whispercpp 1.3.7 → 1.3.8

data/ext/sources/ggml/src/ggml-openvino/utils.h

@@ -1,4 +1,3 @@
-#include "ggml-backend-impl.h"
 #include "ggml-decoder.h"
 #include "ggml-impl.h"
 
@@ -45,6 +44,7 @@ struct graph_key_hash {
 
 struct decoder_runtime_ctx {
     decoder_runtime_ctx(std::shared_ptr<std::mutex> mutex) : mutex(std::move(mutex)) {}
+
     std::shared_ptr<std::mutex> mutex;
     std::shared_ptr<GgmlOvDecoder> ptr;
 };
@@ -64,11 +64,7 @@ struct ov_runtime_context {
     std::map<std::string, std::string> kv_state_input_name_map;
     std::atomic<int> backend_count;
 
-    ov_runtime_context() :
-        device("CPU"),
-        stateful(false),
-        stateful_kv_size(0),
-        backend_count(0) {}
+    ov_runtime_context() : device("CPU"), stateful(false), stateful_kv_size(0), backend_count(0) {}
 
     void clear_caches() {
         std::lock_guard<std::mutex> lock(ctx_mutex);
@@ -87,6 +83,8 @@ enum ggml_status ov_graph_compute_static(struct ggml_cgraph * cgraph, std::share
 
 size_t checksum(const void * data, size_t size);
 
+bool save_ggml_tensor_data_to_txt(const ggml_tensor * tensor, const std::string & file_path);
+
 void print_input_tensor_info(const std::string & name, const ov::Tensor & tensor);
 
 void print_output_tensor_info(const std::string & name, const ov::Tensor & tensor, const void * output_dst);
@@ -117,8 +115,6 @@ std::vector<T> pad_input(const ggml_tensor * tensor, size_t padded_rows, size_t
                         padded_rows, padded_cols, pad_value);
 }
 
-void set_zero_diagonal(std::vector<float> & matrix, size_t rows, size_t cols);
-
 const ggml_tensor * get_inp_pos_tensor(struct ggml_cgraph * cgraph);
 
 bool get_is_prefill(const ggml_tensor * inp_pos);
@@ -137,6 +133,13 @@ ov::Tensor create_ov_output_tensor(std::shared_ptr<GgmlOvDecoder> ggml_decoder,
 
 bool is_naive(struct ggml_cgraph * cgraph);
 
+/**
+ * @brief Heuristically checks whether the given computation graph is a split-model fragment.
+ * @param cgraph Pointer to the GGML computation graph to analyze.
+ * @return true if the graph is identified as split; otherwise false.
+ */
+bool is_model_splitted(struct ggml_cgraph * cgraph);
+
 enum ggml_status naive_compute(struct ggml_cgraph * cgraph,
                                ov::Core & core,
                                const std::string & device,

data/ext/sources/ggml/src/ggml-quants.c

@@ -71,6 +71,44 @@ void quantize_row_q1_0_ref(const float * GGML_RESTRICT x, block_q1_0 * GGML_REST
     }
 }
 
+void quantize_row_q2_0_ref(const float * GGML_RESTRICT x, block_q2_0 * GGML_RESTRICT y, int64_t k) {
+    static const int qk = QK2_0;
+
+    assert(k % qk == 0);
+
+    const int nb = k / qk;
+
+    for (int i = 0; i < nb; i++) {
+        // Compute scale as max absolute value in the block
+        float amax = 0.0f;
+        for (int j = 0; j < qk; j++) {
+            const float a = fabsf(x[i*qk + j]);
+            if (a > amax) amax = a;
+        }
+        const float d = amax;
+        const float id = d > 0.0f ? 1.0f / d : 0.0f;
+
+        y[i].d = GGML_FP32_TO_FP16(d);
+
+        // Clear quant bytes
+        for (int j = 0; j < qk / 4; ++j) {
+            y[i].qs[j] = 0;
+        }
+
+        // Encode 2-bit values: round(w/d) clamped to [-1, 2], then add 1
+        // 00 (-1) = -scale, 01 (0) = 0, 10 (+1) = +scale, 11 (+2) = 2*scale
+        for (int j = 0; j < qk; ++j) {
+            const float w = x[i*qk + j];
+            int q = (int)roundf(w * id) + 1;
+            if (q < 0) q = 0;
+            if (q > 3) q = 3;
+            const int byte_index = j / 4;
+            const int bit_offset = (j % 4) * 2;
+            y[i].qs[byte_index] |= ((uint8_t)q << bit_offset);
+        }
+    }
+}
+
 // reference implementation for deterministic creation of model files
 void quantize_row_q4_0_ref(const float * GGML_RESTRICT x, block_q4_0 * GGML_RESTRICT y, int64_t k) {
     static const int qk = QK4_0;
@@ -398,6 +436,26 @@ void dequantize_row_q1_0(const block_q1_0 * GGML_RESTRICT x, float * GGML_RESTRI
     }
 }
 
