         __m128 max4 = __lsx_vfmax_s( lasx_extractf128( max_abs, 1 ), lasx_extractf128( max_abs, 0) );
         max4 = __lsx_vfmax_s( max4, (__m128)__lsx_vpickod_d((__m128i) max4, (__m128i)max4 ) );
         __m128 tmp = max4;
-        max4 = __lsx_vfmax_s( max4, (__m128)__lsx_vextrins_w((__m128i)tmp, (__m128i)max4, 0x10 ));
+        max4 = __lsx_vfmax_s( max4, (__m128)__lsx_vextrins_w((__m128i)tmp, (__m128i)max4, 0x1 ));
         const float max_scalar = ((v4f32)max4)[0];
 
         // Quantize these floats

ggml/src/ggml-cpu/kleidiai/kernels.cpp

@@ -22,9 +22,94 @@
 
 #include "kai_common.h"
 
+#include "simd-mappings.h"
+
 #include "kernels.h"
 
 #define NELEMS(x) sizeof(x) / sizeof(*x)
+
+static const size_t INT4_PER_BYTE = 2;
+static const size_t INT4_BITS     = 4;
+static const int Q4_0_ZERO_POINT  = 8;
+const size_t INT4_PER_UINT16      = 4;
+
+static void dequantize_row_qsi4c32pscalef16(
+    const void *packed_data,
+    int32_t row_idx,
+    int64_t nc,
+    float *out,
+    size_t nr_pack,
+    size_t packed_row_stride,
+    size_t kr,
+    size_t bl,
+    size_t num_bytes_multiplier
+) {
+    size_t group_idx = row_idx / nr_pack;
+    size_t row_in_group = row_idx % nr_pack;
+    const uint8_t *packed_group = (const uint8_t *)packed_data + group_idx * packed_row_stride;
+    size_t num_blocks = nc / bl;
+    const uint8_t *block_ptr = packed_group;
+
+    for (size_t b = 0; b < num_blocks; ++b) {
+        uint16_t scale_f16 = *((const uint16_t *)(block_ptr + row_in_group * num_bytes_multiplier));
+        float scale = GGML_CPU_FP16_TO_FP32(scale_f16);
+
+        const uint8_t *segment_ptr = block_ptr + nr_pack * num_bytes_multiplier;
+        size_t num_segments = bl / kr;
+        size_t num_bytes_per_segment = kr / INT4_PER_BYTE;
+
+        for (size_t s = 0; s < num_segments; ++s) {
+            const uint8_t *seg_base = segment_ptr + s * nr_pack * num_bytes_per_segment;
+            const uint8_t *qbytes = seg_base + row_in_group * num_bytes_per_segment;
+            for (size_t k = 0; k < num_bytes_per_segment; ++k) {
+                uint8_t byte = qbytes[k] ^ 0x88;
+                int x0 = (byte & 0x0F) - Q4_0_ZERO_POINT;
+                int x1 = (byte >> INT4_BITS) - Q4_0_ZERO_POINT;
+                out[b * bl + s * num_bytes_per_segment + k] = x0 * scale;
+                out[b * bl + s * num_bytes_per_segment + k + bl/2] = x1 * scale;
+            }
+        }
+        block_ptr += nr_pack * num_bytes_multiplier + num_segments * nr_pack * num_bytes_per_segment;
+    }
+}
+
+static void dequantize_row_qsi4c32ps1s0scalef16(
+    const void *packed_data,
+    int32_t row_idx,
+    int64_t k,
+    float *out,
+    size_t nr,
+    size_t packed_row_stride,
+    size_t kr,
+    size_t bl,
+    size_t num_bytes_multiplier
+) {
+    const size_t num_blocks = k / bl;
+    const size_t bl4 = bl / INT4_PER_UINT16;
+
+    size_t group_idx = row_idx / nr;
+    size_t row_in_group = row_idx % nr;
+
+    const uint8_t *packed_group = (const uint8_t *)packed_data + group_idx * packed_row_stride;
+    const uint16_t *qdata = (const uint16_t *)packed_group;
+    const uint16_t *scales = (const uint16_t *)(packed_group + packed_row_stride - (nr * num_blocks * num_bytes_multiplier));
+
+    for (size_t block_idx = 0; block_idx < num_blocks; ++block_idx) {
+        uint16_t scale_f16 = scales[row_in_group + block_idx * nr];
+        float scale = GGML_CPU_FP16_TO_FP32(scale_f16);
+
+        for (size_t bl4_idx = 0; bl4_idx < bl4; ++bl4_idx) {
+            uint16_t q = qdata[(block_idx * bl4 + bl4_idx) * nr + row_in_group];
+
+            for (size_t qidx = 0; qidx < INT4_PER_UINT16; ++qidx) {
+                int v = ((q >> (qidx * 4)) & 0xF) - Q4_0_ZERO_POINT;
+                out[block_idx * bl + bl4_idx * INT4_BITS + qidx] = v * scale;
+            }
+        }
+    }
+    GGML_UNUSED(kr);
+}
+
 static ggml_kleidiai_kernels gemm_gemv_kernels[] = {
 #if defined(__ARM_FEATURE_SME)
     {
@@ -63,8 +148,10 @@ static ggml_kleidiai_kernels gemm_gemv_kernels[] = {
             /* .pack_func             = */ kai_run_lhs_quant_pack_qsi8d32p_f32_neon,
         },
         /* .rhs_info = */ {
-            /* .packed_size = */ kai_get_rhs_packed_size_rhs_pack_nxk_qsi4c32ps1s0scalef16_qsu4c32s16s0_neon,
-            /* .pack_func   = */ kai_run_rhs_pack_nxk_qsi4c32ps1s0scalef16_qsu4c32s16s0_neon,
+            /* .packed_size   = */ kai_get_rhs_packed_size_rhs_pack_nxk_qsi4c32ps1s0scalef16_qsu4c32s16s0_neon,
+            /* .packed_stride = */ kai_get_rhs_packed_stride_rhs_pack_nxk_qsi4c32ps1s0scalef16_qsu4c32s16s0_neon,
+            /* .pack_func     = */ kai_run_rhs_pack_nxk_qsi4c32ps1s0scalef16_qsu4c32s16s0_neon,
+            /* .to_float      = */ dequantize_row_qsi4c32ps1s0scalef16,
         },
         /* .required_cpu       = */ CPU_FEATURE_SME,
         /* .lhs_type           = */ GGML_TYPE_F32,
@@ -107,8 +194,10 @@ static ggml_kleidiai_kernels gemm_gemv_kernels[] = {
             /* .pack_func             = */ kai_run_lhs_pack_bf16p2vlx2_f32_sme,
         },
         /* .rhs_info = */ {
-            /* .packed_size = */ kai_get_rhs_packed_size_rhs_pack_kxn_bf16p2vlx2b_f32_x32_sme,
-            /* .pack_func   = */ kai_run_rhs_pack_kxn_bf16p2vlx2b_f32_x32_sme,
+            /* .packed_size   = */ kai_get_rhs_packed_size_rhs_pack_kxn_bf16p2vlx2b_f32_x32_sme,
+            /* .packed_stride = */ NULL,
+            /* .pack_func     = */ kai_run_rhs_pack_kxn_bf16p2vlx2b_f32_x32_sme,
+            /* .to_float      = */ NULL,
         },
         /* .required_cpu       = */ CPU_FEATURE_SME,
         /* .lhs_type           = */ GGML_TYPE_F32,
@@ -154,8 +243,10 @@ static ggml_kleidiai_kernels gemm_gemv_kernels[] = {
             /* .pack_func             = */ kai_run_lhs_quant_pack_qsi8d32p_f32,
         },
         /* .rhs_info = */ {
-            /* .packed_size = */ kai_get_rhs_packed_size_rhs_pack_nxk_qsi4c32pscalef16_qsu4c32s16s0,
-            /* .pack_func   = */ kai_run_rhs_pack_nxk_qsi4c32pscalef16_qsu4c32s16s0,
+            /* .packed_size   = */ kai_get_rhs_packed_size_rhs_pack_nxk_qsi4c32pscalef16_qsu4c32s16s0,
+            /* .packed_stride = */ kai_get_rhs_packed_stride_rhs_pack_nxk_qsi4c32pscalef16_qsu4c32s16s0,
+            /* .pack_func     = */ kai_run_rhs_pack_nxk_qsi4c32pscalef16_qsu4c32s16s0,
+            /* .to_float      = */ dequantize_row_qsi4c32pscalef16,
         },
         /* .required_cpu       = */ CPU_FEATURE_DOTPROD,
         /* .lhs_type           = */ GGML_TYPE_F32,
@@ -200,8 +291,10 @@ static ggml_kleidiai_kernels gemm_gemv_kernels[] = {
             /* .pack_func             = */ kai_run_lhs_quant_pack_qsi8d32p_f32,
         },
         /* .rhs_info = */ {
-            /* .packed_size = */ kai_get_rhs_packed_size_rhs_pack_nxk_qsi4c32pscalef16_qsu4c32s16s0,
-            /* .pack_func   = */ kai_run_rhs_pack_nxk_qsi4c32pscalef16_qsu4c32s16s0,
+            /* .packed_size   = */ kai_get_rhs_packed_size_rhs_pack_nxk_qsi4c32pscalef16_qsu4c32s16s0,
+            /* .packed_stride = */ kai_get_rhs_packed_stride_rhs_pack_nxk_qsi4c32pscalef16_qsu4c32s16s0,
+            /* .pack_func     = */ kai_run_rhs_pack_nxk_qsi4c32pscalef16_qsu4c32s16s0,
+            /* .to_float      = */ dequantize_row_qsi4c32pscalef16,
         },
         /* .required_cpu       = */ CPU_FEATURE_DOTPROD | CPU_FEATURE_I8MM,
         /* .lhs_type           = */ GGML_TYPE_F32,
@@ -247,8 +340,10 @@ static ggml_kleidiai_kernels gemm_gemv_kernels[] = {
             /* .pack_func             = */ kai_run_lhs_quant_pack_qsi8d32p_f32,
         },
         /* .rhs_info = */ {
-            /* .packed_size = */ kai_get_rhs_packed_size_rhs_pack_nxk_qsi4c32pscalef16_qsu4c32s16s0,
-            /* .pack_func   = */ kai_run_rhs_pack_nxk_qsi4c32pscalef16_qsu4c32s16s0,
+            /* .packed_size   = */ kai_get_rhs_packed_size_rhs_pack_nxk_qsi4c32pscalef16_qsu4c32s16s0,
+            /* .packed_stride = */ kai_get_rhs_packed_stride_rhs_pack_nxk_qsi4c32pscalef16_qsu4c32s16s0,
+            /* .pack_func     = */ kai_run_rhs_pack_nxk_qsi4c32pscalef16_qsu4c32s16s0,
+            /* .to_float      = */ dequantize_row_qsi4c32pscalef16,
         },
         /* .required_cpu       = */ CPU_FEATURE_DOTPROD | CPU_FEATURE_I8MM,
         /* .lhs_type           = */ GGML_TYPE_F32,
@@ -293,8 +388,10 @@ static ggml_kleidiai_kernels gemm_gemv_kernels[] = {
             /* .pack_func             = */ kai_run_lhs_quant_pack_qsi8d32p_f32,
         },
         /* .rhs_info = */ {
-            /* .packed_size = */ kai_get_rhs_packed_size_rhs_pack_nxk_qsi4c32pscalef16_qsu4c32s16s0,
-            /* .pack_func   = */ kai_run_rhs_pack_nxk_qsi4c32pscalef16_qsu4c32s16s0,
+            /* .packed_size   = */ kai_get_rhs_packed_size_rhs_pack_nxk_qsi4c32pscalef16_qsu4c32s16s0,
+            /* .packed_stride = */ kai_get_rhs_packed_stride_rhs_pack_nxk_qsi4c32pscalef16_qsu4c32s16s0,
+            /* .pack_func     = */ kai_run_rhs_pack_nxk_qsi4c32pscalef16_qsu4c32s16s0,
+            /* .to_float      = */ dequantize_row_qsi4c32pscalef16,
         },
         /* .required_cpu       = */ CPU_FEATURE_DOTPROD,
         /* .lhs_type           = */ GGML_TYPE_F32,

ggml/src/ggml-cpu/kleidiai/kernels.h

@@ -71,12 +71,15 @@ struct rhs_packing_info {
         std::function<size_t(size_t n, size_t k, size_t nr, size_t kr, size_t bl)>,
         std::function<size_t(size_t n, size_t k)>
     > packed_size;
+    size_t (*packed_stride)(size_t k, size_t nr, size_t kr, size_t bl);
     std::variant<
         std::function<void(size_t num_groups, size_t n, size_t k, size_t nr, size_t kr, size_t sr, size_t bl, const uint8_t* rhs,
             const float* bias, void* rhs_packed, size_t extra_bytes, const struct kai_rhs_pack_qs4cxs1s0_param* params)>,
         std::function<void(size_t num_groups, size_t n, size_t k, size_t nr, size_t kr, size_t sr, size_t rhs_stride, const void* rhs,
             const void* bias, const void* scale, void* rhs_packed, size_t extra_bytes, const void* params)>
     > pack_func;
+    void (*to_float)(const void *packed_data, int32_t row_idx, int64_t nc, float *out, size_t nr_pack, size_t packed_row_stride,
+          size_t kr, size_t bl, size_t num_bytes_multiplier);
 };
 
 struct ggml_kleidiai_kernels {

ggml/src/ggml-cpu/kleidiai/kleidiai.cpp

@@ -40,6 +40,17 @@ struct ggml_kleidiai_context {
     ggml_kleidiai_kernels * kernels;
 } static ctx = { CPU_FEATURE_NONE, NULL };
 
