whispercpp 1.3.5 → 1.3.7

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

@@ -58,6 +58,10 @@ static enum ggml_status ggml_zdnn_graph_compute(ggml_backend_t backend, ggml_cgr
             continue;
         }
 
+        if ((node->flags & GGML_TENSOR_FLAG_COMPUTE) == 0) {
+            continue;
+        }
+
         bool ok = ggml_zdnn_compute_forward(ctx, node);
         if (!ok) {
             GGML_LOG_ERROR("%s: unsupported op %s (%s)\n",
@@ -309,6 +313,8 @@ static ggml_backend_buffer_i ggml_backend_zdnn_buffer_i = {
     /* .memset_tensor = */ ggml_backend_zdnn_buffer_memset_tensor,
     /* .set_tensor    = */ ggml_backend_zdnn_buffer_set_tensor,
     /* .get_tensor    = */ ggml_backend_zdnn_buffer_get_tensor,
+    /* .set_tensor_2d = */ NULL,
+    /* .get_tensor_2d = */ NULL,
     /* .cpy_tensor    = */ NULL,
     /* .clear         = */ ggml_backend_zdnn_buffer_clear,
     /* .reset         = */ NULL,
@@ -368,7 +374,8 @@ static size_t ggml_backend_zdnn_buffer_type_get_alignment(ggml_backend_buffer_ty
 }
 
 static bool ggml_backend_zdnn_buffer_type_is_host(ggml_backend_buffer_type_t buft) {
-    return true;
+    /* while it resides in host memory, additional transformation is needed */
+    return false;
 
     GGML_UNUSED(buft);
 }
@@ -412,20 +419,22 @@ static enum ggml_status ggml_backend_zdnn_graph_compute(ggml_backend_t backend,
 }
 
 static ggml_backend_i ggml_backend_zdnn_i = {
-    /* .get_name           = */ ggml_backend_zdnn_name,
-    /* .free               = */ ggml_backend_zdnn_free,
-    /* .set_tensor_async   = */ NULL,
-    /* .get_tensor_async   = */ NULL,
-    /* .cpy_tensor_async   = */ NULL,
-    /* .synchronize        = */ NULL,
-    /* .graph_plan_create  = */ NULL,
-    /* .graph_plan_free    = */ NULL,
-    /* .graph_plan_update  = */ NULL,
-    /* .graph_plan_compute = */ NULL,
-    /* .graph_compute      = */ ggml_backend_zdnn_graph_compute,
-    /* .event_record       = */ NULL,
-    /* .event_wait         = */ NULL,
-    /* .graph_optimize     = */ NULL,
+    /* .get_name               = */ ggml_backend_zdnn_name,
+    /* .free                   = */ ggml_backend_zdnn_free,
+    /* .set_tensor_async       = */ NULL,
+    /* .get_tensor_async       = */ NULL,
+    /* .set_tensor_2d_async    = */ NULL,
+    /* .get_tensor_2d_async    = */ NULL,
+    /* .cpy_tensor_async       = */ NULL,
+    /* .synchronize            = */ NULL,
+    /* .graph_plan_create      = */ NULL,
+    /* .graph_plan_free        = */ NULL,
+    /* .graph_plan_update      = */ NULL,
+    /* .graph_plan_compute     = */ NULL,
+    /* .graph_compute          = */ ggml_backend_zdnn_graph_compute,
+    /* .event_record           = */ NULL,
+    /* .event_wait             = */ NULL,
+    /* .graph_optimize         = */ NULL,
 };
 
 static ggml_guid_t ggml_backend_zdnn_guid(void) {

data/ext/sources/ggml/src/ggml-zendnn/CMakeLists.txt

@@ -1,12 +1,19 @@
 ggml_add_backend_library(ggml-zendnn
                          ggml-zendnn.cpp)
 
-# Get ZenDNN path
 if (NOT DEFINED ZENDNN_ROOT OR ZENDNN_ROOT STREQUAL "")
     set(ZENDNN_ROOT "$ENV{ZENDNN_ROOT}")
 endif()
 