+void dequantize_row_q2_0(const block_q2_0 * GGML_RESTRICT x, float * GGML_RESTRICT y, int64_t k) {
+    static const int qk = QK2_0;
+
+    assert(k % qk == 0);
+
+    const int nb = k / qk;
+
+    for (int i = 0; i < nb; i++) {
+        const float d = GGML_FP16_TO_FP32(x[i].d);
+
+        for (int j = 0; j < qk; ++j) {
+            const int byte_index = j / 4;
+            const int bit_offset = (j % 4) * 2;
+            const uint8_t q = (x[i].qs[byte_index] >> bit_offset) & 0x03;
+            // 00=-1, 01=0, 10=+1, 11=+2
+            y[i*qk + j] = ((int)q - 1) * d;
+        }
+    }
+}
+
 void dequantize_row_q4_0(const block_q4_0 * GGML_RESTRICT x, float * GGML_RESTRICT y, int64_t k) {
     static const int qk = QK4_0;
 
@@ -2052,6 +2110,20 @@ size_t quantize_q1_0(const float * GGML_RESTRICT src, void * GGML_RESTRICT dst,
     return nrow * row_size;
 }
 
+size_t quantize_q2_0(const float * GGML_RESTRICT src, void * GGML_RESTRICT dst, int64_t nrow, int64_t n_per_row, const float * quant_weights) {
+    if (!quant_weights) {
+        quantize_row_q2_0_ref(src, dst, (int64_t)nrow*n_per_row);
+        return nrow * ggml_row_size(GGML_TYPE_Q2_0, n_per_row);
+    }
+    size_t row_size = ggml_row_size(GGML_TYPE_Q2_0, n_per_row);
+    char * qrow = (char *)dst;
+    for (int64_t row = 0; row < nrow; ++row) {
+        quantize_row_q2_0_ref(src, (block_q2_0*)qrow, n_per_row);
+        src += n_per_row;
+        qrow += row_size;
+    }
+    return nrow * row_size;
+}
 
 size_t quantize_q4_0(const float * GGML_RESTRICT src, void * GGML_RESTRICT dst, int64_t nrow, int64_t n_per_row, const float * quant_weights) {
     if (!quant_weights) {
@@ -5461,6 +5533,10 @@ bool ggml_validate_row_data(enum ggml_type type, const void * data, size_t nbyte
             {
                 VALIDATE_ROW_DATA_D_F16_IMPL(block_q1_0, data, nb);
             } break;
+        case GGML_TYPE_Q2_0:
+            {
+                VALIDATE_ROW_DATA_D_F16_IMPL(block_q2_0, data, nb);
+            } break;
         case GGML_TYPE_Q4_0:
             {
                 VALIDATE_ROW_DATA_D_F16_IMPL(block_q4_0, data, nb);

data/ext/sources/ggml/src/ggml-quants.h

@@ -15,6 +15,7 @@ extern "C" {
 
 // Quantization
 GGML_API void quantize_row_q1_0_ref(const float * GGML_RESTRICT x, block_q1_0 * GGML_RESTRICT y, int64_t k);
+GGML_API void quantize_row_q2_0_ref(const float * GGML_RESTRICT x, block_q2_0 * GGML_RESTRICT y, int64_t k);
 GGML_API void quantize_row_q4_0_ref(const float * GGML_RESTRICT x, block_q4_0 * GGML_RESTRICT y, int64_t k);
 GGML_API void quantize_row_q4_1_ref(const float * GGML_RESTRICT x, block_q4_1 * GGML_RESTRICT y, int64_t k);
 GGML_API void quantize_row_q5_0_ref(const float * GGML_RESTRICT x, block_q5_0 * GGML_RESTRICT y, int64_t k);
@@ -43,6 +44,7 @@ GGML_API void quantize_row_iq2_s_ref  (const float * GGML_RESTRICT x, block_iq2_
 