+static const char* cpu_feature_to_string(cpu_feature f) {
+    switch (f) {
+        case CPU_FEATURE_NONE:    return "NONE";
+        case CPU_FEATURE_DOTPROD: return "DOTPROD";
+        case CPU_FEATURE_I8MM:    return "I8MM";
+        case CPU_FEATURE_SVE:     return "SVE";
+        case CPU_FEATURE_SME:     return "SME";
+        default:                  return "UNKNOWN";
+    }
+}
+
 static void init_kleidiai_context(void) {
 
     ggml_critical_section_start();
@@ -62,6 +73,11 @@ static void init_kleidiai_context(void) {
             ctx.features |= ggml_cpu_has_sme() ? CPU_FEATURE_SME : CPU_FEATURE_NONE;
         }
         ctx.kernels = ggml_kleidiai_select_kernels_q4_0(ctx.features);
+#ifndef NDEBUG
+        if (ctx.kernels) {
+            GGML_LOG_DEBUG("kleidiai: using kernel with CPU feature %s\n", cpu_feature_to_string(ctx.kernels->required_cpu));
+        }
+#endif
     }
     ggml_critical_section_end();
 }
@@ -102,6 +118,9 @@ static void transpose_f32kxn_f16nxk(size_t n, size_t k, float * dst, const uint1
 
 class tensor_traits : public ggml::cpu::tensor_traits {
     bool work_size(int /* n_threads */, const struct ggml_tensor * op, size_t & size) override {
+        if (op->op != GGML_OP_MUL_MAT) {
+            return false;
+        }
         ggml_kleidiai_kernels *kernels = ggml_kleidiai_select_kernels(ctx.features, op);
         GGML_ASSERT(kernels);
         kernel_info * kernel = op->src[1]->ne[1] == 1 ? &kernels->gemv : &kernels->gemm;
@@ -135,6 +154,10 @@ class tensor_traits : public ggml::cpu::tensor_traits {
             } else if (dst->src[0]->type == GGML_TYPE_F16) {
                 return compute_forward_kv_cache(params, dst);
             }
+        } else if (dst->op == GGML_OP_GET_ROWS) {
+            if (dst->src[0]->type == GGML_TYPE_Q4_0) {
+                return compute_forward_get_rows(params, dst);
+            }
         }
         return false;
     }
@@ -270,6 +293,8 @@ class tensor_traits : public ggml::cpu::tensor_traits {
     }
 
     bool compute_forward_q4_0(struct ggml_compute_params * params, struct ggml_tensor * dst) {
+        GGML_ASSERT(dst->src[0]->type == GGML_TYPE_Q4_0);
+
         const ggml_tensor * src0 = dst->src[0];
         const ggml_tensor * src1 = dst->src[1];
 
@@ -342,8 +367,49 @@ class tensor_traits : public ggml::cpu::tensor_traits {
         return true;
     }
 
+    bool compute_forward_get_rows(struct ggml_compute_params * params, struct ggml_tensor * dst) {
+        GGML_ASSERT(dst->src[0]->type == GGML_TYPE_Q4_0);
+        GGML_ASSERT(ctx.kernels);
+
+        const ggml_tensor * src0 = dst->src[0];
+        const ggml_tensor * src1 = dst->src[1];
+
+        GGML_TENSOR_BINARY_OP_LOCALS
+
+        rhs_packing_info * rhs_info = &ctx.kernels->rhs_info;
+        kernel_info * kernel        = &ctx.kernels->gemm;
+
+        const int64_t nc     = ne00;
+        const int64_t nr     = ggml_nelements(src1);
+
+        const size_t block_rows = kernel->get_nr();
+        const size_t kr         = kernel->get_kr();
+
+        const size_t num_bytes_multiplier = sizeof(uint16_t);
+        const size_t packed_stride = rhs_info->packed_stride(nc, block_rows, kr, QK4_0);
+
+        const int ith = params->ith;
+        const int nth = params->nth;
+
+        const int dr = (nr + nth - 1) / nth;
+        const int ir0 = dr * ith;
+        const int ir1 = MIN(ir0 + dr, nr);
+
+        for (int64_t i = ir0; i < ir1; ++i) {
+            GGML_ASSERT(src1->type == GGML_TYPE_I32);
+            int64_t row_idx = ((const int32_t *)src1->data)[i];
+            GGML_ASSERT(row_idx >= 0 && row_idx < src0->ne[1]);
+
+            float *out = (float *)((char *)dst->data + i * nb1);
+            rhs_info->to_float(src0->data, row_idx, nc, out, block_rows, packed_stride, kr, QK4_0, num_bytes_multiplier);
+        }
+
+        return true;
+    }
+
 public:
     int repack(struct ggml_tensor * tensor, const void * data, size_t data_size) {
+        GGML_ASSERT(tensor->type == GGML_TYPE_Q4_0);
         GGML_ASSERT(ctx.kernels);
         const size_t n = tensor->ne[1];
         const size_t k = tensor->ne[0];
@@ -351,17 +417,12 @@ public:
         size_t kr      = ctx.kernels->gemm.get_kr();
         size_t sr      = ctx.kernels->gemm.get_sr();
 
-#ifndef NDEBUG
-        const size_t repacked_size = variant_call<size_t>(ctx.kernels->rhs_info.packed_size, n, k, nr, kr, QK4_0);
-        GGML_ASSERT(repacked_size <= data_size && "repacked size larger than the packed size!");
-#endif
         struct kai_rhs_pack_qs4cxs1s0_param params;
         params.lhs_zero_point = 1;
         params.rhs_zero_point = 8;
         variant_call<void>(ctx.kernels->rhs_info.pack_func, 1, n, k, nr, kr, sr, QK4_0, (const uint8_t*)data, nullptr, tensor->data, 0, &params);
 
         return 0;
-
         GGML_UNUSED(data_size);
     }
 };
@@ -375,8 +436,8 @@ static ggml::cpu::tensor_traits * get_tensor_traits(ggml_backend_buffer_t, struc
 static enum ggml_status ggml_backend_cpu_kleidiai_buffer_init_tensor(ggml_backend_buffer_t buffer, struct ggml_tensor * tensor) {
     tensor->extra = (void *) ggml::cpu::kleidiai::get_tensor_traits(buffer, tensor);
 
-    GGML_UNUSED(buffer);
     return GGML_STATUS_SUCCESS;
+    GGML_UNUSED(buffer);
 }
 
 static void ggml_backend_cpu_kleidiai_buffer_set_tensor(ggml_backend_buffer_t buffer, struct ggml_tensor * tensor,
@@ -418,18 +479,35 @@ static size_t ggml_backend_cpu_kleidiai_buffer_type_get_alignment(ggml_backend_b
     GGML_UNUSED(buft);
 }
 
+static size_t ggml_backend_cpu_kleidiai_buffer_type_get_alloc_size(ggml_backend_buffer_type_t buft, const struct ggml_tensor * tensor) {
+    GGML_ASSERT(tensor->type == GGML_TYPE_Q4_0);
+    GGML_ASSERT(ctx.kernels);
+
+    const size_t n  = tensor->ne[1];
+    const size_t k  = tensor->ne[0];
+    const size_t nr = ctx.kernels->gemm.get_nr();
+    const size_t kr = ctx.kernels->gemm.get_kr();
+
+    return variant_call<size_t>(ctx.kernels->rhs_info.packed_size, n, k, nr, kr, QK4_0);
+
+    GGML_UNUSED(buft);
+}
+
 namespace ggml::cpu::kleidiai {
 class extra_buffer_type : ggml::cpu::extra_buffer_type {
     bool supports_op(ggml_backend_dev_t, const struct ggml_tensor * op) override {
-        if (op->op == GGML_OP_MUL_MAT &&
+        if ((op->op == GGML_OP_MUL_MAT || op->op == GGML_OP_GET_ROWS) &&
             op->src[0]->type == GGML_TYPE_Q4_0 &&
             op->src[0]->buffer &&
             (ggml_n_dims(op->src[0]) == 2) &&
             op->src[0]->buffer->buft == ggml_backend_cpu_kleidiai_buffer_type() && ctx.kernels) {
+            if (op->op == GGML_OP_GET_ROWS && op->src[1]->ne[0] != 8) {
+                return false;
+            }
             if (op->src[1]->buffer && !ggml_backend_buft_is_host(op->src[1]->buffer->buft)) {
                 return false;
             }
-            if (op->src[1]->type == GGML_TYPE_F32 &&
+            if ((op->src[1]->type == GGML_TYPE_F32 || op->src[1]->type == GGML_TYPE_I32) &&
                 ggml_ne(op->src[1], 2) == 1 && ggml_ne(op->src[1], 3) == 1) {
                 return true;
             }
@@ -438,7 +516,7 @@ class extra_buffer_type : ggml::cpu::extra_buffer_type {
     }
 
     ggml::cpu::tensor_traits * get_tensor_traits(const struct ggml_tensor * op) override {
-        if (op->op == GGML_OP_MUL_MAT) {
+        if (op->op == GGML_OP_MUL_MAT || op->op == GGML_OP_GET_ROWS) {
             if (op->src[0]->buffer && op->src[0]->buffer->buft == ggml_backend_cpu_kleidiai_buffer_type()) {
                 return (ggml::cpu::tensor_traits *) op->src[0]->extra;
             }
@@ -469,7 +547,7 @@ ggml_backend_buffer_type_t ggml_backend_cpu_kleidiai_buffer_type(void) {
                            /* .alloc_buffer     = */ ggml_backend_cpu_kleidiai_buffer_type_alloc_buffer,
                            /* .get_alignment    = */ ggml_backend_cpu_kleidiai_buffer_type_get_alignment,
                            /* .get_max_size     = */ nullptr,  // defaults to SIZE_MAX
-                           /* .get_alloc_size   = */ nullptr,  // defaults to ggml_nbytes
+                           /* .get_alloc_size   = */ ggml_backend_cpu_kleidiai_buffer_type_get_alloc_size,
                            /* .is_host          = */ nullptr,
                            },
         /* .device  = */ ggml_backend_reg_dev_get(ggml_backend_cpu_reg(), 0),

ggml/src/ggml-cuda/CMakeLists.txt

@@ -102,12 +102,12 @@ if (CUDAToolkit_FOUND)
     if (GGML_STATIC)
         if (WIN32)
             # As of 12.3.1 CUDA Toolkit for Windows does not offer a static cublas library
-            target_link_libraries(ggml-cuda PRIVATE CUDA::cudart_static CUDA::cublas CUDA::cublasLt)
+            target_link_libraries(ggml-cuda PRIVATE CUDA::cudart_static CUDA::cublas)
         else ()
-            target_link_libraries(ggml-cuda PRIVATE  CUDA::cudart_static CUDA::cublas_static CUDA::cublasLt_static)
+            target_link_libraries(ggml-cuda PRIVATE  CUDA::cudart_static CUDA::cublas_static)
         endif()
     else()
-        target_link_libraries(ggml-cuda PRIVATE CUDA::cudart CUDA::cublas CUDA::cublasLt)
+        target_link_libraries(ggml-cuda PRIVATE CUDA::cudart CUDA::cublas)
     endif()
 
     if (GGML_CUDA_NO_VMM)

ggml/src/ggml-cuda/cpy-utils.cuh

@@ -2,24 +2,13 @@
 
 #include "ggml-common.h"
 
-static __device__ __forceinline__ void convert_f32_f32(const float * src, float * dst) {
-    *dst = *src;
-}
-
-static __device__ __forceinline__ void convert_f32_f16(const float * src, half * dst) {
-    *dst = __float2half(*src);
-}
-
-static __device__ __forceinline__ void convert_f32_bf16(const float * src, nv_bfloat16 * dst) {
-    *dst = *src;
-}
-
-static __device__ __forceinline__ void convert_f16_f16(const half * src, half * dst) {
-    *dst = *src;
-}
-
-static __device__ __forceinline__ void convert_f16_f32(const half * src, float * dst) {
-    *dst = *src;
+template<typename src_t, typename dst_t>
+static __device__ __forceinline__ void convert_flt(const src_t * src, dst_t * dst) {
+    if constexpr (std::is_same_v<src_t, dst_t>) {
+        *dst = *src;
+    } else {
+        *dst = float(*src);
+    }
 }
 
 static __device__ __forceinline__ int best_index_int8(int n, const int8_t * val, float x) {
@@ -230,22 +219,7 @@ static __device__ void cpy_blck_f32_iq4_nl(const char * cxi, char * cdsti) {
     quantize_f32_iq4_nl_block((const float *)cxi, (block_iq4_nl *)cdsti);
 }
 
