@novastera-oss/llamarn 0.2.5 → 0.2.6

package/cpp/llama.cpp/ggml/src/CMakeLists.txt

@@ -196,6 +196,7 @@ add_library(ggml-base
             ../include/ggml-opt.h
             ../include/gguf.h
             ggml.c
+            ggml.cpp
             ggml-alloc.c
             ggml-backend.cpp
             ggml-opt.cpp
@@ -226,6 +227,7 @@ function(ggml_add_backend_library backend)
         set_target_properties(${backend} PROPERTIES LIBRARY_OUTPUT_DIRECTORY ${CMAKE_RUNTIME_OUTPUT_DIRECTORY})
         target_compile_definitions(${backend} PRIVATE GGML_BACKEND_DL)
         add_dependencies(ggml ${backend})
+        install(TARGETS ${backend} LIBRARY DESTINATION ${CMAKE_INSTALL_BINDIR})
     else()
         add_library(${backend} ${ARGN})
         target_link_libraries(ggml PUBLIC ${backend})

package/cpp/llama.cpp/ggml/src/ggml-backend.cpp

@@ -1340,7 +1340,10 @@ static bool ggml_backend_sched_alloc_splits(ggml_backend_sched_t sched) {
     // allocate graph
     if (backend_ids_changed || !ggml_gallocr_alloc_graph(sched->galloc, &sched->graph)) {
         // the re-allocation may cause the split inputs to be moved to a different address
-        ggml_backend_sched_synchronize(sched);
+        // synchronize without ggml_backend_sched_synchronize to avoid changing cur_copy
+        for (int i = 0; i < sched->n_backends; i++) {
+            ggml_backend_synchronize(sched->backends[i]);
+        }
 #ifndef NDEBUG
         GGML_LOG_DEBUG("%s: failed to allocate graph, reserving (backend_ids_changed = %d)\n", __func__, backend_ids_changed);
 #endif
@@ -1564,7 +1567,6 @@ bool ggml_backend_sched_alloc_graph(ggml_backend_sched_t sched, struct ggml_cgra
 
     ggml_backend_sched_split_graph(sched, graph);
 
-
     if (!ggml_backend_sched_alloc_splits(sched)) {
         return false;
     }
@@ -1598,9 +1600,12 @@ void ggml_backend_sched_synchronize(ggml_backend_sched_t sched) {
     for (int i = 0; i < sched->n_backends; i++) {
         ggml_backend_synchronize(sched->backends[i]);
     }
-    // reset the current copy to 0 so that the graphs will be similar during generation
-    // necessary for CUDA graphs
-    sched->cur_copy = 0;
+    if (!sched->is_alloc) {
+        // if the graph is not already allocated, always use copy 0 after a synchronization
+        // this ensures that during generation the same copy is used every time,
+        // which avoids changes in the graph that could cause CUDA or other graphs to be disabled
+        sched->cur_copy = 0;
+    }
 }
 
 void ggml_backend_sched_set_eval_callback(ggml_backend_sched_t sched, ggml_backend_sched_eval_callback callback, void * user_data) {

package/cpp/llama.cpp/ggml/src/ggml-blas/CMakeLists.txt

@@ -81,7 +81,7 @@ if (BLAS_FOUND)
     target_link_libraries     (ggml-blas PRIVATE ${BLAS_LIBRARIES})
     target_include_directories(ggml-blas PRIVATE ${BLAS_INCLUDE_DIRS})
 else()
-    message(ERROR "BLAS not found, please refer to "
-                  "https://cmake.org/cmake/help/latest/module/FindBLAS.html#blas-lapack-vendors"
-                  " to set correct GGML_BLAS_VENDOR")
+    message(FATAL_ERROR "BLAS not found, please refer to "
+                        "https://cmake.org/cmake/help/latest/module/FindBLAS.html#blas-lapack-vendors"
+                        " to set correct GGML_BLAS_VENDOR")
 endif()