 // Dequantization
 GGML_API void dequantize_row_q1_0(const block_q1_0 * GGML_RESTRICT x, float * GGML_RESTRICT y, int64_t k);
+GGML_API void dequantize_row_q2_0(const block_q2_0 * GGML_RESTRICT x, float * GGML_RESTRICT y, int64_t k);
 GGML_API void dequantize_row_q4_0(const block_q4_0 * GGML_RESTRICT x, float * GGML_RESTRICT y, int64_t k);
 GGML_API void dequantize_row_q4_1(const block_q4_1 * GGML_RESTRICT x, float * GGML_RESTRICT y, int64_t k);
 GGML_API void dequantize_row_q5_0(const block_q5_0 * GGML_RESTRICT x, float * GGML_RESTRICT y, int64_t k);
@@ -93,6 +95,7 @@ GGML_API size_t quantize_q4_K(const float * GGML_RESTRICT src, void * GGML_RESTR
 GGML_API size_t quantize_q5_K(const float * GGML_RESTRICT src, void * GGML_RESTRICT dst, int64_t nrows, int64_t n_per_row, const float * imatrix);
 GGML_API size_t quantize_q6_K(const float * GGML_RESTRICT src, void * GGML_RESTRICT dst, int64_t nrows, int64_t n_per_row, const float * imatrix);
 GGML_API size_t quantize_q1_0(const float * GGML_RESTRICT src, void * GGML_RESTRICT dst, int64_t nrows, int64_t n_per_row, const float * imatrix);
+GGML_API size_t quantize_q2_0(const float * GGML_RESTRICT src, void * GGML_RESTRICT dst, int64_t nrows, int64_t n_per_row, const float * imatrix);
 GGML_API size_t quantize_q4_0(const float * GGML_RESTRICT src, void * GGML_RESTRICT dst, int64_t nrows, int64_t n_per_row, const float * imatrix);
 GGML_API size_t quantize_q4_1(const float * GGML_RESTRICT src, void * GGML_RESTRICT dst, int64_t nrows, int64_t n_per_row, const float * imatrix);
 GGML_API size_t quantize_q5_0(const float * GGML_RESTRICT src, void * GGML_RESTRICT dst, int64_t nrows, int64_t n_per_row, const float * imatrix);

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

@@ -39,8 +39,8 @@ if (WIN32)
         set(CMAKE_CXX_COMPILER "icx")
         set(CMAKE_CXX_COMPILER_ID "IntelLLVM")
     endif()
-    # Level Zero SDK path for Windows (only when GGML_SYCL_SUPPORT_LEVEL_ZERO is enabled)
-    if(GGML_SYCL_SUPPORT_LEVEL_ZERO)
+    # Level Zero SDK path for Windows (only when GGML_SYCL_SUPPORT_LEVEL_ZERO_API is enabled)
+    if(GGML_SYCL_SUPPORT_LEVEL_ZERO_API)
         if(DEFINED ENV{LEVEL_ZERO_V1_SDK_PATH})
             set(LEVEL_ZERO_V1_SDK_PATH $ENV{LEVEL_ZERO_V1_SDK_PATH})
             if(EXISTS "${LEVEL_ZERO_V1_SDK_PATH}")
@@ -105,8 +105,8 @@ endif()
 
 target_compile_options(ggml-sycl PRIVATE "-Wno-narrowing")
 
-message(STATUS "GGML_SYCL_SUPPORT_LEVEL_ZERO ${GGML_SYCL_SUPPORT_LEVEL_ZERO}")
-if (GGML_SYCL_SUPPORT_LEVEL_ZERO)
+message(STATUS "GGML_SYCL_SUPPORT_LEVEL_ZERO_API ${GGML_SYCL_SUPPORT_LEVEL_ZERO_API}")
+if (GGML_SYCL_SUPPORT_LEVEL_ZERO_API)
     # Link against Level Zero loader for direct device memory allocation.
     # Avoids sycl::malloc_device triggering DMA-buf/TTM system RAM staging
     # in the xe kernel driver during multi-GPU inference.
@@ -114,7 +114,7 @@ if (GGML_SYCL_SUPPORT_LEVEL_ZERO)
     find_library(ZE_LOADER_LIB ze_loader HINTS ${ONEAPI_ROOT}/lib ${LEVEL_ZERO_V1_SDK_LIB_PATH} ENV LD_LIBRARY_PATH)
     if(ZE_LOADER_LIB AND LEVEL_ZERO_INCLUDE_DIR)
         target_link_libraries(ggml-sycl PRIVATE ${ZE_LOADER_LIB})
-        target_compile_definitions(ggml-sycl PRIVATE GGML_SYCL_SUPPORT_LEVEL_ZERO)
+        target_compile_definitions(ggml-sycl PRIVATE GGML_SYCL_SUPPORT_LEVEL_ZERO_API)
         message(STATUS "Level Zero loader found: ${ZE_LOADER_LIB}")
         message(STATUS "Level Zero headers found: ${LEVEL_ZERO_INCLUDE_DIR}")
     else()

data/ext/sources/ggml/src/ggml-sycl/backend.hpp

@@ -14,9 +14,11 @@
 #define GGML_SYCL_BACKEND_HPP
 
 #include "binbcast.hpp"
+#include "col2im-1d.hpp"
 #include "common.hpp"
 #include "concat.hpp"
 #include "conv.hpp"
+#include "conv3d.hpp"
 #include "convert.hpp"
 #include "count-equal.hpp"
 #include "cpy.hpp"