-static __device__ void cpy_1_f32_f32(const char * cxi, char * cdsti) {
-    convert_f32_f32((const float *)cxi, (float *)cdsti);
-}
-
-static __device__ void cpy_1_f32_f16(const char * cxi, char * cdsti) {
-    convert_f32_f16((const float *)cxi, (half *)cdsti);
-}
-
-static __device__ void cpy_1_f32_bf16(const char * cxi, char * cdsti) {
-    convert_f32_bf16((const float *)cxi, (nv_bfloat16 *)cdsti);
-}
-
-static __device__ void cpy_1_f16_f16(const char * cxi, char * cdsti) {
-    convert_f16_f16((const half *)cxi, (half *)cdsti);
-}
-
-static __device__ void cpy_1_f16_f32(const char * cxi, char * cdsti) {
-    convert_f16_f32((const half *)cxi, (float *)cdsti);
+template<typename src_t, typename dst_t>
+static __device__ void cpy_1_flt(const char * cxi, char * cdsti) {
+    convert_flt((const src_t *)cxi, (dst_t *)cdsti);
 }

ggml/src/ggml-cuda/cpy.cu

@@ -8,10 +8,10 @@
 typedef void (*cpy_kernel_t)(const char * cx, char * cdst);
 
 template <cpy_kernel_t cpy_1>
-static __global__ void cpy_f32_f16(const char * cx, char * cdst_direct, const int ne,
-                                   const int ne00, const int ne01, const int ne02, const int nb00, const int nb01, const int nb02,
-                                   const int nb03, const int ne10, const int ne11, const int ne12, const int nb10, const int nb11,
-                                   const int nb12, const int nb13, char ** cdst_indirect, int graph_cpynode_index) {
+static __global__ void cpy_flt(const char * cx, char * cdst_direct, const int ne,
+                               const int ne00, const int ne01, const int ne02, const int nb00, const int nb01, const int nb02,
+                               const int nb03, const int ne10, const int ne11, const int ne12, const int nb10, const int nb11,
+                               const int nb12, const int nb13, char ** cdst_indirect, int graph_cpynode_index) {
     const int64_t i = blockDim.x*blockIdx.x + threadIdx.x;
 
     if (i >= ne) {
@@ -139,43 +139,14 @@ void ggml_cuda_cpy_dest_ptrs_copy(ggml_cuda_graph * cuda_graph, char ** host_des
 #endif
 }
 
-static void ggml_cpy_f16_f32_cuda(
+template<typename src_t, typename dst_t>
+static void ggml_cpy_flt_cuda(
     const char * cx, char * cdst, const int ne,
     const int ne00, const int ne01, const int ne02, const int nb00, const int nb01, const int nb02,
     const int nb03, const int ne10, const int ne11, const int ne12, const int nb10, const int nb11, const int nb12, const int nb13, cudaStream_t stream, char ** cdst_indirect, int & graph_cpynode_index) {
 
     const int num_blocks = (ne + CUDA_CPY_BLOCK_SIZE - 1) / CUDA_CPY_BLOCK_SIZE;
-    cpy_f32_f16<cpy_1_f16_f32><<<num_blocks, CUDA_CPY_BLOCK_SIZE, 0, stream>>>
-        (cx, cdst, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, cdst_indirect, graph_cpynode_index++);
-}
-
-static void ggml_cpy_f32_f32_cuda(
-    const char * cx, char * cdst, const int ne,
-    const int ne00, const int ne01, const int ne02, const int nb00, const int nb01, const int nb02,
-    const int nb03, const int ne10, const int ne11, const int ne12, const int nb10, const int nb11, const int nb12, const int nb13, cudaStream_t stream, char ** cdst_indirect, int & graph_cpynode_index) {
-
-    const int num_blocks = (ne + CUDA_CPY_BLOCK_SIZE - 1) / CUDA_CPY_BLOCK_SIZE;
-    cpy_f32_f16<cpy_1_f32_f32><<<num_blocks, CUDA_CPY_BLOCK_SIZE, 0, stream>>>
-        (cx, cdst, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, cdst_indirect, graph_cpynode_index++);
-}
-
-static void ggml_cpy_f32_bf16_cuda(
-    const char * cx, char * cdst, const int ne,
-    const int ne00, const int ne01, const int ne02, const int nb00, const int nb01, const int nb02,
-    const int nb03, const int ne10, const int ne11, const int ne12, const int nb10, const int nb11, const int nb12, const int nb13, cudaStream_t stream, char ** cdst_indirect, int & graph_cpynode_index) {
-
-    const int num_blocks = (ne + CUDA_CPY_BLOCK_SIZE - 1) / CUDA_CPY_BLOCK_SIZE;
-    cpy_f32_f16<cpy_1_f32_bf16><<<num_blocks, CUDA_CPY_BLOCK_SIZE, 0, stream>>>
-        (cx, cdst, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, cdst_indirect, graph_cpynode_index++);
-}
-
-static void ggml_cpy_f32_f16_cuda(
-    const char * cx, char * cdst, const int ne,
-    const int ne00, const int ne01, const int ne02, const int nb00, const int nb01, const int nb02,
-    const int nb03, const int ne10, const int ne11, const int ne12, const int nb10, const int nb11, const int nb12, const int nb13, cudaStream_t stream, char ** cdst_indirect, int & graph_cpynode_index) {
-
-    const int num_blocks = (ne + CUDA_CPY_BLOCK_SIZE - 1) / CUDA_CPY_BLOCK_SIZE;
-    cpy_f32_f16<cpy_1_f32_f16><<<num_blocks, CUDA_CPY_BLOCK_SIZE, 0, stream>>>
+    cpy_flt<cpy_1_flt<src_t, dst_t>><<<num_blocks, CUDA_CPY_BLOCK_SIZE, 0, stream>>>
         (cx, cdst, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, cdst_indirect, graph_cpynode_index++);
 }
 
@@ -307,16 +278,6 @@ static void ggml_cpy_f32_iq4_nl_cuda(
         (cx, cdst, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, cdst_indirect, graph_cpynode_index++);
 }
 
-static void ggml_cpy_f16_f16_cuda(
-    const char * cx, char * cdst, const int ne,
-    const int ne00, const int ne01, const int ne02, const int nb00, const int nb01, const int nb02,
-    const int nb03, const int ne10, const int ne11, const int ne12, const int nb10, const int nb11, const int nb12, const int nb13, cudaStream_t stream, char ** cdst_indirect, int & graph_cpynode_index) {
-
-    const int num_blocks = (ne + CUDA_CPY_BLOCK_SIZE - 1) / CUDA_CPY_BLOCK_SIZE;
-    cpy_f32_f16<cpy_1_f16_f16><<<num_blocks, CUDA_CPY_BLOCK_SIZE, 0, stream>>>
-        (cx, cdst, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, cdst_indirect, graph_cpynode_index++);
-}
-
 void ggml_cuda_cpy(ggml_backend_cuda_context & ctx, const ggml_tensor * src0, ggml_tensor * src1, bool disable_indirection_for_this_node) {
     const int64_t ne = ggml_nelements(src0);
     GGML_ASSERT(ne == ggml_nelements(src1));
@@ -372,11 +333,11 @@ void ggml_cuda_cpy(ggml_backend_cuda_context & ctx, const ggml_tensor * src0, gg
             CUDA_CHECK(cudaMemcpyAsync(src1_ddc, src0_ddc, ggml_nbytes(src0), cudaMemcpyDeviceToDevice, main_stream));
         }
     } else if (src0->type == GGML_TYPE_F32 && src1->type == GGML_TYPE_F32) {
-        ggml_cpy_f32_f32_cuda (src0_ddc, src1_ddc, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, main_stream, dest_ptrs_d, graph_cpynode_index);
+        ggml_cpy_flt_cuda<float, float> (src0_ddc, src1_ddc, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, main_stream, dest_ptrs_d, graph_cpynode_index);
     } else if (src0->type == GGML_TYPE_F32 && src1->type == GGML_TYPE_BF16) {
-        ggml_cpy_f32_bf16_cuda(src0_ddc, src1_ddc, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, main_stream, dest_ptrs_d, graph_cpynode_index);
+        ggml_cpy_flt_cuda<float, nv_bfloat16> (src0_ddc, src1_ddc, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, main_stream, dest_ptrs_d, graph_cpynode_index);
     } else if (src0->type == GGML_TYPE_F32 && src1->type == GGML_TYPE_F16) {
-        ggml_cpy_f32_f16_cuda (src0_ddc, src1_ddc, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, main_stream, dest_ptrs_d, graph_cpynode_index);
+        ggml_cpy_flt_cuda<float, half> (src0_ddc, src1_ddc, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, main_stream, dest_ptrs_d, graph_cpynode_index);
     } else if (src0->type == GGML_TYPE_F32 && src1->type == GGML_TYPE_Q8_0) {
         ggml_cpy_f32_q8_0_cuda(src0_ddc, src1_ddc, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, main_stream, dest_ptrs_d, graph_cpynode_index);
     } else if (src0->type == GGML_TYPE_Q8_0 && src1->type == GGML_TYPE_F32) {
@@ -403,9 +364,17 @@ void ggml_cuda_cpy(ggml_backend_cuda_context & ctx, const ggml_tensor * src0, gg
     } else if (src0->type == GGML_TYPE_Q5_1 && src1->type == GGML_TYPE_F32) {
         ggml_cpy_q5_1_f32_cuda(src0_ddc, src1_ddc, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, main_stream, dest_ptrs_d, graph_cpynode_index);
     } else if (src0->type == GGML_TYPE_F16 && src1->type == GGML_TYPE_F16) {
-        ggml_cpy_f16_f16_cuda (src0_ddc, src1_ddc, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, main_stream, dest_ptrs_d, graph_cpynode_index);
+        ggml_cpy_flt_cuda<half, half> (src0_ddc, src1_ddc, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, main_stream, dest_ptrs_d, graph_cpynode_index);
+    } else if (src0->type == GGML_TYPE_F16 && src1->type == GGML_TYPE_BF16) {
+        ggml_cpy_flt_cuda<half, nv_bfloat16> (src0_ddc, src1_ddc, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, main_stream, dest_ptrs_d, graph_cpynode_index);
     } else if (src0->type == GGML_TYPE_F16 && src1->type == GGML_TYPE_F32) {
-        ggml_cpy_f16_f32_cuda (src0_ddc, src1_ddc, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, main_stream, dest_ptrs_d, graph_cpynode_index);
+        ggml_cpy_flt_cuda<half, float> (src0_ddc, src1_ddc, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, main_stream, dest_ptrs_d, graph_cpynode_index);
+    } else if (src0->type == GGML_TYPE_BF16 && src1->type == GGML_TYPE_BF16) {
+        ggml_cpy_flt_cuda<nv_bfloat16, nv_bfloat16> (src0_ddc, src1_ddc, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, main_stream, dest_ptrs_d, graph_cpynode_index);
+    } else if (src0->type == GGML_TYPE_BF16 && src1->type == GGML_TYPE_F16) {
+        ggml_cpy_flt_cuda<nv_bfloat16, half> (src0_ddc, src1_ddc, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, main_stream, dest_ptrs_d, graph_cpynode_index);
+    } else if (src0->type == GGML_TYPE_BF16 && src1->type == GGML_TYPE_F32) {
+        ggml_cpy_flt_cuda<nv_bfloat16, float> (src0_ddc, src1_ddc, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, main_stream, dest_ptrs_d, graph_cpynode_index);
     } else {
         GGML_ABORT("%s: unsupported type combination (%s to %s)\n", __func__,
                 ggml_type_name(src0->type), ggml_type_name(src1->type));
@@ -430,11 +399,11 @@ void* ggml_cuda_cpy_fn(const ggml_tensor * src0, ggml_tensor * src1) {
     if (src0->type == src1->type && ggml_is_contiguous(src0) && ggml_is_contiguous(src1)) {
         return nullptr;
     } else if (src0->type == GGML_TYPE_F32 && src1->type == GGML_TYPE_F32) {
-        return (void*) cpy_f32_f16<cpy_1_f32_f32>;
+        return (void*) cpy_flt<cpy_1_flt<float, float>>;
     } else if (src0->type == GGML_TYPE_F32 && src1->type == GGML_TYPE_BF16) {
-        return (void*) cpy_f32_f16<cpy_1_f32_bf16>;
+        return (void*) cpy_flt<cpy_1_flt<float, nv_bfloat16>>;
     } else if (src0->type == GGML_TYPE_F32 && src1->type == GGML_TYPE_F16) {
-        return (void*) cpy_f32_f16<cpy_1_f32_f16>;
+        return (void*) cpy_flt<cpy_1_flt<float, half>>;
     } else if (src0->type == GGML_TYPE_F32 && src1->type == GGML_TYPE_Q8_0) {
         return (void*) cpy_f32_q<cpy_blck_f32_q8_0, QK8_0>;
     } else if (src0->type == GGML_TYPE_Q8_0 && src1->type == GGML_TYPE_F32) {
@@ -458,9 +427,17 @@ void* ggml_cuda_cpy_fn(const ggml_tensor * src0, ggml_tensor * src1) {
     } else if (src0->type == GGML_TYPE_Q5_1 && src1->type == GGML_TYPE_F32) {
         return (void*) cpy_q_f32<cpy_blck_q_f32<dequantize_q5_1, QK5_1>, QK5_1>;
     } else if (src0->type == GGML_TYPE_F16 && src1->type == GGML_TYPE_F16) {
-        return (void*) cpy_f32_f16<cpy_1_f32_f16>;
+        return (void*) cpy_flt<cpy_1_flt<half, half>>;
+    } else if (src0->type == GGML_TYPE_F16 && src1->type == GGML_TYPE_BF16) {
+        return (void*) cpy_flt<cpy_1_flt<half, nv_bfloat16>>;
     } else if (src0->type == GGML_TYPE_F16 && src1->type == GGML_TYPE_F32) {
-        return (void*) cpy_f32_f16<cpy_1_f16_f32>;
+        return (void*) cpy_flt<cpy_1_flt<half, float>>;
+    } else if (src0->type == GGML_TYPE_BF16 && src1->type == GGML_TYPE_F16) {
+        return (void*) cpy_flt<cpy_1_flt<nv_bfloat16, half>>;
+    } else if (src0->type == GGML_TYPE_BF16 && src1->type == GGML_TYPE_BF16) {
+        return (void*) cpy_flt<cpy_1_flt<nv_bfloat16, nv_bfloat16>>;
+    } else if (src0->type == GGML_TYPE_BF16 && src1->type == GGML_TYPE_F32) {
+        return (void*) cpy_flt<cpy_1_flt<nv_bfloat16, float>>;
     } else {
         GGML_ABORT("%s: unsupported type combination (%s to %s)\n", __func__,
                 ggml_type_name(src0->type), ggml_type_name(src1->type));

ggml/src/ggml-cuda/ggml-cuda.cu

@@ -55,6 +55,7 @@
 #include <cstddef>
 #include <cstdint>
 #include <float.h>
+#include <initializer_list>
 #include <limits>
 #include <map>
 #include <memory>
@@ -2765,6 +2766,39 @@ static void update_cuda_graph_executable(ggml_backend_cuda_context * cuda_ctx) {
 }
 #endif
 