-# Check if path is still empty or OFF
+if (BUILD_SHARED_LIBS)
+    set(ZENDNN_SHARED_LIB ON)
+    set(ZENDNN_ARCHIVE_LIB OFF)
+else()
+    set(ZENDNN_SHARED_LIB OFF)
+    set(ZENDNN_ARCHIVE_LIB ON)
+endif()
+
+# Download and build ZenDNN if not provided
 if (NOT ZENDNN_ROOT OR ZENDNN_ROOT STREQUAL "" OR ZENDNN_ROOT STREQUAL "OFF")
     message(STATUS "ZENDNN_ROOT not set. Automatically downloading and building ZenDNN...")
     message(STATUS "This will take several minutes on first build...")
@@ -21,7 +28,7 @@ if (NOT ZENDNN_ROOT OR ZENDNN_ROOT STREQUAL "" OR ZENDNN_ROOT STREQUAL "OFF")
     ExternalProject_Add(
         zendnn
         GIT_REPOSITORY https://github.com/amd/ZenDNN.git
-        GIT_TAG zendnnl
+        GIT_TAG 253b94ce0d7e9284c265fefb485714944caff9d3    # ZenDNN-2026-WW19
         PREFIX      ${ZENDNN_PREFIX}
         SOURCE_DIR  ${ZENDNN_SOURCE_DIR}
         BINARY_DIR  ${ZENDNN_BUILD_DIR}
@@ -32,7 +39,9 @@ if (NOT ZENDNN_ROOT OR ZENDNN_ROOT STREQUAL "" OR ZENDNN_ROOT STREQUAL "OFF")
             -DZENDNNL_BUILD_DOXYGEN=OFF
             -DZENDNNL_BUILD_GTEST=OFF
             -DZENDNNL_BUILD_BENCHDNN=OFF
-            # Enable ALL matmul algorithm backends
+            -DZENDNNL_DEPENDS_FBGEMM=OFF
+            -DZENDNNL_LIB_BUILD_ARCHIVE=${ZENDNN_ARCHIVE_LIB}
+            -DZENDNNL_LIB_BUILD_SHARED=${ZENDNN_SHARED_LIB}
             -DZENDNNL_DEPENDS_AOCLDLP=ON
             -DZENDNNL_DEPENDS_ONEDNN=ON
             -DZENDNNL_DEPENDS_LIBXSMM=ON
@@ -45,47 +54,37 @@ if (NOT ZENDNN_ROOT OR ZENDNN_ROOT STREQUAL "" OR ZENDNN_ROOT STREQUAL "OFF")
         LOG_INSTALL ON
     )
 
-    # Add dependency so ZenDNN builds before our library
     add_dependencies(ggml-zendnn zendnn)
-
-    # Set ZENDNN_ROOT to the installation directory
     set(ZENDNN_ROOT ${ZENDNN_INSTALL_DIR})
-
     message(STATUS "ZenDNN will be built to: ${ZENDNN_ROOT}")
 else()
     message(STATUS "Using custom ZenDNN installation at: ${ZENDNN_ROOT}")
 endif()
 
-# ZenDNN headers + libs
 target_include_directories(ggml-zendnn PRIVATE
     ${ZENDNN_ROOT}/zendnnl/include
-    ${ZENDNN_ROOT}/deps/aocldlp/include
-    ${ZENDNN_ROOT}/deps/aoclutils/include
     ${ZENDNN_ROOT}/deps/json/include
-    ${ZENDNN_ROOT}/deps/libxsmm/include
+    ${ZENDNN_ROOT}/deps/aoclutils/include
+    ${ZENDNN_ROOT}/deps/aocldlp/include
     ${ZENDNN_ROOT}/deps/onednn/include
-)
+    ${ZENDNN_ROOT}/deps/libxsmm/include)
 