data/ext/sources/ggml/src/ggml-sycl/binbcast.cpp

@@ -287,6 +287,18 @@ inline void ggml_sycl_op_bin_bcast(ggml_backend_sycl_context & ctx, const ggml_t
              ne10, ne11, ne12, ne13, ne0, ne1, ne2, ne3, nb00, nb01, nb02, nb03, nb10, nb11, nb12, nb13, nb0, nb1, nb2,
              nb3, ggml_is_contiguous(src0), ggml_is_contiguous(src1), ggml_is_permuted(src0), ggml_is_permuted(src1),
              main_stream);
+#ifdef GGML_SYCL_HAS_BF16
+    } else if (src0->type == GGML_TYPE_BF16 && src1->type == GGML_TYPE_BF16 && dst->type == GGML_TYPE_BF16) {
+        op()((const sycl::ext::oneapi::bfloat16 *) src0->data, (const sycl::ext::oneapi::bfloat16 *) src1->data,
+             (sycl::ext::oneapi::bfloat16 *) dst->data, ne00, ne01, ne02, ne03, ne10, ne11, ne12, ne13, ne0, ne1, ne2,
+             ne3, nb00, nb01, nb02, nb03, nb10, nb11, nb12, nb13, nb0, nb1, nb2, nb3, ggml_is_contiguous(src0),
+             ggml_is_contiguous(src1), ggml_is_permuted(src0), ggml_is_permuted(src1), main_stream);
+    } else if (src0->type == GGML_TYPE_BF16 && src1->type == GGML_TYPE_F32 && dst->type == GGML_TYPE_BF16) {
+        op()((const sycl::ext::oneapi::bfloat16 *) src0->data, (const float *) src1->data,
+             (sycl::ext::oneapi::bfloat16 *) dst->data, ne00, ne01, ne02, ne03, ne10, ne11, ne12, ne13, ne0, ne1, ne2,
+             ne3, nb00, nb01, nb02, nb03, nb10, nb11, nb12, nb13, nb0, nb1, nb2, nb3, ggml_is_contiguous(src0),
+             ggml_is_contiguous(src1), ggml_is_permuted(src0), ggml_is_permuted(src1), main_stream);
+#endif
     } else {
         fprintf(stderr, "%s: unsupported types: dst: %s, src0: %s, src1: %s\n", __func__, ggml_type_name(dst->type),
                 ggml_type_name(src0->type), ggml_type_name(src1->type));

data/ext/sources/ggml/src/ggml-sycl/col2im-1d.cpp

@@ -0,0 +1,102 @@
+#include "col2im-1d.hpp"
+
+template <typename T>
+static void col2im_1d_sycl(
+        const T * col,
+        T * dst,
+        const int T_in,
+        const sycl::uint3 T_out_fd,
+        const int K,
+        const int K_OC,
+        const int32_t s0,
+        const int32_t p0,
+        const int total,
+        dpct::queue_ptr stream) {
+
+    const uint32_t block_size = SYCL_COL2IM_1D_BLOCK_SIZE;
+    const uint32_t num_blocks = (uint32_t) ((total + block_size - 1) / block_size);
+
+    stream->parallel_for(
+        sycl::nd_range<3>(
+            sycl::range<3>(1, 1, num_blocks * block_size),
+            sycl::range<3>(1, 1, block_size)),
+        [=](sycl::nd_item<3> item_ct1) {
+            const int idx = (int) item_ct1.get_global_id(2);
+            if (idx >= total) {
+                return;
+            }
+
+            const sycl::uint2 qr = fast_div_modulo((uint32_t) idx, T_out_fd);
+            const int oc    = (int) qr.x();
+            const int t_out = (int) qr.y();
+            const int t_abs = t_out + p0;
+
+            int t_in_min = (t_abs - K + s0) / s0;
+            if (t_in_min < 0) {
+                t_in_min = 0;
+            }
+            int t_in_max = t_abs / s0;
+            if (t_in_max >= T_in) {
+                t_in_max = T_in - 1;
+            }
+
+            float sum = 0.0f;
+            for (int t_in = t_in_min; t_in <= t_in_max; ++t_in) {
+                const int k = t_abs - t_in * s0;
+                sum += static_cast<float>(col[(oc * K + k) + t_in * K_OC]);
+            }
+
+            dst[idx] = static_cast<T>(sum);
+        });
+}
+
+void ggml_sycl_op_col2im_1d(ggml_backend_sycl_context & ctx, ggml_tensor * dst) {
+    const ggml_tensor * src0 = dst->src[0];
+
+    GGML_ASSERT(src0 != nullptr);
+    GGML_ASSERT(ggml_is_contiguous(src0));
+    GGML_ASSERT(src0->type == dst->type);
+
+    const int32_t s0 = ((const int32_t *) dst->op_params)[0];
+    const int32_t OC = ((const int32_t *) dst->op_params)[1];
+    const int32_t p0 = ((const int32_t *) dst->op_params)[2];
+
+    const int K_OC  = (int) src0->ne[0];
+    const int T_in  = (int) src0->ne[1];
+    const int K     = K_OC / OC;
+    const int T_out = (int) dst->ne[0];
+
+    GGML_ASSERT(OC > 0);
+    GGML_ASSERT(K_OC % OC == 0);
+
+    const sycl::uint3 T_out_fd = init_fastdiv_values((uint32_t) T_out);
+
+    const int total = T_out * OC;
+
+    dpct::queue_ptr stream = ctx.stream();
+
+    switch (src0->type) {
+        case GGML_TYPE_F32:
+            col2im_1d_sycl<float>(
+                (const float *) src0->data,
+                (float *) dst->data,
+                T_in, T_out_fd, K, K_OC, s0, p0, total, stream);
+            break;
+        case GGML_TYPE_F16:
+            col2im_1d_sycl<sycl::half>(
+                (const sycl::half *) src0->data,
+                (sycl::half *) dst->data,
+                T_in, T_out_fd, K, K_OC, s0, p0, total, stream);
+            break;
+#ifdef GGML_SYCL_HAS_BF16
+        case GGML_TYPE_BF16:
+            col2im_1d_sycl<sycl::ext::oneapi::bfloat16>(
+                (const sycl::ext::oneapi::bfloat16 *) src0->data,
+                (sycl::ext::oneapi::bfloat16 *) dst->data,
+                T_in, T_out_fd, K, K_OC, s0, p0, total, stream);
+            break;
+#endif
+        default:
+            GGML_ABORT("col2im_1d: unsupported type %d", src0->type);
+    }
+}