package/cpp/llama.cpp/ggml/src/ggml-cpu/ggml-cpu.c

@@ -2418,12 +2418,32 @@ static bool ggml_thread_apply_priority(int32_t prio) {
     // This is up to the applications.
     DWORD p = THREAD_PRIORITY_NORMAL;
     switch (prio) {
+        case GGML_SCHED_PRIO_LOW:      p = THREAD_PRIORITY_BELOW_NORMAL;  break;
         case GGML_SCHED_PRIO_NORMAL:   p = THREAD_PRIORITY_NORMAL;        break;
         case GGML_SCHED_PRIO_MEDIUM:   p = THREAD_PRIORITY_ABOVE_NORMAL;  break;
         case GGML_SCHED_PRIO_HIGH:     p = THREAD_PRIORITY_HIGHEST;       break;
         case GGML_SCHED_PRIO_REALTIME: p = THREAD_PRIORITY_TIME_CRITICAL; break;
     }
 
+    if (prio != GGML_SCHED_PRIO_LOW) {
+        // Tell Windows that this thread should not be throttled (needs its own CPU core).
+        // Newer Windows 11 versions aggresively park (offline) CPU cores and often place
+        // all our threads onto the first 4 cores which results in terrible performance with
+        // n_threads > 4
+        #if _WIN32_WINNT >= 0x0602
+        THREAD_POWER_THROTTLING_STATE t;
+        ZeroMemory(&t, sizeof(t));
+        t.Version     = THREAD_POWER_THROTTLING_CURRENT_VERSION;
+        t.ControlMask = THREAD_POWER_THROTTLING_EXECUTION_SPEED;
+        t.StateMask   = 0;
+
+        if (!SetThreadInformation(GetCurrentThread(), ThreadPowerThrottling, &t, sizeof(t))) {
+            GGML_LOG_DEBUG("failed to disable thread power throttling %d : (%d)\n", prio, (int) GetLastError());
+            return false;
+        }
+        #endif
+    }
+
     if (prio == GGML_SCHED_PRIO_NORMAL) {
         // Keep inherited policy/priority
         return true;
@@ -2451,6 +2471,8 @@ static bool ggml_thread_apply_priority(int32_t prio) {
     struct sched_param p;
     int32_t policy = SCHED_OTHER;
     switch (prio) {
+        // TODO: there seems to be no way to set lower prio on Apple platforms
+        case GGML_SCHED_PRIO_LOW:      policy = SCHED_OTHER; p.sched_priority = 0;  break;
         case GGML_SCHED_PRIO_NORMAL:   policy = SCHED_OTHER; p.sched_priority = 0;  break;
         case GGML_SCHED_PRIO_MEDIUM:   policy = SCHED_FIFO;  p.sched_priority = 40; break;
         case GGML_SCHED_PRIO_HIGH:     policy = SCHED_FIFO;  p.sched_priority = 80; break;
@@ -2507,6 +2529,7 @@ static bool ggml_thread_apply_priority(int32_t prio) {
     struct sched_param p;
     int32_t policy = SCHED_OTHER;
     switch (prio) {
+        case GGML_SCHED_PRIO_LOW:      policy = SCHED_BATCH; p.sched_priority = 0;  break;
         case GGML_SCHED_PRIO_NORMAL:   policy = SCHED_OTHER; p.sched_priority = 0;  break;
         case GGML_SCHED_PRIO_MEDIUM:   policy = SCHED_FIFO;  p.sched_priority = 40; break;
         case GGML_SCHED_PRIO_HIGH:     policy = SCHED_FIFO;  p.sched_priority = 80; break;

package/cpp/llama.cpp/ggml/src/ggml-cuda/common.cuh

@@ -635,6 +635,7 @@ struct ggml_cuda_device_info {
         int     nsm;                // number of streaming multiprocessors
         size_t  smpb;               // max. shared memory per block
         size_t  smpbo;              // max. shared memory per block (with opt-in)
+        bool    integrated;         // Device is integrated as opposed to discrete
         bool    vmm;                // virtual memory support
         size_t  vmm_granularity;    // granularity of virtual memory
         size_t  total_vram;