-target_link_directories(ggml-zendnn PRIVATE
-    ${ZENDNN_ROOT}/zendnnl/lib
-    ${ZENDNN_ROOT}/deps/aocldlp/lib
-    ${ZENDNN_ROOT}/deps/aoclutils/lib
-    ${ZENDNN_ROOT}/deps/libxsmm/lib
-    ${ZENDNN_ROOT}/deps/onednn/lib
-)
+if (ZENDNN_SHARED_LIB)
+    target_link_directories(ggml-zendnn PRIVATE ${ZENDNN_ROOT}/zendnnl/lib)
+    target_link_libraries(ggml-zendnn PRIVATE zendnnl)
+elseif (ZENDNN_ARCHIVE_LIB)
+    target_link_libraries(ggml-zendnn PRIVATE
+        ${ZENDNN_ROOT}/zendnnl/lib/libzendnnl_archive.a
+        ${ZENDNN_ROOT}/deps/aoclutils/${CMAKE_INSTALL_LIBDIR}/libaoclutils.a
+        ${ZENDNN_ROOT}/deps/aoclutils/${CMAKE_INSTALL_LIBDIR}/libau_cpuid.a
+        ${ZENDNN_ROOT}/deps/aocldlp/lib/libaocl-dlp.a
+        ${ZENDNN_ROOT}/deps/onednn/${CMAKE_INSTALL_LIBDIR}/libdnnl.a
+        ${ZENDNN_ROOT}/deps/libxsmm/lib/libxsmm.a
+        ${ZENDNN_ROOT}/deps/libxsmm/lib/libxsmmext.a
+        ${ZENDNN_ROOT}/deps/libxsmm/lib/libxsmmnoblas.a)
+endif()
 
-target_link_libraries(ggml-zendnn PRIVATE
-    zendnnl_archive    # ZenDNN main
-    aocl-dlp           # AOCL libraries
-    aoclutils
-    au_cpuid
-    dnnl               # OneDNN
-    xsmm               # libxsmm small matrix math
-    xsmmext
-    xsmmnoblas
-    m
-    pthread
-)
+target_link_libraries(ggml-zendnn PRIVATE m pthread)
 
 if (GGML_OPENMP)
     target_link_libraries(ggml-zendnn PRIVATE OpenMP::OpenMP_CXX)

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

@@ -2,7 +2,10 @@
 
 #include "ggml-backend-impl.h"
 #include "ggml-impl.h"
-#include "ggml-cpu.h"
+
+#define GGML_COMMON_DECL_CPP
+#include "ggml-common.h"
+
 #include "zendnnl.hpp"
 
 #include <cstring>
@@ -20,6 +23,8 @@ zendnnl::common::data_type_t ggml_to_zendnn_type() {
         return zendnnl::common::data_type_t::f32;
     } else if constexpr (std::is_same_v<T, ggml_bf16_t>) {
         return zendnnl::common::data_type_t::bf16;
+    } else if constexpr (std::is_same_v<T, block_q8_0>) {
+        return zendnnl::common::data_type_t::s8;
     } else {
         return zendnnl::common::data_type_t::none;
     }
@@ -42,13 +47,25 @@ static bool ggml_zendnn_matmul(ggml_backend_zendnn_context * ctx, int64_t m, int
                                const TA * A, int64_t lda, const TB * B, int64_t ldb, TC * C,
                                int64_t ldc) {
 
-    zendnnl::lowoha::lowoha_params params;
+    zendnnl::lowoha::matmul::matmul_params params;
     params.dtypes.src = ggml_to_zendnn_type<TB>();
     params.dtypes.wei = ggml_to_zendnn_type<TA>();
     params.dtypes.dst = ggml_to_zendnn_type<TC>();
     params.num_threads = ctx->n_threads;
 
-    zendnnl::lowoha::status_t status = zendnnl::lowoha::matmul_direct(
+    zendnnl::lowoha::matmul::matmul_batch_params_t batch_params;
+
+    if constexpr (std::is_same_v<TA, block_q8_0>) {
+        params.dtypes.compute = zendnnl::common::data_type_t::s8;
+        const int64_t num_groups = k / QK8_0;
+        params.dynamic_quant = true;
+        params.quant_params.src_scale.buff = nullptr;
+        params.quant_params.src_scale.dt   = zendnnl::common::data_type_t::bf16;
+        params.quant_params.src_scale.dims = {n, num_groups};
+        params.packing.pack_format_b = 1;
+    }
+
+    zendnnl::error_handling::status_t status = zendnnl::lowoha::matmul::matmul_direct(
         'r', false, true,   // row-major, don't transpose B, transpose A (because it's column-major)
         n,                  // M: rows of B and C
         m,                  // N: cols of A^T and C
@@ -60,18 +77,18 @@ static bool ggml_zendnn_matmul(ggml_backend_zendnn_context * ctx, int64_t m, int
         0.0f,               // beta
         C, ldc,             // output C[n,m]
         true,               // is_weights_const
-        {},                 // batch_params
+        batch_params,       // batch_params
         params              // params
     );
 