+static bool ggml_cuda_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;
+    }
+
+    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];
+
+        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;
+        }
+
+        //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;
+        }
+
+        //rms_norm kernel assumes contigous rows
+        if (!ggml_is_contiguous_rows(mul->src[0]) || !ggml_is_contiguous_rows(mul->src[1])) {
+            return false;
+        }
+    }
+
+    return true;
+}
+
 static void evaluate_and_capture_cuda_graph(ggml_backend_cuda_context * cuda_ctx, ggml_cgraph * cgraph,
     bool & graph_evaluated_or_captured, bool & use_cuda_graph, bool & cuda_graph_update_required) {
     // flag used to determine whether it is an integrated_gpu
@@ -2774,6 +2808,7 @@ static void evaluate_and_capture_cuda_graph(ggml_backend_cuda_context * cuda_ctx
         // Only perform the graph execution if CUDA graphs are not enabled, or we are capturing the graph.
         // With the use of CUDA graphs, the execution will be performed by the graph launch.
         if (!use_cuda_graph || cuda_graph_update_required) {
+
             for (int i = 0; i < cgraph->n_nodes; i++) {
                 ggml_tensor * node = cgraph->nodes[i];
 
@@ -2781,6 +2816,12 @@ static void evaluate_and_capture_cuda_graph(ggml_backend_cuda_context * cuda_ctx
                     continue;
                 }
 
+                static bool disable_fusion = (getenv("GGML_CUDA_DISABLE_FUSION") != nullptr);
+                if (!disable_fusion && ggml_cuda_can_fuse(cgraph, i, { GGML_OP_RMS_NORM, GGML_OP_MUL })) {
+                    ggml_cuda_op_rms_norm_fused(*cuda_ctx, node, cgraph->nodes[i+1]);
+                    i++;
+                    continue;
+                }
 #ifndef NDEBUG
                 assert(node->buffer->buft == ggml_backend_cuda_buffer_type(cuda_ctx->device));
                 for (int j = 0; j < GGML_MAX_SRC; j++) {
@@ -3242,13 +3283,9 @@ static bool ggml_backend_cuda_device_supports_op(ggml_backend_dev_t dev, const g
             {
                 ggml_type src0_type = op->src[0]->type;
                 ggml_type src1_type = op->src[1]->type;
-                if (src0_type == GGML_TYPE_F32 && src1_type == GGML_TYPE_F32) {
-                    return true;
-                }
-                if (src0_type == GGML_TYPE_F32 && src1_type == GGML_TYPE_BF16) {
-                    return true;
-                }
-                if (src0_type == GGML_TYPE_F32 && src1_type == GGML_TYPE_F16) {
+                if ((src0_type == GGML_TYPE_F32 || src0_type == GGML_TYPE_BF16 || src0_type == GGML_TYPE_F16) &&
+                    (src1_type == GGML_TYPE_F32 || src1_type == GGML_TYPE_BF16 || src1_type == GGML_TYPE_F16)
+                ) {
                     return true;
                 }
                 if (src0_type == GGML_TYPE_F32 && src1_type == GGML_TYPE_Q8_0) {
@@ -3284,12 +3321,6 @@ static bool ggml_backend_cuda_device_supports_op(ggml_backend_dev_t dev, const g
                 if (src0_type == GGML_TYPE_F32 && src1_type == GGML_TYPE_IQ4_NL) {
                     return true;
                 }
-                if (src0_type == GGML_TYPE_F16 && src1_type == GGML_TYPE_F16) {
-                    return true;
-                }
-                if (src0_type == GGML_TYPE_F16 && src1_type == GGML_TYPE_F32) {
-                    return true;
-                }
                 if (src0_type == src1_type && ggml_is_contiguous(op->src[0]) && ggml_is_contiguous(op->src[1])) {
                     return true;
                 }
@@ -3370,7 +3401,7 @@ static bool ggml_backend_cuda_device_supports_op(ggml_backend_dev_t dev, const g
             return op->src[0]->ne[1] % 128 == 0;
         }
         case GGML_OP_CONT:
-            return op->src[0]->type != GGML_TYPE_BF16;
+            return true;
         case GGML_OP_DIAG_MASK_INF:
             return true;
         case GGML_OP_SOFT_MAX:

ggml/src/ggml-cuda/im2col.cu

@@ -10,7 +10,7 @@ static  __global__ void im2col_kernel(
         return;
     }
 
-    const int64_t  ksize = OW * (KH > 1 ? KW : 1);
+    const int64_t  ksize = OW * KH;
     const int64_t  kx = i / ksize;
     const int64_t  kd = kx * ksize;
     const int64_t  ky = (i - kd) / OW;

ggml/src/ggml-cuda/norm.cu

@@ -104,10 +104,12 @@ static __global__ void group_norm_f32(const float * x, float * dst, const int gr
     }
 }
 
-template <int block_size>
+template <int block_size, bool do_multiply = false>
 static __global__ void rms_norm_f32(
         const float * x, float * dst, const int ncols, const int64_t stride_row, const int64_t stride_channel,
-        const int64_t stride_sample, const float eps) {
+        const int64_t stride_sample, const float eps, const float * mul = nullptr, const int64_t mul_stride_row = 0,
+        const int64_t mul_stride_channel = 0, const int64_t mul_stride_sample = 0, const int mul_ncols = 0,
+        const int mul_nrows = 0, const int mul_nchannels = 0, const int mul_nsamples = 0) {
     const int nrows     = gridDim.x;
     const int nchannels = gridDim.y;
 
@@ -119,6 +121,13 @@ static __global__ void rms_norm_f32(
     x   += sample*stride_sample + channel*stride_channel + row*stride_row;
     dst += ((sample*nchannels + channel)*nrows + row)*ncols;
 
+    if constexpr (do_multiply) {
+        const int mul_row = row % mul_nrows;
+        const int mul_channel = channel % mul_nchannels;
+        const int mul_sample = sample % mul_nsamples;
+        mul += mul_sample*mul_stride_sample + mul_channel*mul_stride_channel + mul_row*mul_stride_row;
+    }
+
     float tmp = 0.0f; // partial sum for thread in warp
 
     for (int col = tid; col < ncols; col += block_size) {
@@ -145,7 +154,12 @@ static __global__ void rms_norm_f32(
     const float scale = rsqrtf(mean + eps);
 
     for (int col = tid; col < ncols; col += block_size) {
-        dst[col] = scale * x[col];
+        if constexpr (do_multiply) {
+            const int mul_col = col % mul_ncols;
+            dst[col] = scale * x[col] * mul[mul_col];
+        } else {
+            dst[col] = scale * x[col];
+        }
     }
 }
 
@@ -310,10 +324,30 @@ static void rms_norm_f32_cuda(
     const dim3 blocks_num(nrows, nchannels, nsamples);
     if (ncols < 1024) {
         const dim3 block_dims(WARP_SIZE, 1, 1);
-        rms_norm_f32<WARP_SIZE><<<blocks_num, block_dims, 0, stream>>>(x, dst, ncols, stride_row, stride_channel, stride_sample, eps);
+        rms_norm_f32<WARP_SIZE, false><<<blocks_num, block_dims, 0, stream>>>(x, dst, ncols, stride_row, stride_channel, stride_sample, eps);
     } else {
         const dim3 block_dims(1024, 1, 1);
-        rms_norm_f32<1024><<<blocks_num, block_dims, 0, stream>>>(x, dst, ncols, stride_row, stride_channel, stride_sample, eps);
+        rms_norm_f32<1024, false><<<blocks_num, block_dims, 0, stream>>>(x, dst, ncols, stride_row, stride_channel, stride_sample, eps);
+    }
+}
+
+static void rms_norm_mul_f32_cuda(
+        const float * x, const float * mul, float * dst, const int ncols, const int nrows, const int nchannels, const int nsamples,
+        const int64_t stride_row, const int64_t stride_channel, const int64_t stride_sample,
+        const int64_t mul_stride_row, const int64_t mul_stride_channel, const int64_t mul_stride_sample,
+        const int mul_ncols, const int mul_nrows, const int mul_nchannels, const int mul_nsamples,
+        const float eps, cudaStream_t stream) {
+    const dim3 blocks_num(nrows, nchannels, nsamples);
+    if (mul == nullptr) {
+        rms_norm_f32_cuda(x, dst, ncols, nrows, nchannels, nsamples, stride_row, stride_channel, stride_sample, eps, stream);
+        return;
+    }
+    if (ncols < 1024) {
+        const dim3 block_dims(WARP_SIZE, 1, 1);
+        rms_norm_f32<WARP_SIZE, true><<<blocks_num, block_dims, 0, stream>>>(x, dst, ncols, stride_row, stride_channel, stride_sample, eps, mul, mul_stride_row, mul_stride_channel, mul_stride_sample, mul_ncols, mul_nrows, mul_nchannels, mul_nsamples);
+    } else {
+        const dim3 block_dims(1024, 1, 1);
+        rms_norm_f32<1024, true><<<blocks_num, block_dims, 0, stream>>>(x, dst, ncols, stride_row, stride_channel, stride_sample, eps, mul, mul_stride_row, mul_stride_channel, mul_stride_sample, mul_ncols, mul_nrows, mul_nchannels, mul_nsamples);
     }
 }
 
@@ -407,6 +441,59 @@ void ggml_cuda_op_rms_norm(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
     rms_norm_f32_cuda(src0_d, dst_d, ne00, ne01, ne02, ne03, s01, s02, s03, eps, stream);
 }
 
+void ggml_cuda_op_rms_norm_fused(ggml_backend_cuda_context & ctx, ggml_tensor * dst, ggml_tensor * mul_tensor) {
+    const ggml_tensor * rms_norm_src = (ggml_tensor *) dst->src[0];
+    float eps = 0.0f;
+
+    memcpy(&eps, dst->op_params, sizeof(float));
+
+    const float * src0_d = (const float *) rms_norm_src->data;
+    const float * mul_d = nullptr;
+    const ggml_tensor * mul_src = nullptr;
+
+    if (mul_tensor->src[0] == dst) {
+        mul_d = (float *) mul_tensor->src[1]->data;
+        mul_src = mul_tensor->src[1];
+    } else if(mul_tensor->src[1] == dst) {
+        mul_d = (float *) mul_tensor->src[0]->data;
+        mul_src = mul_tensor->src[0];
+    } else {
+        GGML_ASSERT(false);
+    }
+
+    float * dst_d = (float *) mul_tensor->data;
+    cudaStream_t stream = ctx.stream();
+
+    GGML_ASSERT(rms_norm_src->type == GGML_TYPE_F32);
+    GGML_ASSERT(dst->type == GGML_TYPE_F32);
+    GGML_ASSERT(mul_tensor->type == GGML_TYPE_F32);
+    GGML_ASSERT(eps >= 0.0f);
+
+    const int64_t ne00 = rms_norm_src->ne[0];
+    const int64_t ne01 = rms_norm_src->ne[1];
+    const int64_t ne02 = rms_norm_src->ne[2];
+    const int64_t ne03 = rms_norm_src->ne[3];
+
+    const size_t ts0 = ggml_type_size(rms_norm_src->type);
+    GGML_ASSERT(rms_norm_src->nb[0] == ts0);
+    const int64_t s01 = rms_norm_src->nb[1] / ts0;
+    const int64_t s02 = rms_norm_src->nb[2] / ts0;
+    const int64_t s03 = rms_norm_src->nb[3] / ts0;
+
+    const size_t ts_mul = ggml_type_size(mul_src->type);
+    GGML_ASSERT(mul_src->nb[0] == ts_mul);
+    const int64_t mul_s01 = mul_src->nb[1] / ts_mul;
+    const int64_t mul_s02 = mul_src->nb[2] / ts_mul;
+    const int64_t mul_s03 = mul_src->nb[3] / ts_mul;
+
+    const int mul_ncols     = mul_src->ne[0];
+    const int mul_nrows     = mul_src->ne[1];
+    const int mul_nchannels = mul_src->ne[2];
+    const int mul_nsamples  = mul_src->ne[3];
+
+    rms_norm_mul_f32_cuda(src0_d, mul_d, dst_d, ne00, ne01, ne02, ne03, s01, s02, s03, mul_s01, mul_s02, mul_s03, mul_ncols, mul_nrows, mul_nchannels, mul_nsamples, eps, stream);
+}
+
 void ggml_cuda_op_rms_norm_back(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
     const ggml_tensor * grad  = dst->src[0]; // gradients
     const ggml_tensor * src0f = dst->src[1]; // src0 from forward pass

ggml/src/ggml-cuda/norm.cuh

@@ -6,6 +6,8 @@ void ggml_cuda_op_group_norm(ggml_backend_cuda_context & ctx, ggml_tensor * dst)
 
 void ggml_cuda_op_rms_norm(ggml_backend_cuda_context & ctx, ggml_tensor * dst);
 