data/ext/sources/ggml/src/ggml-sycl/col2im-1d.hpp

@@ -0,0 +1,8 @@
+#ifndef GGML_SYCL_COL2IM_1D_HPP
+#define GGML_SYCL_COL2IM_1D_HPP
+
+#include "common.hpp"
+
+void ggml_sycl_op_col2im_1d(ggml_backend_sycl_context & ctx, ggml_tensor * dst);
+
+#endif // GGML_SYCL_COL2IM_1D_HPP

data/ext/sources/ggml/src/ggml-sycl/common.cpp

@@ -12,7 +12,7 @@
 
 #include "common.hpp"
 #include <sycl/backend.hpp>
-#ifdef GGML_SYCL_SUPPORT_LEVEL_ZERO
+#ifdef GGML_SYCL_SUPPORT_LEVEL_ZERO_API
 #include <level_zero/ze_api.h>
 #endif
 
@@ -59,7 +59,7 @@ bool gpu_has_xmx(sycl::device &dev) {
     return dev.has(sycl::aspect::ext_intel_matrix);
 }
 
-static int ggml_sycl_get_env(const char *env_name, int default_val) {
+int ggml_sycl_get_env(const char *env_name, int default_val) {
     char *user_device_string = getenv(env_name);
     int user_number = default_val;
 
@@ -84,9 +84,9 @@ int64_t downsample_sycl_global_range(int64_t accumulate_block_num, int64_t block
   return sycl_down_blk_size;
 }
 
-#ifdef GGML_SYCL_SUPPORT_LEVEL_ZERO
+#ifdef GGML_SYCL_SUPPORT_LEVEL_ZERO_API
 static bool ggml_sycl_use_level_zero_device_alloc(sycl::queue &q) {
-    return ggml_sycl_get_env("GGML_SYCL_ENABLE_LEVEL_ZERO", 1) &&
+    return g_ggml_sycl_use_level_zero_api &&
         q.get_device().is_gpu() &&
         q.get_backend() == sycl::backend::ext_oneapi_level_zero;
 }
@@ -94,10 +94,8 @@ static bool ggml_sycl_use_level_zero_device_alloc(sycl::queue &q) {
 