-    if (status != zendnnl::lowoha::status_t::success) {
+    if (status != zendnnl::error_handling::status_t::success) {
         GGML_LOG_ERROR("%s, ZenDNN matmul failed: status=%d\n", __func__, static_cast<int>(status));
         return false;
     }
     return true;
 }
 
-static bool ggml_zendnn_sgemm(ggml_backend_zendnn_context * ctx, int64_t m, int64_t n, int64_t k,
+static bool ggml_zendnn_gemm(ggml_backend_zendnn_context * ctx, int64_t m, int64_t n, int64_t k,
                               const void * A, int64_t lda, const void * B, int64_t ldb, void * C,
                               int64_t ldc, int Atype, int Btype, int Ctype) {
 
@@ -108,6 +125,14 @@ static bool ggml_zendnn_sgemm(ggml_backend_zendnn_context * ctx, int64_t m, int6
                     (const ggml_bf16_t *)B, ldb,
                     (float *)C, ldc);
             return false;
+        case GGML_TYPE_Q8_0:
+            if (Btype != GGML_TYPE_F32 || Ctype != GGML_TYPE_F32)
+                return false;
+            return ggml_zendnn_matmul<block_q8_0, float, float>(
+                ctx, m, n, k,
+                (const block_q8_0 *)A, lda,
+                (const float *)B, ldb,
+                (float *)C, ldc);
         default:
             return false; // unsupported type
     }
@@ -122,8 +147,8 @@ static void ggml_zendnn_compute_forward_mul_mat(
 
     GGML_TENSOR_BINARY_OP_LOCALS
 
-    ggml_type         const vec_dot_type = ggml_get_type_traits_cpu(src0->type)->vec_dot_type;
-    ggml_from_float_t const from_float = ggml_get_type_traits_cpu(vec_dot_type)->from_float;
+    ggml_type         const vec_dot_type = src0->type;
+    ggml_from_float_t const from_float = ggml_get_type_traits(vec_dot_type)->from_float_ref;
 
     GGML_ASSERT(ne0 == ne01);
     GGML_ASSERT(ne1 == ne11);
@@ -145,7 +170,9 @@ static void ggml_zendnn_compute_forward_mul_mat(
     const int64_t r3 = ne13/ne03;
 
     void * work_data = ctx->work_data.get();
-    if (src1->type != vec_dot_type) {
+
+    // ZenDNN requires FP32 for dynamic quantization, so conversion is skipped
+    if (src1->type != vec_dot_type && src0->type != GGML_TYPE_Q8_0) {
         const size_t nbw1 = ggml_row_size(vec_dot_type, ne10);
         const size_t nbw2 = nbw1 * ne11;
         const size_t nbw3 = nbw2 * ne12;
@@ -171,9 +198,9 @@ static void ggml_zendnn_compute_forward_mul_mat(
 