+void ggml_cuda_op_rms_norm_fused(ggml_backend_cuda_context & ctx, ggml_tensor * dst, ggml_tensor * mul_tensor);
+
 void ggml_cuda_op_rms_norm_back(ggml_backend_cuda_context & ctx, ggml_tensor * dst);
 
 void ggml_cuda_op_l2_norm(ggml_backend_cuda_context & ctx, ggml_tensor * dst);

ggml/src/ggml-cuda/set-rows.cu

@@ -4,24 +4,8 @@
 typedef void (*set_rows_kernel_t)(const char * src, char * dst);
 
 template<typename src_t, typename dst_t>
-__device__ void set_rows_1(const src_t * src_f, dst_t * dst_f) {
-    GGML_UNUSED(src_f);
-    GGML_UNUSED(dst_f);
-}
-
-template<>
-__device__ __forceinline__ void set_rows_1<float, half>(const float * src_f, half * dst_h) {
-    convert_f32_f16(src_f, dst_h);
-}
-
-template<>
-__device__ __forceinline__ void set_rows_1<float, nv_bfloat16>(const float * src_f, nv_bfloat16 * dst_b) {
-    convert_f32_bf16(src_f, dst_b);
-}
-
-template<>
-__device__ __forceinline__ void set_rows_1<float, float>(const float * src_f, float * dst_f) {
-    convert_f32_f32(src_f, dst_f);
+__device__ __forceinline__ void set_rows_1(const src_t * src_f, dst_t * dst_f) {
+    convert_flt(src_f, dst_f);
 }
 
 // Generic quantized set_rows kernel template

ggml/src/ggml-opencl/CMakeLists.txt

@@ -105,6 +105,8 @@ set(GGML_OPENCL_KERNELS
     pad
     repeat
     mul_mat_f16_f32
+    conv2d
+    conv2d_f16_f32
 )
 
 foreach (K ${GGML_OPENCL_KERNELS})

ggml/src/ggml-opencl/ggml-opencl.cpp

@@ -390,6 +390,9 @@ 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;
 
@@ -441,6 +444,9 @@ 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;
 
@@ -1478,6 +1484,47 @@ static void load_cl_kernels(ggml_backend_opencl_context *backend_ctx, ggml_cl_ve
         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
@@ -2361,6 +2408,10 @@ static bool ggml_opencl_supports_op(ggml_backend_dev_t dev, const struct ggml_te
                    op->src[0]->ne[3] == 1 && op->ne[3] == 1;
         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:
@@ -4998,6 +5049,82 @@ static void ggml_cl_mul_mat_f16_f32_tiled(ggml_backend_t backend, const ggml_ten
     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);
@@ -6752,6 +6879,12 @@ 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;

ggml/src/ggml-opencl/kernels/conv2d.cl

@@ -0,0 +1,185 @@
+#ifdef USE_FP16
+#pragma OPENCL EXTENSION cl_khr_fp16 : enable
+#define T_FLOAT half
+#define T_FLOAT4 half4
+#define VSTORE_T_FLOAT4(data, offset, p) vstore_half4_rte(data, offset, p)
+#else
+#define T_FLOAT float
+#define T_FLOAT4 float4
+#define VSTORE_T_FLOAT4(data, offset, p) vstore4(data, offset, p)
+#endif
+
+#if defined(cl_qcom_reqd_sub_group_size)
+#pragma OPENCL EXTENSION cl_qcom_reqd_sub_group_size : enable
+#define REQD_SUBGROUP_SIZE_128 __attribute__((qcom_reqd_sub_group_size("full")))
+#else
+#define REQD_SUBGROUP_SIZE_128
+#endif
+
+#define T_ACCUM float4
+#define VEC_SIZE 4
+
+#define BS_K 64
+#define BS_NPQ 64
+#define BS_CRS 16
+
+#define TS_K 4
+#define TS_NPQ 8
+
+#define WG_K (BS_K / TS_K)
+#define WG_NPQ (BS_NPQ / TS_NPQ)
+
+#define BS_NPQ_VEC (BS_NPQ / VEC_SIZE)
+#define TS_NPQ_VEC (TS_NPQ / VEC_SIZE)
+
+static inline uint splitWork(uint work_size, uint block_size){
+    return (work_size + block_size - 1) / block_size;
+}
+
+REQD_SUBGROUP_SIZE_128
+kernel void kernel_conv_2d(
+    global void* p_knl,
+    ulong off_knl,
+    global void* p_src,
+    ulong off_src,
+    global void* p_dst,
+    ulong off_dst,
+    local void* shared,
+    uint Cout, uint Cin, uint N,
+    uint KW, uint KH, uint W, uint H, uint OW, uint OH,
+    uint s0, uint s1, uint p0, uint p1, uint d0, uint d1,
+    uint nb01, uint nb02, uint nb03,
+    uint nb11, uint nb12, uint nb13,
+    uint nb1, uint nb2, uint nb3
+) {
+    global T_FLOAT* knl_data = (global T_FLOAT*) ((global char*)p_knl + off_knl);
+    global T_FLOAT* src_data = (global T_FLOAT*) ((global char*)p_src + off_src);
+    global T_FLOAT* dst_data = (global T_FLOAT*) ((global char*)p_dst + off_dst);
+
+    const uint K = Cout;
+    const uint CRS = Cin*KH*KW;
+    const uint NPQ = N*OH*OW;
+
+    const uint lid_k = get_local_id(0);
+    const uint lid_npq = get_local_id(1);
+    const uint tid = lid_npq * WG_K + lid_k;
+
+    const uint B_idx_K = get_group_id(0);
+    const uint B_idx_NPQ = get_group_id(1);
+
+    const uint offset_k = B_idx_K * BS_K;
+    const uint offset_npq = B_idx_NPQ * BS_NPQ;
+
+    local T_FLOAT* Ash = (local T_FLOAT*)shared;
+    local T_FLOAT4* Bsh = (local T_FLOAT4*) &Ash[BS_K * BS_CRS];
+
+    T_ACCUM regC[TS_K][TS_NPQ_VEC];
+    for (int i = 0; i < TS_K; ++i) {
+        for (int j = 0; j < TS_NPQ_VEC; ++j) {
+            regC[i][j] = (T_ACCUM)(0.0f);
+        }
+    }
+
+    const uint NB_CRS = splitWork(CRS, BS_CRS);
+
+    for (uint B_idx_CRS = 0; B_idx_CRS < NB_CRS; ++B_idx_CRS) {
+        const uint offset_crs = B_idx_CRS * BS_CRS;
+
+        for (int i = tid; i < BS_K * BS_CRS; i += (WG_K * WG_NPQ)) {
+            const uint k_l = i / BS_CRS;
+            const uint crs_l = i % BS_CRS;
+            const uint k_g = offset_k + k_l;
+            const uint crs_g = offset_crs + crs_l;
+
+            if (k_g < K && crs_g < CRS) {
+                const uint Cin_idx = crs_g / (KW*KH);
+                const uint KH_idx = (crs_g - Cin_idx*KW*KH) / KW;
+                const uint KW_idx = crs_g - Cin_idx*KW*KH - KH_idx*KW;
+                const uint knl_idx = KW_idx + KH_idx*nb01 + Cin_idx*nb02 + k_g*nb03;
+                Ash[k_l * BS_CRS + crs_l] = knl_data[knl_idx];
+            } else {
+                Ash[k_l * BS_CRS + crs_l] = (T_FLOAT)0.0f;
+            }
+        }
+
+        for (int i = tid; i < BS_CRS * BS_NPQ_VEC; i += (WG_K * WG_NPQ)) {
+            const uint crs_l = i / BS_NPQ_VEC;
+            const uint npq_l_vec = i % BS_NPQ_VEC;
+            const uint crs_g = offset_crs + crs_l;
+
+            T_FLOAT4 val = (T_FLOAT4)(0.0f);
+            if (crs_g < CRS) {
+                const uint Cin_idx = crs_g / (KW * KH);
+                const uint KH_idx = (crs_g - Cin_idx * KW * KH) / KW;
+                const uint KW_idx = crs_g - Cin_idx * KW * KH - KH_idx * KW;
+                for (int v = 0; v < VEC_SIZE; ++v) {
+                    const uint npq_g = offset_npq + npq_l_vec * VEC_SIZE + v;
+                    if (npq_g < NPQ) {
+                        const uint N_idx = npq_g / (OH * OW);
+                        const uint pq_idx = npq_g % (OH * OW);
+                        const uint OH_idx = pq_idx / OW;
+                        const uint OW_idx = pq_idx % OW;
+                        const int H_idx = (int)(OH_idx * s1 + KH_idx * d1 - p1);
+                        const int W_idx = (int)(OW_idx * s0 + KW_idx * d0 - p0);
+
+                        if (H_idx >= 0 && H_idx < H && W_idx >= 0 && W_idx < W) {
+                            const uint src_idx = W_idx + H_idx * nb11 + Cin_idx * nb12 + N_idx * nb13;
+                            ((T_FLOAT*)&val)[v] = src_data[src_idx];
+                        }
+                    }
+                }
+            }
+            Bsh[crs_l * BS_NPQ_VEC + npq_l_vec] = val;
+        }
+
+        barrier(CLK_LOCAL_MEM_FENCE);
+
+        #pragma unroll
+        for (uint crs_l = 0; crs_l < BS_CRS; ++crs_l) {
+            T_FLOAT regA[TS_K];
+            for (uint k_l_reg = 0; k_l_reg < TS_K; ++k_l_reg) {
+                regA[k_l_reg] = Ash[(lid_k * TS_K + k_l_reg) * BS_CRS + crs_l];
+            }
+
+            for (uint npq_l_vec_reg = 0; npq_l_vec_reg < TS_NPQ_VEC; ++npq_l_vec_reg) {
+                T_FLOAT4 regB = Bsh[crs_l * BS_NPQ_VEC + lid_npq * TS_NPQ_VEC + npq_l_vec_reg];
+                for (uint k_l_reg = 0; k_l_reg < TS_K; ++k_l_reg) {
+                    regC[k_l_reg][npq_l_vec_reg] = mad(convert_float(regA[k_l_reg]), convert_float4(regB), regC[k_l_reg][npq_l_vec_reg]);
+                }
+            }
+        }
+        barrier(CLK_LOCAL_MEM_FENCE);
+    }
+
+    for (uint k_l_reg = 0; k_l_reg < TS_K; ++k_l_reg) {
+        const uint k_g = offset_k + lid_k * TS_K + k_l_reg;
+        if (k_g >= K) continue;
+
+        for (uint npq_l_vec_reg = 0; npq_l_vec_reg < TS_NPQ_VEC; ++npq_l_vec_reg) {
+            const uint npq_g_base = offset_npq + (lid_npq * TS_NPQ_VEC + npq_l_vec_reg) * VEC_SIZE;
+
+            const uint N_idx = npq_g_base / (OH * OW);
+            const uint pq_idx = npq_g_base % (OH * OW);
+            const uint OH_idx = pq_idx / OW;
+            const uint OW_idx = pq_idx % OW;
+
+            if (nb1 == OW && OW_idx + VEC_SIZE <= OW && npq_g_base + VEC_SIZE <= NPQ) {
+                const uint dst_idx = OW_idx + OH_idx*nb1 + k_g*nb2 + N_idx*nb3;
+                VSTORE_T_FLOAT4(regC[k_l_reg][npq_l_vec_reg], 0, &dst_data[dst_idx]);
+            } else {
+                T_ACCUM res = regC[k_l_reg][npq_l_vec_reg];
+                for (int v = 0; v < VEC_SIZE; ++v) {
+                    const uint npq_g = npq_g_base + v;
+                    if (npq_g < NPQ) {
+                        const uint N_idx_s = npq_g / (OH*OW);
+                        const uint pq_idx_s = npq_g % (OH*OW);
+                        const uint OH_idx_s = pq_idx_s / OW;
+                        const uint OW_idx_s = pq_idx_s % OW;
+                        const uint dst_idx_s = OW_idx_s + OH_idx_s*nb1 + k_g*nb2 + N_idx_s*nb3;
+                        dst_data[dst_idx_s] = (T_FLOAT)(((float*)&res)[v]);
+                    }
+                }
+            }
+        }
+    }
+}