 // Use Level Zero zeMemAllocDevice to avoid sycl::malloc_device triggering
 // DMA-buf/TTM system RAM staging in the xe kernel driver during multi-GPU inference.
-// The decision is made from the queue and runtime env because large buffers can be
-// allocated before ggml_check_sycl() initializes g_ggml_sycl_enable_level_zero.
 void * ggml_sycl_malloc_device(size_t size, sycl::queue &q) {
-#ifdef GGML_SYCL_SUPPORT_LEVEL_ZERO
+#ifdef GGML_SYCL_SUPPORT_LEVEL_ZERO_API
     if (ggml_sycl_use_level_zero_device_alloc(q)) {
         void *ptr = nullptr;
         auto ze_ctx = sycl::get_native<sycl::backend::ext_oneapi_level_zero>(q.get_context());
@@ -129,7 +127,7 @@ void * ggml_sycl_malloc_device(size_t size, sycl::queue &q) {
 
 void ggml_sycl_free_device(void *ptr, sycl::queue &q) {
     if (!ptr) return;
-#ifdef GGML_SYCL_SUPPORT_LEVEL_ZERO
+#ifdef GGML_SYCL_SUPPORT_LEVEL_ZERO_API
     if (ggml_sycl_use_level_zero_device_alloc(q)) {
         auto ze_ctx = sycl::get_native<sycl::backend::ext_oneapi_level_zero>(q.get_context());
         zeMemFree(ze_ctx, ptr);

data/ext/sources/ggml/src/ggml-sycl/common.hpp

@@ -59,9 +59,10 @@ void ggml_sycl_host_free(void* ptr);
 
 
 extern int g_ggml_sycl_debug;
-extern int g_ggml_sycl_disable_optimize;
+extern int g_ggml_sycl_enable_optimize;
 extern int g_ggml_sycl_prioritize_dmmv;
 extern int g_ggml_sycl_enable_flash_attention;
+extern int g_ggml_sycl_dev2dev_memcpy;
 
 
 #if defined(__clang__) && __has_builtin(__builtin_expect)
@@ -126,6 +127,11 @@ enum ggml_sycl_backend_gpu_mode {
   SYCL_MUL_GPU_MODE
 };
 
+enum ggml_sycl_dev2dev_memcpy_mode {
+  DEV2DEV_MEMCPY_SYCL = 0,
+  DEV2DEV_MEMCPY_L0 = 1,
+};
+
 static_assert(sizeof(sycl::half) == sizeof(ggml_fp16_t), "wrong fp16 size");
 
 static void crash() {
@@ -225,10 +231,12 @@ struct sycl_device_info {
     int max_wg_per_cu; // max work groups per compute unit - refer to
                        // cudaOccupancyMaxActiveBlocksPerMultiprocessor
     bool    vmm;                // virtual memory support
+    bool    l0_discrete_gpu;    // Level Zero backend and not an integrated GPU
     size_t  vmm_granularity;    // granularity of virtual memory
     size_t  total_vram;
     sycl_hw_info hw_info;
     optimize_feature opt_feature;
+    bool    usm_system_support; // support for USM system allocations
 };
 
 
@@ -316,12 +324,17 @@ struct ggml_tensor_extra_gpu {
   optimize_feature optimized_feature;
 };
 
-extern int g_ggml_sycl_enable_level_zero;
+extern int g_ggml_sycl_use_level_zero_api;
 void * ggml_sycl_malloc_device(size_t size, sycl::queue &q);
 void ggml_sycl_free_device(void *ptr, sycl::queue &q);
 
 void release_extra_gpu(ggml_tensor_extra_gpu * extra, std::vector<queue_ptr> streams={});
 
+struct mmid_row_mapping {
+    int32_t i1;
+    int32_t i2;
+};
+
 namespace sycl_ex = sycl::ext::oneapi::experimental;
 struct ggml_backend_sycl_context {
     int device;
@@ -419,6 +432,8 @@ struct ggml_backend_sycl_context {
 
     std::unique_ptr<ggml_sycl_pool> host_pools[GGML_SYCL_MAX_DEVICES];
 
+    std::vector<mmid_row_mapping> mmid_row_mapping_host;
+
     static std::unique_ptr<ggml_sycl_pool> new_pool_for_device(queue_ptr qptr, int device);
 
     static std::unique_ptr<ggml_sycl_pool> new_pool_for_host(queue_ptr qptr, int device);
@@ -644,6 +659,8 @@ constexpr size_t ceil_div(const size_t m, const size_t n) {
 
 bool gpu_has_xmx(sycl::device &dev);
 
+int ggml_sycl_get_env(const char *env_name, int default_val);
+
 template <int N, class T> std::string debug_get_array_str(const std::string & prefix, const T array[N]) {
     if (LIKELY(!g_ggml_sycl_debug)) {
         return "";

data/ext/sources/ggml/src/ggml-sycl/concat.cpp

@@ -10,6 +10,8 @@
 // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
 //
 
+#include "ggml.h"
+
 #include "concat.hpp"
 