     for (int64_t i13 = 0; i13 < ne13; i13++) {
         for (int64_t i12 = 0; i12 < ne12; i12++) {
-            const void* wdata = src1->type == vec_dot_type ? src1->data : work_data;
+            const void* wdata = (src1->type == vec_dot_type || src0->type == GGML_TYPE_Q8_0) ? src1->data : work_data;
             const size_t row_size = ggml_row_size(vec_dot_type, ne10);
-            if (!ggml_zendnn_sgemm(ctx,
+            if (!ggml_zendnn_gemm(ctx,
                                   ne01,     // m
                                   ne11,     // n
                                   ne10,     // k
@@ -184,9 +211,179 @@ static void ggml_zendnn_compute_forward_mul_mat(
                                   static_cast<char *>(dst->data) + i12*nb2 + i13*nb3,
                                   ne01,     // ldc
                                   src0->type,
-                                  vec_dot_type,
+                                  src0->type == GGML_TYPE_Q8_0 ? GGML_TYPE_F32 : vec_dot_type,
                                   dst->type))
-                GGML_ABORT("%s: ZenDNN sgemm failed\n", __func__);
+                GGML_ABORT("%s: ZenDNN gemm failed\n", __func__);
+        }
+    }
+}
+
+struct mmid_row_mapping {
+    int32_t i1;
+    int32_t i2;
+};
+
+static void ggml_zendnn_compute_forward_mul_mat_id(
+    ggml_backend_zendnn_context * ctx,
+    ggml_tensor * dst) {
+
+    const ggml_tensor * src0 = dst->src[0];  // expert weights
+    const ggml_tensor * src1 = dst->src[1];  // inputs
+    const ggml_tensor * ids  = dst->src[2];  // expert ids
+
+    GGML_TENSOR_BINARY_OP_LOCALS
+
+    // exit for no tokens to process
+    if (ne2 == 0 || ne11 == 0) {
+        return;
+    }
+
+    ggml_type         const vec_dot_type = src0->type;
+    ggml_from_float_t const from_float = ggml_get_type_traits(vec_dot_type)->from_float_ref;
+
+    // we don't support permuted src0 or src1
+    GGML_ASSERT(nb00 == ggml_type_size(src0->type));
+    GGML_ASSERT(nb10 == ggml_type_size(src1->type));
+
+    // dst cannot be transposed or permuted
+    GGML_ASSERT(nb0 == sizeof(float));
+    GGML_ASSERT(nb0 <= nb1);
+    GGML_ASSERT(nb1 <= nb2);
+    GGML_ASSERT(nb2 <= nb3);
+
+    GGML_ASSERT(ne03 == 1);
+    GGML_ASSERT(ne13 == 1);
+    GGML_ASSERT(ne3  == 1);
+
+    // row groups
+    const int n_ids = ids->ne[0]; // n_expert_used
+    const int n_as  = ne02;       // n_experts
+
+    std::vector<int64_t> matrix_row_counts(n_as, 0);
+    std::vector<std::vector<mmid_row_mapping>> matrix_rows(n_as);
+
+    int64_t max_rows = 0;
+    // group rows by expert (preprocessing step)
+    for (int64_t iid1 = 0; iid1 < ids->ne[1]; ++iid1) {
+        for (int id = 0; id < n_ids; ++id) {
+            const int32_t i02 = *(const int32_t *)((const char *)ids->data + iid1*ids->nb[1] + id*ids->nb[0]);
+
+            GGML_ASSERT(i02 >= 0 && i02 < n_as);
+
+            matrix_rows[i02].push_back({id, iid1});
+            matrix_row_counts[i02]++;
+            if (matrix_row_counts[i02] > max_rows) {
+                max_rows = matrix_row_counts[i02];
+            }
+        }
+    }
+
+    if (max_rows == 0) {
+        return; // no rows to process
+    }
+
+    const size_t row_size = ggml_row_size(vec_dot_type, ne10);
+
+    // size for converting src1 rows to vec_dot_type if needed
+    const size_t nbw1 = row_size;
+    const size_t nbw2 = nbw1 * ne11;
+    const size_t nbw3 = nbw2 * ne12;
+    const size_t src1_conv_size = (src1->type != vec_dot_type && src0->type != GGML_TYPE_Q8_0) ? ne13 * nbw3 : 0;
+
+    // For Q8_0, src1 is always F32; the gather buffer must hold F32 rows (ne10*4 bytes),
+    // not Q8_0-encoded rows (row_size ≈ ne10/32*34 bytes) — they differ by ~4x.
+    const size_t f32_row_size = (size_t)ne10 * sizeof(float);
+    const size_t gather_row_size = (src0->type == GGML_TYPE_Q8_0) ? f32_row_size : row_size;
+
+    // size for MoE gather/scatter buffers
+    const size_t wdata_cur_size = max_rows * gather_row_size;
+    const size_t dst_cur_size = max_rows * ggml_row_size(dst->type, ne01);
+
+    // allocate single buffer for all needs
+    const size_t total_size = src1_conv_size + wdata_cur_size + dst_cur_size;
+    if (ctx->work_size < total_size) {
+        ctx->work_data.reset(new char[total_size]);
+        ctx->work_size = total_size;
+    }
+
+    // partition the buffer
+    char * work_data = ctx->work_data.get();
+    char * wdata_cur = work_data + src1_conv_size;
+    char * dst_cur = wdata_cur + wdata_cur_size;
+
+    // ZenDNN requires FP32 for dynamic quantization, so conversion is skipped
+    if (src1->type != vec_dot_type && src0->type != GGML_TYPE_Q8_0) {
+        GGML_ASSERT(src1->type == GGML_TYPE_F32);
+
+        #pragma omp parallel for collapse(3) num_threads(ctx->n_threads) schedule(static)
+        for (int64_t i13 = 0; i13 < ne13; ++i13) {
+            for (int64_t i12 = 0; i12 < ne12; ++i12) {
+                for (int64_t i11 = 0; i11 < ne11; ++i11) {
+                    const float * src1_f32 = (float *)((char *)src1->data + i11*nb11 + i12*nb12 + i13*nb13);
+                    void * src1_conv = (char *)work_data + i11*nbw1 + i12*nbw2 + i13*nbw3;
+                    from_float(src1_f32, src1_conv, ne10);
+                }
+            }
+        }
+    }
+
+    const void * wdata = (src1->type == vec_dot_type || src0->type == GGML_TYPE_Q8_0) ? src1->data : work_data;
+
+    // process each expert with gather -> gemm -> scatter pattern
+    for (int64_t cur_a = 0; cur_a < n_as; ++cur_a) {
+        const int64_t cne1 = matrix_row_counts[cur_a];
+
+        if (cne1 == 0) {
+            continue;
+        }
+
+        const char * src0_cur = (const char *) src0->data + cur_a*nb02;
+
+        // gather input rows for this expert
+        #pragma omp parallel for num_threads(ctx->n_threads) schedule(static)
+        for (int64_t ir1 = 0; ir1 < cne1; ++ir1) {
+            const mmid_row_mapping & row_mapping = matrix_rows[cur_a][ir1];
+            const int64_t id = row_mapping.i1;
+            const int64_t i11 = id % ne11;
+            const int64_t i12 = row_mapping.i2;
+
+            std::memcpy(
+                wdata_cur + ir1 * gather_row_size,
+                (const char *) wdata + (i11 + i12*ne11) * gather_row_size,
+                gather_row_size
+            );
+        }
+
+        // batched gemm for all tokens in this expert
+        if (!ggml_zendnn_gemm(ctx,
+                              ne01,       // m
+                              cne1,       // n
+                              ne10,       // k
+                              src0_cur,
+                              ne00,       // lda
+                              wdata_cur,
+                              ne10,       // ldb
+                              dst_cur,
+                              ne01,       // ldc
+                              src0->type,
+                              src0->type == GGML_TYPE_Q8_0 ? GGML_TYPE_F32 : vec_dot_type,
+                              dst->type)) {
+            GGML_ABORT("%s: ZenDNN gemm failed\n", __func__);
+        }
+
+        // scatter output rows to destination
+        #pragma omp parallel for num_threads(ctx->n_threads) schedule(static)
+        for (int64_t ir1 = 0; ir1 < cne1; ++ir1) {
+            const mmid_row_mapping & row_mapping = matrix_rows[cur_a][ir1];
+            const int64_t id = row_mapping.i1;
+            const int64_t i1 = id;
+            const int64_t i2 = row_mapping.i2;
+
+            std::memcpy(
+                (char *) dst->data + i1*nb1 + i2*nb2,
+                dst_cur + ir1 * ggml_row_size(dst->type, ne01),
+                ggml_row_size(dst->type, ne01)
+            );
         }
     }
 }
@@ -211,10 +408,17 @@ static ggml_status ggml_backend_zendnn_graph_compute(ggml_backend_t backend, ggm
     for (int i = 0; i < cgraph->n_nodes; i++) {
         struct ggml_tensor * node = cgraph->nodes[i];
 