ggml/src/ggml-opencl/kernels/conv2d_f16_f32.cl

@@ -0,0 +1,176 @@
+#pragma OPENCL EXTENSION cl_khr_fp16 : enable
+
+#if defined(cl_qcom_reqd_sub_group_size)
+#pragma OPENCL EXTENSION cl_qcom_reqd_sub_group_size : enable
+#define REQD_SUBGROUP_SIZE_128 __attribute__((qcom_reqd_sub_group_size("full")))
+#else
+#define REQD_SUBGROUP_SIZE_128
+#endif
+
+#define T_ACCUM float4
+#define VEC_SIZE 4
+
+#define BS_K 64
+#define BS_NPQ 64
+#define BS_CRS 16
+
+#define TS_K 4
+#define TS_NPQ 8
+
+#define WG_K (BS_K / TS_K)
+#define WG_NPQ (BS_NPQ / TS_NPQ)
+
+#define BS_NPQ_VEC (BS_NPQ / VEC_SIZE)
+#define TS_NPQ_VEC (TS_NPQ / VEC_SIZE)
+
+static inline uint splitWork(uint work_size, uint block_size){
+    return (work_size + block_size - 1) / block_size;
+}
+
+REQD_SUBGROUP_SIZE_128
+kernel void kernel_conv_2d(
+    global void* p_knl,
+    ulong off_knl,
+    global void* p_src,
+    ulong off_src,
+    global void* p_dst,
+    ulong off_dst,
+    local void* shared,
+    uint Cout, uint Cin, uint N,
+    uint KW, uint KH, uint W, uint H, uint OW, uint OH,
+    uint s0, uint s1, uint p0, uint p1, uint d0, uint d1,
+    uint nb01, uint nb02, uint nb03,
+    uint nb11, uint nb12, uint nb13,
+    uint nb1, uint nb2, uint nb3
+) {
+    global half* knl_data = (global half*) ((global char*)p_knl + off_knl);
+    global float* src_data = (global float*) ((global char*)p_src + off_src);
+    global float* dst_data = (global float*) ((global char*)p_dst + off_dst);
+
+    const uint K = Cout;
+    const uint CRS = Cin*KH*KW;
+    const uint NPQ = N*OH*OW;
+
+    const uint lid_k = get_local_id(0);
+    const uint lid_npq = get_local_id(1);
+    const uint tid = lid_npq * WG_K + lid_k;
+
+    const uint B_idx_K = get_group_id(0);
+    const uint B_idx_NPQ = get_group_id(1);
+
+    const uint offset_k = B_idx_K * BS_K;
+    const uint offset_npq = B_idx_NPQ * BS_NPQ;
+
+    local half* Ash = (local half*)shared;
+    local float4* Bsh = (local float4*) &Ash[BS_K * BS_CRS];
+
+    T_ACCUM regC[TS_K][TS_NPQ_VEC];
+    for (int i = 0; i < TS_K; ++i) {
+        for (int j = 0; j < TS_NPQ_VEC; ++j) {
+            regC[i][j] = (T_ACCUM)(0.0f);
+        }
+    }
+
+    const uint NB_CRS = splitWork(CRS, BS_CRS);
+
+    for (uint B_idx_CRS = 0; B_idx_CRS < NB_CRS; ++B_idx_CRS) {
+        const uint offset_crs = B_idx_CRS * BS_CRS;
+
+        for (int i = tid; i < BS_K * BS_CRS; i += (WG_K * WG_NPQ)) {
+            const uint k_l = i / BS_CRS;
+            const uint crs_l = i % BS_CRS;
+            const uint k_g = offset_k + k_l;
+            const uint crs_g = offset_crs + crs_l;
+
+            if (k_g < K && crs_g < CRS) {
+                const uint Cin_idx = crs_g / (KW*KH);
+                const uint KH_idx = (crs_g - Cin_idx*KW*KH) / KW;
+                const uint KW_idx = crs_g - Cin_idx*KW*KH - KH_idx*KW;
+                const uint knl_idx = KW_idx + KH_idx*nb01 + Cin_idx*nb02 + k_g*nb03;
+                Ash[k_l * BS_CRS + crs_l] = knl_data[knl_idx];
+            } else {
+                Ash[k_l * BS_CRS + crs_l] = (half)0.0f;
+            }
+        }
+
+        for (int i = tid; i < BS_CRS * BS_NPQ_VEC; i += (WG_K * WG_NPQ)) {
+            const uint crs_l = i / BS_NPQ_VEC;
+            const uint npq_l_vec = i % BS_NPQ_VEC;
+            const uint crs_g = offset_crs + crs_l;
+
+            float4 val = (float4)(0.0f);
+            if (crs_g < CRS) {
+                const uint Cin_idx = crs_g / (KW * KH);
+                const uint KH_idx = (crs_g - Cin_idx * KW * KH) / KW;
+                const uint KW_idx = crs_g - Cin_idx * KW * KH - KH_idx * KW;
+                for (int v = 0; v < VEC_SIZE; ++v) {
+                    const uint npq_g = offset_npq + npq_l_vec * VEC_SIZE + v;
+                    if (npq_g < NPQ) {
+                        const uint N_idx = npq_g / (OH * OW);
+                        const uint pq_idx = npq_g % (OH * OW);
+                        const uint OH_idx = pq_idx / OW;
+                        const uint OW_idx = pq_idx % OW;
+                        const int H_idx = (int)(OH_idx * s1 + KH_idx * d1 - p1);
+                        const int W_idx = (int)(OW_idx * s0 + KW_idx * d0 - p0);
+
+                        if (H_idx >= 0 && H_idx < H && W_idx >= 0 && W_idx < W) {
+                            const uint src_idx = W_idx + H_idx * nb11 + Cin_idx * nb12 + N_idx * nb13;
+                            ((float*)&val)[v] = src_data[src_idx];
+                        }
+                    }
+                }
+            }
+            Bsh[crs_l * BS_NPQ_VEC + npq_l_vec] = val;
+        }
+
+        barrier(CLK_LOCAL_MEM_FENCE);
+
+        #pragma unroll
+        for (uint crs_l = 0; crs_l < BS_CRS; ++crs_l) {
+            half regA[TS_K];
+            for (uint k_l_reg = 0; k_l_reg < TS_K; ++k_l_reg) {
+                regA[k_l_reg] = Ash[(lid_k * TS_K + k_l_reg) * BS_CRS + crs_l];
+            }
+
+            for (uint npq_l_vec_reg = 0; npq_l_vec_reg < TS_NPQ_VEC; ++npq_l_vec_reg) {
+                float4 regB = Bsh[crs_l * BS_NPQ_VEC + lid_npq * TS_NPQ_VEC + npq_l_vec_reg];
+                for (uint k_l_reg = 0; k_l_reg < TS_K; ++k_l_reg) {
+                    regC[k_l_reg][npq_l_vec_reg] = mad(convert_float(regA[k_l_reg]), regB, regC[k_l_reg][npq_l_vec_reg]);
+                }
+            }
+        }
+        barrier(CLK_LOCAL_MEM_FENCE);
+    }
+
+    for (uint k_l_reg = 0; k_l_reg < TS_K; ++k_l_reg) {
+        const uint k_g = offset_k + lid_k * TS_K + k_l_reg;
+        if (k_g >= K) continue;
+
+        for (uint npq_l_vec_reg = 0; npq_l_vec_reg < TS_NPQ_VEC; ++npq_l_vec_reg) {
+            const uint npq_g_base = offset_npq + (lid_npq * TS_NPQ_VEC + npq_l_vec_reg) * VEC_SIZE;
+
+            const uint N_idx = npq_g_base / (OH * OW);
+            const uint pq_idx = npq_g_base % (OH * OW);
+            const uint OH_idx = pq_idx / OW;
+            const uint OW_idx = pq_idx % OW;
+
+            if (nb1 == OW && OW_idx + VEC_SIZE <= OW && npq_g_base + VEC_SIZE <= NPQ) {
+                const uint dst_idx = OW_idx + OH_idx*nb1 + k_g*nb2 + N_idx*nb3;
+                vstore4(regC[k_l_reg][npq_l_vec_reg], 0, &dst_data[dst_idx]);
+            } else {
+                T_ACCUM res = regC[k_l_reg][npq_l_vec_reg];
+                for (int v = 0; v < VEC_SIZE; ++v) {
+                    const uint npq_g = npq_g_base + v;
+                    if (npq_g < NPQ) {
+                        const uint N_idx_s = npq_g / (OH*OW);
+                        const uint pq_idx_s = npq_g % (OH*OW);
+                        const uint OH_idx_s = pq_idx_s / OW;
+                        const uint OW_idx_s = pq_idx_s % OW;
+                        const uint dst_idx_s = OW_idx_s + OH_idx_s*nb1 + k_g*nb2 + N_idx_s*nb3;
+                        dst_data[dst_idx_s] = ((float*)&res)[v];
+                    }
+                }
+            }
+        }
+    }
+}

ggml/src/ggml-opencl/kernels/im2col_f16.cl

@@ -31,7 +31,7 @@ kernel void kernel_im2col_f16(
     src1 = (global float*)((global char*)src1 + offset1);
     dst = (global half*)((global char*)dst + offsetd);
 
-    long  ksize = OW * (KH > 1 ? KW : 1);
+    long  ksize = OW * KH;
     long  kx = i / ksize;
     long  kd = kx * ksize;
     long  ky = (i - kd) / OW;

ggml/src/ggml-opencl/kernels/im2col_f32.cl

@@ -31,7 +31,7 @@ kernel void kernel_im2col_f32(
     src1 = (global float*)((global char*)src1 + offset1);
     dst = (global float*)((global char*)dst + offsetd);
 
-    long  ksize = OW * (KH > 1 ? KW : 1);
+    long  ksize = OW * KH;
     long  kx = i / ksize;
     long  kd = kx * ksize;
     long  ky = (i - kd) / OW;

ggml/src/ggml-sycl/im2col.cpp

@@ -26,7 +26,7 @@ static void im2col_kernel(const float * x, T * dst, int64_t batch_offset, int64_
 
     // make each work-item deal with more elements since sycl global range can not exceed max int
     for (int64_t i = global_id; i < pelements; i += (work_group_size * item_ct1.get_group_range(2))) {
-        const int64_t ksize = OW * (KH > 1 ? KW : 1);
+        const int64_t ksize = OW * KH;
         const int64_t kx    = i / ksize;
         const int64_t kd    = kx * ksize;
         const int64_t ky    = (i - kd) / OW;

ggml/src/ggml-vulkan/ggml-vulkan.cpp

@@ -10248,7 +10248,7 @@ static bool ggml_vk_can_fuse(const struct ggml_cgraph * cgraph, int node_idx, st
         }
         // if rms_norm is the B operand, then we don't handle broadcast
         if (rms_norm == mul->src[1] &&
-            mul->src[0]->ne[1] != rms_norm->ne[1]) {
+            !ggml_are_same_shape(mul->src[0], rms_norm)) {
             return false;
         }
         // rms_norm shader assumes contiguous rows

ggml/src/ggml-vulkan/vulkan-shaders/im2col.comp

@@ -40,12 +40,10 @@ void main() {
     const uint src_base = ic * p.offset_delta + batch * p.batch_offset;
     const uint dst_base = ((batch * p.OH + oh) * p.OW) * p.CHW + ic * (p.KW * p.KH);
     const int oh_s1 = int(oh) * p.s1;
-    const uint ksize = p.OW * (p.KH > 1 ? p.KW : 1);
+    const uint ksize = p.OW * p.KH;
 
     const uint base_linear_idx = gidx * NUM_ITER;
 
-    const uint max_ky = ksize / p.OW;
-
     uint current_kx = base_linear_idx / ksize;
     const uint rem = base_linear_idx - (current_kx * ksize);
     uint current_ky = rem / p.OW;
@@ -76,7 +74,7 @@ void main() {
 
         if (++current_ix == p.OW) {
             current_ix = 0;
-            if (++current_ky == max_ky) {
+            if (++current_ky == p.KH) {
                 current_ky = 0;
                 current_kx++;
             }

ggml/src/ggml-vulkan/vulkan-shaders/rms_norm.comp

@@ -50,8 +50,14 @@ void main() {
     const FLOAT_TYPE scale = inversesqrt(mean + FLOAT_TYPE(p.param1));
 
     if (do_multiply) {
-        [[unroll]] for (uint col = tid; col < ncols; col += BLOCK_SIZE) {
-            data_d[d_offset + col] = D_TYPE(scale * FLOAT_TYPE(data_a[a_offset + col]) * FLOAT_TYPE(data_b[b_offset + col]));
+        if (ncols > p.ne10) {
+            [[unroll]] for (uint col = tid; col < ncols; col += BLOCK_SIZE) {
+                data_d[d_offset + col] = D_TYPE(scale * FLOAT_TYPE(data_a[a_offset + col]) * FLOAT_TYPE(data_b[b_offset + fastmod(col, p.ne10)]));
+            }
+        } else {
+            [[unroll]] for (uint col = tid; col < ncols; col += BLOCK_SIZE) {
+                data_d[d_offset + col] = D_TYPE(scale * FLOAT_TYPE(data_a[a_offset + col]) * FLOAT_TYPE(data_b[b_offset + col]));
+            }
         }
     } else {
         [[unroll]] for (uint col = tid; col < ncols; col += BLOCK_SIZE) {

gguf-py/gguf/metadata.py

@@ -144,6 +144,10 @@ class Metadata:
         # Quick hack to fix the Norway problem
         # https://hitchdev.com/strictyaml/why/implicit-typing-removed/
         yaml_content = yaml_content.replace("- no\n", "- \"no\"\n")
+        # yaml should use 2 spaces insted of tab
+        # this issue has came up with the Qwen/Qwen3-235B-A22B-Instruct-2507 model card
+        #    (I've also sent a pr tp fix the modelcard too)
+        yaml_content = yaml_content.replace("\t", "  ")
 
         if yaml_content:
             data = yaml.safe_load(yaml_content)

src/llama-model.cpp

@@ -1544,7 +1544,11 @@ void llama_model::load_hparams(llama_model_loader & ml) {
                 ml.get_key(LLM_KV_TOKEN_SHIFT_COUNT,                      hparams.token_shift_count, false);
 