 static inline size_t elem_size(ggml_type t) {
@@ -192,11 +194,29 @@ void ggml_sycl_op_concat(ggml_backend_sycl_context & ctx, ggml_tensor *dst) {
     case GGML_TYPE_F32:
         concat_impl_sycl<float>(ctx, dst);
         break;
+    case GGML_TYPE_F16:
+        concat_impl_sycl<sycl::half>(ctx, dst);
+        break;
+#ifdef GGML_SYCL_HAS_BF16
+    case GGML_TYPE_BF16:
+        concat_impl_sycl<sycl::ext::oneapi::bfloat16>(ctx, dst);
+        break;
+#endif
     case GGML_TYPE_I32:
         concat_impl_sycl<int32_t>(ctx, dst);
         break;
+    case GGML_TYPE_I16:
+        concat_impl_sycl<int16_t>(ctx, dst);
+        break;
+    case GGML_TYPE_I64:
+        concat_impl_sycl<int64_t>(ctx, dst);
+        break;
+    case GGML_TYPE_I8:
+        concat_impl_sycl<int8_t>(ctx, dst);
+        break;
     default:
-    GGML_ASSERT(false && "ggml_sycl_op_concat: unsupported type");
+        fprintf(stderr, "%s: unsupported types: dst: %s\n", __func__, ggml_type_name(dst->type));
+        GGML_ASSERT(false);
     break;
     }
 }