package/cpp/llama.cpp/ggml/src/ggml-cuda/fattn-mma-f16.cuh

@@ -1246,7 +1246,7 @@ static __global__ void flash_attn_ext_f16(
         NO_DEVICE_CODE;
         return;
     }
-#endif __CUDA_ARCH__ == GGML_CUDA_CC_TURING
+#endif // __CUDA_ARCH__ == GGML_CUDA_CC_TURING
 
     static_assert(!mla || DKQ >= DV, "MLA needs DKQ >= DV");
 

package/cpp/llama.cpp/ggml/src/ggml-cuda/ggml-cuda.cu

@@ -243,10 +243,10 @@ static ggml_cuda_device_info ggml_cuda_init() {
 
         info.default_tensor_split[id] = total_vram;
         total_vram += prop.totalGlobalMem;
-
-        info.devices[id].nsm       = prop.multiProcessorCount;
-        info.devices[id].smpb      = prop.sharedMemPerBlock;
-        info.devices[id].warp_size = prop.warpSize;
+        info.devices[id].integrated = prop.integrated;
+        info.devices[id].nsm        = prop.multiProcessorCount;
+        info.devices[id].smpb       = prop.sharedMemPerBlock;
+        info.devices[id].warp_size  = prop.warpSize;
 #if defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__)
         info.devices[id].smpbo = prop.sharedMemPerBlock;
 
@@ -1065,6 +1065,10 @@ static const char * ggml_backend_cuda_host_buffer_type_name(ggml_backend_buffer_
     GGML_UNUSED(buft);
 }
 
+static bool ggml_backend_buft_is_cuda_host(ggml_backend_buffer_type_t buft) {
+    return buft->iface.get_name == ggml_backend_cuda_host_buffer_type_name;
+}
+
 static void ggml_backend_cuda_host_buffer_free_buffer(ggml_backend_buffer_t buffer) {
     CUDA_CHECK(cudaFreeHost(buffer->context));
 }
@@ -2641,6 +2645,8 @@ static void update_cuda_graph_executable(ggml_backend_cuda_context * cuda_ctx) {
 
 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
+    const bool integrated = ggml_cuda_info().devices[cuda_ctx->device].integrated;
 
     while (!graph_evaluated_or_captured) {
         // Only perform the graph execution if CUDA graphs are not enabled, or we are capturing the graph.
@@ -2659,7 +2665,7 @@ static void evaluate_and_capture_cuda_graph(ggml_backend_cuda_context * cuda_ctx
                     if (node->src[j] != nullptr) {
                         assert(node->src[j]->buffer);
                         assert(node->src[j]->buffer->buft == ggml_backend_cuda_buffer_type(cuda_ctx->device) ||
-                               ggml_backend_buft_is_cuda_split(node->src[j]->buffer->buft));
+                               ggml_backend_buft_is_cuda_split(node->src[j]->buffer->buft) || (integrated && ggml_backend_buft_is_cuda_host(node->src[j]->buffer->buft)));
                     }
                 }
 #endif
@@ -2994,9 +3000,12 @@ static bool ggml_backend_cuda_device_supports_op(ggml_backend_dev_t dev, const g
             {
                 struct ggml_tensor * a = op->src[0];
                 struct ggml_tensor * b = op->src[1];
-                // for small weight matrices the active device can end up without any rows, don't use row split in those cases
-                // this avoids some edge cases (and the performance would not be good anyways)
                 if (a->buffer && ggml_backend_buft_is_cuda_split(a->buffer->buft)) {
+                    if (a->ne[2] > 1 || a->ne[3] > 1) {
+                        return false;
+                    }
+                    // for small weight matrices the active device can end up without any rows, don't use row split in those cases
+                    // this avoids some edge cases (and the performance would not be good anyways)
                     ggml_backend_cuda_split_buffer_type_context * buft_ctx = (ggml_backend_cuda_split_buffer_type_context *) a->buffer->buft->context;
                     int64_t row_low;
                     int64_t row_high;
@@ -3263,7 +3272,9 @@ static bool ggml_backend_cuda_device_supports_op(ggml_backend_dev_t dev, const g
 }
 