                 switch (hparams.n_layer) {
-                    case 12: type = LLM_TYPE_190M; break;
+                    case 12:
+                        switch (hparams.n_embd) {
+                            case 768: type = LLM_TYPE_190M; break;
+                            default: type = LLM_TYPE_UNKNOWN;
+                        } break;
                     case 24:
                         switch (hparams.n_embd) {
                             case 1024: type = LLM_TYPE_450M; break;
@@ -1557,7 +1561,17 @@ void llama_model::load_hparams(llama_model_loader & ml) {
                             case 3584: type = LLM_TYPE_7B; break;
                             default: type = LLM_TYPE_UNKNOWN;
                         } break;
-                    case 32: type = LLM_TYPE_2_9B; break; // RWKV-7-World
+                    case 32:
+                        switch (hparams.n_embd) {
+                            case 2560: type = LLM_TYPE_2_9B; break;
+                            case 4096: type = LLM_TYPE_7B; break;
+                            default: type = LLM_TYPE_UNKNOWN;
+                        } break;
+                    case 61:
+                        switch (hparams.n_embd) {
+                            case 4096: type = LLM_TYPE_14B; break;
+                            default: type = LLM_TYPE_UNKNOWN;
+                        } break;
                     default: type = LLM_TYPE_UNKNOWN;
                 }
             } break;

tests/test-backend-ops.cpp

@@ -2641,6 +2641,7 @@ struct test_rms_norm_mul_add : public test_case {
     const ggml_type type;
     const std::array<int64_t, 4> ne;
     const float eps;
+    const bool broadcast;
 
     std::string op_desc(ggml_tensor * t) override {
         GGML_UNUSED(t);
@@ -2650,18 +2651,21 @@ struct test_rms_norm_mul_add : public test_case {
     bool run_whole_graph() override { return true; }
 
     std::string vars() override {
-        return VARS_TO_STR3(type, ne, eps);
+        return VARS_TO_STR4(type, ne, eps, broadcast);
     }
 
     test_rms_norm_mul_add(ggml_type type = GGML_TYPE_F32,
             std::array<int64_t, 4> ne = {64, 5, 4, 3},
-            float eps = 1e-6f)
-        : type(type), ne(ne), eps(eps) {}
+            float eps = 1e-6f, bool broadcast = false)
+        : type(type), ne(ne), eps(eps), broadcast(broadcast) {}
 
     ggml_tensor * build_graph(ggml_context * ctx) override {
-        ggml_tensor * a = ggml_new_tensor(ctx, type, 4, ne.data());
+        std::array<int64_t, 4> broadcast_dims = {ne[0]*2, ne[1]*3, ne[2]*3, ne[3]*4};
+
+        ggml_tensor * a = ggml_new_tensor(ctx, type, 4, broadcast ? broadcast_dims.data() : ne.data());
         ggml_tensor * b = ggml_new_tensor(ctx, type, 4, ne.data());
         ggml_tensor * c = ggml_new_tensor(ctx, type, 4, ne.data());
+
         ggml_set_param(a);
         ggml_set_name(a, "a");
         ggml_set_param(b);
@@ -5093,6 +5097,7 @@ static std::vector<std::unique_ptr<test_case>> make_test_cases_eval() {
     test_cases.emplace_back(new test_im2col(GGML_TYPE_F32, GGML_TYPE_F16, GGML_TYPE_F16, {12, 12, 2, 2048}, {3, 3, 2, 2048}, 1, 1, 1, 1, 1, 1, true));
     test_cases.emplace_back(new test_im2col(GGML_TYPE_F32, GGML_TYPE_F16, GGML_TYPE_F16, {12, 12, 1, 2560}, {3, 3, 1, 2560}, 1, 1, 1, 1, 1, 1, true));
     test_cases.emplace_back(new test_im2col(GGML_TYPE_F32, GGML_TYPE_F16, GGML_TYPE_F16, {12, 12, 2, 2560}, {3, 3, 2, 2560}, 1, 1, 1, 1, 1, 1, true));
+    test_cases.emplace_back(new test_im2col(GGML_TYPE_F32, GGML_TYPE_F16, GGML_TYPE_F16, {5, 5, 1, 32}, {3, 4, 1, 32}, 1, 1, 0, 0, 1, 1, true));
 
 // Conv_2D test cases
 #ifdef DETAILED_TESTS
@@ -5353,6 +5358,7 @@ static std::vector<std::unique_ptr<test_case>> make_test_cases_eval() {
     }
     for (float eps : {0.0f, 1e-6f, 1e-4f, 1e-1f, 1.0f}) {
         test_cases.emplace_back(new test_rms_norm_mul_add(GGML_TYPE_F32, {64, 5, 4, 3}, eps));
+        test_cases.emplace_back(new test_rms_norm_mul_add(GGML_TYPE_F32, {64, 5, 4, 3}, eps, true));
     }
 
     test_cases.emplace_back(new test_l2_norm(GGML_TYPE_F32, {64, 5, 4, 3}, 1e-12f));

tools/imatrix/README.md

@@ -1,34 +1,92 @@
 # llama.cpp/tools/imatrix
 
 Compute an importance matrix for a model and given text dataset. Can be used during quantization to enhance the quality of the quantized models.
-More information is available here: https://github.com/ggml-org/llama.cpp/pull/4861
+More information is available in <https://github.com/ggml-org/llama.cpp/pull/4861>.
 
 ## Usage
 
 ```
 ./llama-imatrix \
-    -m model.gguf -f some-text.txt [-o imatrix.gguf] [--process-output] \
-    [--no-ppl] [--chunk 123] [--output-frequency 10] [--save-frequency 0] \
-    [--in-file imatrix-prev-0.gguf --in-file imatrix-prev-1.gguf ...] \
-    [--parse-special]
+    -m model.gguf -f some-text.txt [-o imatrix.gguf] [--no-ppl] \
+    [--process-output] [--chunk 123] [--save-frequency 0] [--output-frequency 10] \
+    [--in-file imatrix-prev-0.gguf --in-file imatrix-prev-1.gguf ...] [--parse-special] \
+    [--show-statistics] [...]
 ```
 
-Here `-m` with a model name and `-f` with a file containing training data (such as e.g. `wiki.train.raw`) are mandatory.
+Here `-m | --model` with a model name and `-f | --file` with a file containing calibration data (such as e.g. `wiki.train.raw`) are mandatory.
 The parameters in square brackets are optional and have the following meaning:
-* `-o` (or `--output-file`) specifies the name of the file where the computed data will be stored. If missing `imatrix.gguf` is used.
-* `--verbosity` specifies the verbosity level. If set to `0`, no output other than the perplexity of the processed chunks will be generated. If set to `1`, each time the results are saved a message is written to `stderr`. If `>=2`, a message is output each time data is collected for any tensor. Default verbosity level is `1`.
-* `--output-frequency` specifies how often the so far computed result is saved to disk. Default is 10 (i.e., every 10 chunks)
+
+* `-h | --help` shows usage information and exits.
+* `-lv | --verbosity` specifies the verbosity level. If set to `0`, no output other than the perplexity of the processed chunks will be generated. If set to `1`, each time the results are saved a message is written to `stderr`. If `>=2`, a message is output each time data is collected for any tensor. Default verbosity level is `1`.
+* `-o | --output-file` specifies the name of the file where the computed data will be stored. If missing `imatrix.gguf` is used.
+* `-ofreq | --output-frequency` specifies how often the so far computed result is saved to disk. Default is 10 (i.e., every 10 chunks)
 * `--save-frequency` specifies how often to save a copy of the imatrix in a separate file. Default is 0 (i.e., never)
-* `--process-output` specifies if data will be collected for the `output.weight` tensor. My experience is that it is better to not utilize the importance matrix when quantizing `output.weight`, so this is set to `false` by default.
+* `--process-output` specifies if data will be collected for the `output.weight` tensor. Typically, it is better not to utilize the importance matrix when quantizing `output.weight`, so this is set to `false` by default.
+* `--in-file` one or more existing imatrix files to load and combine. Useful for merging files from multiple runs/datasets.
+* `--parse-special` enables parsing of special tokens (e.g., `<|im_start|>` in some models). Useful for models with custom tokenizers.
+* `--chunk | --from-chunk` to skip the first `n` chunks of tokens from the input data. Useful for resuming or skipping initial low-quality data.
+* `--chunks` maximum number of chunks to process. Default is -1 for all available chunks.
+* `--no-ppl` disables the calculation of perplexity for the processed chunks. Useful if you want to speed up the processing and do not care about perplexity.
+* `--show-statistics` displays imatrix file's statistics.
 
-For faster computation, make sure to use GPU offloading via the `-ngl` argument
+For faster computation, make sure to use GPU offloading via the `-ngl | --n-gpu-layers` argument.
 
-## Example
+Recent versions of `llama-imatrix` store data in GGUF format by default. For the legacy format, use an extension other than `.gguf` when saving the output file. More information is available in <https://github.com/ggml-org/llama.cpp/pull/9400>.
+
+## Examples
 
 ```bash
-# generate importance matrix (imatrix.gguf)
-./llama-imatrix -m ggml-model-f16.gguf -f train-data.txt -ngl 99
+# generate importance matrix using default filename (imatrix.gguf), offloading 99 layers to GPU
+./llama-imatrix -m ggml-model-f16.gguf -f calibration-data.txt -ngl 99
 
 # use the imatrix to perform a Q4_K_M quantization
 ./llama-quantize --imatrix imatrix.gguf ggml-model-f16.gguf ./ggml-model-q4_k_m.gguf q4_k_m
 ```
+
+```bash
+# generate and save the imatrix using legacy format
+./llama-imatrix -m ggml-model-f16.gguf -f calibration-data.txt -o imatrix-legcy-format.dat -ngl 99
+```
+
+```bash
+# covert legacy (binary) imatrix format to new (GGUF) format
+./llama-imatrix --in-file imatrix-legacy-format.dat -o imatrix-new-format.gguf
+```
+
+```bash
+# combine existing imatrices
+./llama-imatrix --in-file imatrix-prev-0.gguf --in-file imatrix-prev-1.gguf -o imatrix-combined.gguf
+```
+
+```bash
+# skip first 5 chunks, save intermediates every 20 chunks and snapshots every 50, parsing special tokens
+./llama-imatrix -m ggml-model-f16.gguf -f calibration-data.txt --chunk 5 --output-frequency 20 --save-frequency 50 --parse-special
+```
+
+```bash
+# analyse imatrix file and display summary statistics instead of running inference
+./llama-imatrix --in-file imatrix.gguf --show-statistics
+```
+
+`--show-statistics` will display the following statistics:
+
+#### Per tensor
+
+* Σ(Act²): sum of all squared activations (the importance scores)
+* Min & Max: minimum and maximum squared activations values
+* μ & σ: Squared activations' mean and standard deviation
+* % Active: proportion of elements whose average squared activation exceeds a small threshold (1e-5). Helpful to determine how alive/dormant the tensor is during inference
+* N: number of squared activations
+* Entropy: entropy of the squared activation distribution, in bits (standard Shannon entropy measurement) $S = -\sum_{i=1}^N p_i \log_2 p_i$
+* E (norm): Normalized entropy. $E(norm)=\frac{-\sum_{i=1}^N p_i \log_2 p_i}{log_2 N}$. These two metrics can be used to determine how well a prompt "exercises" the model's capabilities
+* ZD Score: z-score distribution as described in _3.1 Layer Importance Scores_ of [Layer-Wise Quantization](https://arxiv.org/abs/2406.17415)
+* CosSim: cosine similarity with respect to the previous layer's tensor. Useful to determine how similar the squared activations of the current layer are to the previous layer's squared activations.
+
+#### Per layer
+
+Weighted averages of Σ(Act²), ZD Score and CosSim are also calculated.
+
+#### Important note on the computed Statistics
+
+When using these statistics, please note that they are computed on the squared activations, **not on the actual (raw) activations**.
+Whilst the results are still useful, they're less realiable than using the raw values, and in the case of the cosine similarity, could be misleading if the tensor contains opposite vectors.

tools/imatrix/imatrix.cpp

@@ -16,6 +16,8 @@
 #include <fstream>
 #include <unordered_map>
 #include <map>
+#include <regex>
+#include <numeric>
 
 #if defined(_MSC_VER)
 #pragma warning(disable: 4244 4267) // possible loss of data
@@ -24,10 +26,10 @@
 static void print_usage(int, char ** argv) {
     LOG("\nexample usage:\n");
     LOG("\n    %s \\\n"
-            "       -m model.gguf -f some-text.txt [-o imatrix.gguf] [--process-output] \\\n"
-            "       [--no-ppl] [--chunk 123] [--output-frequency 10] [--save-frequency 0] \\\n"
-            "       [--in-file imatrix-prev-0.gguf --in-file imatrix-prev-1.gguf ...] \\\n"
-            "       [--parse-special]\n" , argv[0]);
+            "       -m model.gguf -f some-text.txt [-o imatrix.gguf] [--no-ppl] \\\n"
+            "       [--process-output] [--chunk 123] [--save-frequency 0] [--output-frequency 10] \\\n"
+            "       [--in-file imatrix-prev-0.gguf --in-file imatrix-prev-1.gguf ...] [--parse-special] \\\n"
+            "       [--show-statistics] [...]\n" , argv[0]);
     LOG("\n");
 }
 
@@ -40,6 +42,21 @@ struct Stats {
     std::vector<int64_t> counts;
 };
 