+        if ((node->flags & GGML_TENSOR_FLAG_COMPUTE) == 0) {
+            continue;
+        }
+
         switch (node->op) {
             case GGML_OP_MUL_MAT:
                 ggml_zendnn_compute_forward_mul_mat(ctx, node);
                 break;
+            case GGML_OP_MUL_MAT_ID:
+                ggml_zendnn_compute_forward_mul_mat_id(ctx, node);
+                break;
             case GGML_OP_NONE:
             case GGML_OP_RESHAPE:
             case GGML_OP_VIEW:
@@ -237,6 +441,8 @@ static struct ggml_backend_i ggml_backend_zendnn_i = {
     /* .free                    = */ ggml_backend_zendnn_free,
     /* .set_tensor_async        = */ NULL,
     /* .get_tensor_async        = */ NULL,
+    /* .set_tensor_2d_async     = */ NULL,
+    /* .get_tensor_2d_async     = */ NULL,
     /* .cpy_tensor_async        = */ NULL,
     /* .synchronize             = */ NULL,
     /* .graph_plan_create       = */ NULL,
@@ -348,6 +554,12 @@ static ggml_backend_buffer_t ggml_backend_zendnn_device_buffer_from_host_ptr(ggm
     GGML_UNUSED(max_tensor_size);
 }
 