data/ext/sources/ggml/src/ggml-sycl/conv2d-dw.cpp

@@ -0,0 +1,158 @@
+#include "conv2d-dw.hpp"
+
+struct conv2d_dw_params {
+    int in_w, in_h;
+    int out_w, out_h;
+    int kernel_w, kernel_h;
+    int stride_x, stride_y;
+    int padding_x, padding_y;
+    int dilation_x, dilation_y;
+    int channels, batches;
+};
+
+struct conv2d_dw_kernel_bounds {
+    int y_min, y_max;
+    int x_min, x_max;
+};
+
+static inline conv2d_dw_kernel_bounds dw_calculate_kernel_bounds(int out_x, int out_y,
+                                                                  const conv2d_dw_params & p) {
+    conv2d_dw_kernel_bounds bounds;
+    bounds.y_min = sycl::max(0, (p.padding_y - out_y * p.stride_y + p.dilation_y - 1) / p.dilation_y);
+    bounds.y_max = sycl::min(p.kernel_h,
+                             (p.in_h + p.padding_y - out_y * p.stride_y + p.dilation_y - 1) / p.dilation_y);
+    bounds.x_min = sycl::max(0, (p.padding_x - out_x * p.stride_x + p.dilation_x - 1) / p.dilation_x);
+    bounds.x_max = sycl::min(p.kernel_w,
+                             (p.in_w + p.padding_x - out_x * p.stride_x + p.dilation_x - 1) / p.dilation_x);
+    return bounds;
+}
+
+static inline int dw_calculate_input_coord(int out_coord, int kern_coord, int stride, int dilation, int padding) {
+    return out_coord * stride + kern_coord * dilation - padding;
+}
+
+// whcn layout: input/output stored as [N, C, H, W]
+struct dw_whcn_layout {
+    static int input_index(int n, int c, int y, int x, const conv2d_dw_params & p) {
+        return n * (p.channels * p.in_w * p.in_h) + c * p.in_w * p.in_h + y * p.in_w + x;
+    }
+    static int kernel_index(int c, int ky, int kx, const conv2d_dw_params & p) {
+        return c * p.kernel_h * p.kernel_w + ky * p.kernel_w + kx;
+    }
+    static int output_index(int n, int c, int y, int x, const conv2d_dw_params & p) {
+        return n * (p.channels * p.out_w * p.out_h) + c * p.out_w * p.out_h + y * p.out_w + x;
+    }
+    static void unpack_indices(int global_idx, const conv2d_dw_params & p,
+                               int & n, int & c, int & out_y, int & out_x) {
+        out_x  = global_idx % p.out_w;
+        out_y  = (global_idx / p.out_w) % p.out_h;
+        c      = (global_idx / (p.out_w * p.out_h)) % p.channels;
+        n      = global_idx / (p.out_w * p.out_h * p.channels);
+    }
+};
+
+// cwhn layout: input/output stored as [N, H, W, C]
+struct dw_cwhn_layout {
+    static int input_index(int n, int c, int y, int x, const conv2d_dw_params & p) {
+        return n * (p.channels * p.in_w * p.in_h) + (y * p.in_w + x) * p.channels + c;
+    }
+    static int kernel_index(int c, int ky, int kx, const conv2d_dw_params & p) {
+        return (ky * p.kernel_w + kx) * p.channels + c;
+    }
+    static int output_index(int n, int c, int y, int x, const conv2d_dw_params & p) {
+        return n * (p.channels * p.out_w * p.out_h) + y * (p.out_w * p.channels) + x * p.channels + c;
+    }
+    static void unpack_indices(int global_idx, const conv2d_dw_params & p,
+                               int & n, int & c, int & out_y, int & out_x) {
+        c      = global_idx % p.channels;
+        out_x  = (global_idx / p.channels) % p.out_w;
+        out_y  = (global_idx / (p.channels * p.out_w)) % p.out_h;
+        n      = global_idx / (p.channels * p.out_w * p.out_h);
+    }
+};
+
+template <typename Layout>
+static void conv2d_dw_kernel(const float * input, const float * kernel, float * output,
+                             const conv2d_dw_params p, const sycl::nd_item<3> & item_ct1) {
+    const int global_idx     = item_ct1.get_local_id(2) +
+                               item_ct1.get_group(2) * item_ct1.get_local_range(2);
+    const int total_elements = p.batches * p.channels * p.out_h * p.out_w;
+
+    if (global_idx >= total_elements) {
+        return;
+    }
+
+    int n, c, out_y, out_x;
+    Layout::unpack_indices(global_idx, p, n, c, out_y, out_x);
+
+    float acc = 0.0f;
+    const conv2d_dw_kernel_bounds bounds = dw_calculate_kernel_bounds(out_x, out_y, p);
+
+    for (int ky = bounds.y_min; ky < bounds.y_max; ++ky) {
+        const int in_y = dw_calculate_input_coord(out_y, ky, p.stride_y, p.dilation_y, p.padding_y);
+        for (int kx = bounds.x_min; kx < bounds.x_max; ++kx) {
+            const int in_x = dw_calculate_input_coord(out_x, kx, p.stride_x, p.dilation_x, p.padding_x);
+            acc += input[Layout::input_index(n, c, in_y, in_x, p)] *
+                   kernel[Layout::kernel_index(c, ky, kx, p)];
+        }
+    }
+
+    output[Layout::output_index(n, c, out_y, out_x, p)] = acc;
+}
+
+template <typename Layout>
+static void conv2d_dw_sycl(const float * x_d, const float * w_d, float * y_d,
+                            const conv2d_dw_params p, const queue_ptr & stream) {
+    const int total      = p.batches * p.channels * p.out_h * p.out_w;
+    const int num_blocks = (total + SYCL_CONV2D_DW_BLOCK_SIZE - 1) / SYCL_CONV2D_DW_BLOCK_SIZE;
+    const sycl::range<3> block_dims(1, 1, SYCL_CONV2D_DW_BLOCK_SIZE);
+    const sycl::range<3> block_nums(1, 1, num_blocks);
+    stream->parallel_for(sycl::nd_range<3>(block_nums * block_dims, block_dims),
+        [=](sycl::nd_item<3> item_ct1) {
+            conv2d_dw_kernel<Layout>(x_d, w_d, y_d, p, item_ct1);
+        });
+}
+
+void ggml_sycl_op_conv2d_dw(ggml_backend_sycl_context & ctx, ggml_tensor * dst) {
+    scope_op_debug_print scope_dbg_print(__func__, dst, /*num_src=*/2);
+
+    const ggml_tensor * kernel = dst->src[0];
+    const ggml_tensor * input  = dst->src[1];
+
+    GGML_ASSERT(kernel->type == GGML_TYPE_F32 && input->type == GGML_TYPE_F32 && dst->type == GGML_TYPE_F32);
+
+    const float * w_d = (const float *) kernel->data;
+    const float * x_d = (const float *) input->data;
+    float *       y_d = (float *) dst->data;
+
+    const int32_t * p          = (const int32_t *) dst->op_params;
+    const int       stride_x   = p[0];
+    const int       stride_y   = p[1];
+    const int       padding_x  = p[2];
+    const int       padding_y  = p[3];
+    const int       dilation_x = p[4];
+    const int       dilation_y = p[5];
+
+    const int in_w     = input->ne[0];
+    const int in_h     = input->ne[1];
+    const int kernel_w = kernel->ne[0];
+    const int kernel_h = kernel->ne[1];
+    const int out_w    = dst->ne[0];
+    const int out_h    = dst->ne[1];
+    const int channels = dst->ne[2];
+    const int batches  = dst->ne[3];
+
+    const conv2d_dw_params params = { in_w, in_h, out_w, out_h, kernel_w, kernel_h,
+                                      stride_x, stride_y, padding_x, padding_y,
+                                      dilation_x, dilation_y, channels, batches };
+
+    const queue_ptr stream = ctx.stream();
+
+    if (ggml_is_contiguous(input)) {
+        conv2d_dw_sycl<dw_whcn_layout>(x_d, w_d, y_d, params, stream);
+    } else if (ggml_is_contiguous_channels(input)) {
+        conv2d_dw_sycl<dw_cwhn_layout>(x_d, w_d, y_d, params, stream);
+    } else {
+        GGML_ABORT("Unsupported memory layout for conv2d_dw");
+    }
+}