+struct tensor_statistics {
+    std::string tensor;
+    Stats stats;
+    float total_sqract = 0.0f;
+    float mean_sqract  = 0.0f;
+    float max_sqract   = 0.0f;
+    float min_sqract   = 0.0f;
+    int elements       = 0;
+    float stddev       = 0.0f;
+    float active       = 0.0f;
+    float entropy      = 0.0f;
+    float zd           = 0.0f;
+    float cossim       = 0.0f;
+};
+
 class IMatrixCollector {
 public:
     IMatrixCollector() = default;
@@ -49,6 +66,7 @@ public:
     void save_imatrix(int32_t n_chunk = -1) const;
     bool load_imatrix_legacy(const char * fname);
     bool load_imatrix(const char * file_name);
+    const std::unordered_map<std::string, Stats> & get_mstats() const { return m_stats; }
 private:
     std::unordered_map<std::string, Stats> m_stats;
     common_params                          m_params;
@@ -78,6 +96,126 @@ static std::string filter_tensor_name(const char * name) {
     return wname;
 }
 
+static void process_tensor_name(const std::string & input, std::string & layer, std::string & tensor) {
+    std::vector<std::string> name;
+    std::istringstream stream(input);
+    std::string item;
+
+    while (std::getline(stream, item, '.')) {
+        name.push_back(item);
+    }
+    for (size_t i = 0; i < name.size(); ++i) {
+        if (name[i] == "blk" && i + 1 < name.size()) {
+            layer = name[i + 1];
+            break;
+        }
+    }
+    for (size_t i = 0; i < name.size(); ++i) {
+        if (name[i] == "weight" && i > 0) {
+            tensor = name[i - 1];
+            break;
+        }
+    }
+
+    if (tensor.empty()) {
+        tensor = input;
+    }
+    if (layer.empty()) {
+        layer = "-";
+    }
+}
+
+static void compute_statistics(std::vector<tensor_statistics> & tstats, const std::string & name, const Stats & e) {
+    if (e.values.size() % e.counts.size() != 0) {
+        LOG_ERR("%s: activation size mismatch for tensor %s (%zu vs %zu)\n", __func__, name.c_str(), e.counts.size(), e.values.size());
+        return;
+    }
+    if (e.counts.empty()) {
+        LOG_ERR("%s: there are no activations for tensor %s. The imatrix may be suboptimal\n", __func__, name.c_str());
+        return;
+    }
+
+    const int n_mat = e.counts.size();
+    const int row_size = e.values.size() / n_mat;
+
+    std::vector<float> activations;
+    activations.reserve(e.values.size());
+
+    for (int i = 0; i < n_mat; ++i) {
+        for (int j = 0; j < row_size; ++j) {
+            activations.push_back(e.values[i*row_size + j] / e.counts[i]);
+        }
+    }
+
+    const float act_total     = std::accumulate(activations.begin(), activations.end(), 0.0f);
+    const float act_max       = *std::max_element(activations.begin(), activations.end());
+    const float act_min       = *std::min_element(activations.begin(), activations.end());
+    const float act_mean      = act_total / activations.size();
+    const float act_sqr_total = std::inner_product(activations.begin(), activations.end(), activations.begin(), 0.0f);
+    const float act_var       = (act_sqr_total / activations.size()) - (act_mean * act_mean);
+    const float act_dev       = std::sqrt(std::max(0.0f, act_var));
+    float threshold           = 1e-5f;
+    const int inactive_count  = std::count_if(activations.begin(), activations.end(),
+                                               [threshold](const float v) { return fabsf(v) <= threshold; });
+    const float active_ratio  = 1 - static_cast<float>(inactive_count) / activations.size();
+
+    float entropy = 0;
+    if (act_total > 0) {
+        for (const auto act : activations) {
+            if (const float p = act / act_total; p > 0) {
+                entropy -= p * std::log2(p);
+            }
+        }
+    }
+
+    int z_score = 0;
+    if (act_dev > 0.0f) {
+        for (const auto act : activations) {
+            if (const float p = (act - act_mean) / act_dev; p > 1) {
+                z_score++;
+            }
+        }
+    }
+
+    auto & ts = tstats.emplace_back();
+    ts.tensor     = name;
+    ts.stats      = e;
+    ts.total_sqract = act_total;
+    ts.mean_sqract  = act_mean;
+    ts.max_sqract   = act_max;
+    ts.min_sqract   = act_min;
+    ts.elements   = static_cast<int>(activations.size());
+    ts.stddev     = act_dev;
+    ts.active     = active_ratio;
+    ts.entropy    = entropy;
+    ts.zd         = static_cast<float>(z_score) / ts.elements;
+}
+
+static void compute_cossim(std::vector<tensor_statistics> & tstats) {
+    static const std::regex pattern(R"(blk\.(\d+)\.)");
+    for (auto & ts : tstats) {
+        if (std::smatch match; std::regex_search(ts.tensor, match, pattern)) {
+            const int blk = std::stoi(match[1]);
+            std::string tname(ts.tensor);
+            tname.replace(match.position(1), match.length(1), std::to_string(blk-1));
+            auto prev = std::find_if(tstats.begin(), tstats.end(),
+                [tname](const tensor_statistics & t) { return t.tensor == tname; });
+            if (prev != tstats.end()) {
+                const float dp = std::inner_product(ts.stats.values.begin(), ts.stats.values.end(),
+                    prev->stats.values.begin(), 0.0f);
+                const float curr_mag = std::sqrt(std::inner_product(ts.stats.values.begin(), ts.stats.values.end(),
+                    ts.stats.values.begin(), 0.0f));
+                const float prev_mag = std::sqrt(std::inner_product(prev->stats.values.begin(), prev->stats.values.end(),
+                    prev->stats.values.begin(), 0.0f));
+                const float cs = dp / (curr_mag * prev_mag);
+                ts.cossim = cs;
+            }
+        } else {
+            ts.cossim = 0;
+        }
+    }
+}
+
 bool IMatrixCollector::collect_imatrix(struct ggml_tensor * t, bool ask, void * user_data) {
     GGML_UNUSED(user_data);
 
@@ -678,7 +816,6 @@ static bool ik_collect_imatrix(struct ggml_tensor * t, bool ask, void * user_dat
     return g_collector.collect_imatrix(t, ask, user_data);
 }
 
-
 struct results_log_softmax {
     double log_softmax;
     float  logit;
@@ -926,6 +1063,113 @@ static bool compute_imatrix(llama_context * ctx, const common_params & params, c
     return true;
 }
 
+static bool show_statistics(const common_params & params) {
+    std::vector<tensor_statistics> ts;
+    if (params.in_files.empty() || params.in_files.size() > 1) {
+        LOG_ERR("\nError: a single imatrix file is required to compute tensor statistics\n\n");
+        return false;
+    }
+    if (g_collector.load_imatrix(params.in_files[0].c_str())) {
+        for (const auto & [name, stats] :g_collector.get_mstats()) {
+            compute_statistics(ts, name, stats);
+        }
+    } else {
+        LOG_ERR("\nError: %s is not a valid imatrix file\n\n", params.in_files[0].c_str());
+        return false;
+    }
+    if (!ts.empty()) {
+        compute_cossim(ts);
+    } else {
+        LOG_ERR("Error: cannot compute statistics for %s\n\n", params.in_files[0].c_str());
+        return false;
+    }
+
+    struct tensor_comparer {
+        bool operator()(const tensor_statistics & a, const tensor_statistics & b) const {
+            std::string layer, name_a, name_b;
+            ;
+            process_tensor_name(a.tensor, layer, name_a);
+            process_tensor_name(b.tensor, layer, name_b);
+            return name_a < name_b || (name_a == name_b && a.total_sqract > b.total_sqract);
+        }
+    };
+    std::sort(ts.begin(), ts.end(), tensor_comparer());
+
+    struct weighted_stats {
+        float weighted_bias   = 0.0f;
+        float weighted_zd     = 0.0f;
+        float weighted_cossim = 0.0f;
+        int   total_elements  = 0;
+    };
+    std::map<int, weighted_stats> ws;
+
+    LOG_INF("\nComputing statistics for %s (%d tensors)\n", params.in_files[0].c_str(), static_cast<int>(ts.size()));
+    LOG_INF("\n%s\t%s\t%s\t%s\t%s\t%s\t%s\t%s\t%s\t%s\t%s\t%s\t%s\n", " Layer", "       Tensor", "          Σ(Act²)",
+            "  Min", "            Max", "           μ", "   σ", " % Active", "N", "   Entropy", "E (norm)", "ZD",
+            "  CosSim");
+    LOG_INF(
+        "=============================================================================================================="
+        "===========================================================\n");
+    for (const auto & tstat : ts) {
+        std::string layer, name;
+        process_tensor_name(tstat.tensor, layer, name);
+
+        int blk;
+        try {
+            blk = std::stoi(layer);
+        } catch (const std::exception & e) {
+            blk = -1;  // not a block layer
+        }
+
+        LOG_INF("%5s\t%-20s\t%10.2f\t%8.4f\t%11.4f\t%6.2f\t%6.2f\t%8.2f%%\t%6d\t%10.4f\t%6.2f%%\t%10.2f%%\t%8.4f\n",
+                layer.c_str(), name.c_str(), tstat.total_sqract, tstat.min_sqract, tstat.max_sqract, tstat.mean_sqract,
+                tstat.stddev, tstat.active * 100.0f, tstat.elements, tstat.entropy,
+                100.0f * (tstat.entropy / std::log2(tstat.elements)), 100.0f * tstat.zd, tstat.cossim);
+
+        const float weighted_bias   = tstat.elements * tstat.total_sqract;
+        const float weighted_zd     = tstat.elements * tstat.zd;
+        const float weighted_cossim = tstat.elements * tstat.cossim;
+
+        if (ws.find(blk) != ws.end()) {
+            ws[blk].weighted_bias += weighted_bias;
+            ws[blk].weighted_zd += weighted_zd;
+            ws[blk].weighted_cossim += weighted_cossim;
+            ws[blk].total_elements += tstat.elements;
+        } else {
+            weighted_stats temp_ws;
+            temp_ws.weighted_bias   = weighted_bias;
+            temp_ws.weighted_zd     = weighted_zd;
+            temp_ws.weighted_cossim = weighted_cossim;
+            temp_ws.total_elements  = tstat.elements;
+            ws[blk]                 = temp_ws;
+        }
+    }
+
+    const int layers = std::count_if(ws.begin(), ws.end(), [](const auto & kv) { return kv.first >= 0; });
+    LOG_INF("\nComputing weighted average statistics per layer (%d layers)\n", layers);
+    LOG_INF("\n%s\t%s\t%s\t%s\n", "  Layer", "     μΣ(Act²)", "      μZD", "μCosSim");
+    LOG_INF("================================================\n");
+    for (const auto & [first, second] : ws) {
+        const auto & layer = first;
+        const auto & stats = second;
+
+        if (stats.total_elements == 0) {
+            continue;
+        }
+
+        if (layer >= 0) {
+            const float bias   = stats.weighted_bias / stats.total_elements;
+            const float zd     = stats.weighted_zd / stats.total_elements;
+            const float cossim = stats.weighted_cossim / stats.total_elements;
+
+            LOG_INF("%5d\t%14.2f\t%10.4f%%\t%6.4f\n", layer, bias, 100.0f * zd, cossim);
+        }
+    }
+    LOG_INF("\n");
+
+    return true;
+}
+
 int main(int argc, char ** argv) {
     common_params params;
 
@@ -938,6 +1182,13 @@ int main(int argc, char ** argv) {
         return 1;
     }
 
+    if (params.show_statistics) {
+        if (!show_statistics(params)) {
+            return 1;
+        }
+        return 0;
+    }
+
     common_init();
 
     const int32_t n_ctx = params.n_ctx;

tools/mtmd/clip.cpp

@@ -367,8 +367,8 @@ struct clip_ctx {
     std::vector<ggml_backend_t> backend_ptrs;
     std::vector<ggml_backend_buffer_type_t> backend_buft;
 
-    ggml_backend_t backend;
-    ggml_backend_t backend_cpu;
+    ggml_backend_t backend = nullptr;
+    ggml_backend_t backend_cpu = nullptr;
     ggml_backend_buffer_ptr buf;
 
     int max_nodes = 8192;
@@ -384,9 +384,18 @@ struct clip_ctx {
         if (!backend_cpu) {
             throw std::runtime_error("failed to initialize CPU backend");
         }
-        backend = ctx_params.use_gpu
-                    ? ggml_backend_init_by_type(GGML_BACKEND_DEVICE_TYPE_GPU, nullptr)
-                    : nullptr;
+        if (ctx_params.use_gpu) {
+            auto backend_name = std::getenv("MTMD_BACKEND_DEVICE");
+            if (backend_name != nullptr) {
+                backend = ggml_backend_init_by_name(backend_name, nullptr);
+                if (!backend) {
+                    LOG_WRN("%s: Warning: Failed to initialize \"%s\" backend, falling back to default GPU backend\n", __func__, backend_name);
+                }
+            }
+            if (!backend) {
+                backend = ggml_backend_init_by_type(GGML_BACKEND_DEVICE_TYPE_GPU, nullptr);
+            }
+        }
 
         if (backend) {
             LOG_INF("%s: CLIP using %s backend\n", __func__, ggml_backend_name(backend));

tools/server/server.cpp

@@ -253,6 +253,7 @@ struct server_task {
         defaults.sampling    = params_base.sampling;
         defaults.speculative = params_base.speculative;
         defaults.n_keep      = params_base.n_keep;
+        defaults.antiprompt  = params_base.antiprompt;
 
         // enabling this will output extra debug information in the HTTP responses from the server
         params.verbose           = params_base.verbosity > 9;
@@ -490,6 +491,10 @@ struct server_task {
                     }
                 }
             }
+            // set reverse prompt from cli args if not set in the request
+            if (params.antiprompt.empty()) {
+                params.antiprompt = defaults.antiprompt;
+            }
         }
 
         {