+static bool ggml_zendnn_adaptive_fallback_enabled() {
+    static const bool enabled = std::getenv("GGML_ZENDNN_ADAPTIVE_FALLBACK") == nullptr ||
+                                std::atoi(std::getenv("GGML_ZENDNN_ADAPTIVE_FALLBACK")) != 0;
+    return enabled;
+}
+
 static bool ggml_backend_zendnn_device_supports_op(ggml_backend_dev_t dev, const struct ggml_tensor * op) {
     switch (op->op) {
         case GGML_OP_NONE:
@@ -358,6 +570,7 @@ static bool ggml_backend_zendnn_device_supports_op(ggml_backend_dev_t dev, const
             return true;
 
         case GGML_OP_MUL_MAT:
+        case GGML_OP_MUL_MAT_ID:
         {
             const ggml_tensor * weights = op->src[0];
             const ggml_tensor * inputs = op->src[1];
@@ -365,15 +578,39 @@ static bool ggml_backend_zendnn_device_supports_op(ggml_backend_dev_t dev, const
             const int64_t ne10 = inputs->ne[0];
             const int64_t ne0 = op->ne[0];
             const int64_t ne1 = op->ne[1];
-
             const int64_t min_batch = 1;
-            if (!ggml_is_contiguous(weights) || !ggml_is_contiguous(inputs) ||
-                ne0 < min_batch || ne1 < min_batch || ne10 < min_batch) {
+
+            if(!ggml_is_contiguous(weights) || !ggml_is_contiguous(inputs)) {
+                return false;
+            }
+
+            if (ggml_zendnn_adaptive_fallback_enabled()) {
+                const int64_t K = inputs->ne[0];
+                const int64_t N = (inputs->ne[1]*inputs->ne[2]*inputs->ne[3]);
+                const int64_t M = weights->ne[1];
+                if(K <= 256 || N <= 128 || M <= 96) {
                     return false;
+                }
+            }
+            else if (ne0 < min_batch || ne1 < min_batch || ne10 < min_batch) {
+                return false;
+            }
+
+            // MUL_MAT_ID performs best with a moderate number of experts due to its
+            // gather + batched matmul + scatter approach. Future versions will leverage
+            // ZenDNN's grouped_gemm for better scalability with larger expert counts:
+            // https://github.com/amd/ZenDNN/blob/main/docs/operator/lowoha_group_gemm_operator.md
+            if (op->op == GGML_OP_MUL_MAT_ID) {
+                const int64_t n_experts = weights->ne[2];
+                const int64_t max_experts = 32;
+                if (n_experts > max_experts) {
+                    return false;
+                }
             }
             switch (weights->type) {
                 case GGML_TYPE_F32:
                 case GGML_TYPE_BF16:
+                case GGML_TYPE_Q8_0:
                     return true;
                 default:
                     return false;