 static bool ggml_backend_cuda_device_supports_buft(ggml_backend_dev_t dev, ggml_backend_buffer_type_t buft) {
-    return (ggml_backend_buft_is_cuda(buft) || ggml_backend_buft_is_cuda_split(buft)) && buft->device == dev;
+    ggml_backend_cuda_device_context * dev_ctx = (ggml_backend_cuda_device_context *) dev->context;
+    const bool integrated = ggml_cuda_info().devices[dev_ctx->device].integrated;
+    return (((ggml_backend_buft_is_cuda(buft) || ggml_backend_buft_is_cuda_split(buft)) && buft->device == dev) || (integrated && ggml_backend_buft_is_cuda_host(buft)));
 }
 
 static int64_t get_op_batch_size(const ggml_tensor * op) {

package/cpp/llama.cpp/ggml/src/ggml-impl.h

@@ -32,6 +32,8 @@
 extern "C" {
 #endif
 
+void ggml_print_backtrace(void);
+
 #ifndef MIN
 #    define MIN(a, b) ((a) < (b) ? (a) : (b))
 #endif

package/cpp/llama.cpp/ggml/src/ggml-sycl/CMakeLists.txt

@@ -13,7 +13,7 @@ elseif(SUPPORTS_SYCL)
         If you expected the oneAPI Release compiler, please install oneAPI & source it, like:
         source /opt/intel/oneapi/setvars.sh")
 else()
-    message(FATAL_ERROR, "C++ compiler lacks SYCL support.")
+    message(FATAL_ERROR "C++ compiler lacks SYCL support.")
 endif()
 message(STATUS "SYCL found")
 #todo: AOT
@@ -170,7 +170,7 @@ else()
         target_compile_definitions(ggml-sycl PRIVATE GGML_SYCL_NVIDIA)
     elseif (GGML_SYCL_TARGET STREQUAL "AMD")
         if (NOT GGML_SYCL_DEVICE_ARCH)
-            message(ERROR "Can't enable SYCL hip backend, GGML_SYCL_DEVICE_ARCH has not been set.")
+            message(FATAL_ERROR "Can't enable SYCL hip backend, GGML_SYCL_DEVICE_ARCH has not been set.")
         endif()
         target_link_libraries(ggml-sycl PRIVATE ONEMATH::onemath_blas_rocblas)
         target_compile_options(ggml-sycl PRIVATE "-fsycl-targets=amdgcn-amd-amdhsa")

package/cpp/llama.cpp/ggml/src/ggml-sycl/ggml-sycl.cpp

@@ -4257,14 +4257,6 @@ static bool ggml_backend_sycl_device_supports_op(ggml_backend_dev_t dev, const g
         case GGML_OP_SOFT_MAX:
             return true;
         case GGML_OP_ROPE:
-            {
-                const int mode = ((const int32_t *) op->op_params)[2];
-                // mode is not used as a bitmask in practice, the various rope type modes are independent implementations
-                if (mode == GGML_ROPE_TYPE_MROPE) {
-                    return false;
-                }
-                return true;
-            }
         case GGML_OP_IM2COL:
             return true;
         case GGML_OP_UPSCALE:

package/cpp/llama.cpp/ggml/src/ggml-sycl/rope.cpp

@@ -49,10 +49,7 @@ static void rope_norm(const T * x, T * dst, const int ne0, const int ne1, const
 
     if (i0 >= n_dims) {
         const int i = row * ne0 + i0;
-
-        dst[i + 0] = x[i + 0];
-        dst[i + 1] = x[i + 1];
-
+        *reinterpret_cast<sycl::vec<T, 2> *>(dst + i) = *reinterpret_cast<const sycl::vec<T, 2> *>(x + i);
         return;
     }
 
@@ -93,10 +90,7 @@ static void rope_neox(const T * x, T * dst, const int ne0, const int ne1, const
 
     if (i0 >= n_dims) {
         const int i = row * ne0 + i0;
-
-        dst[i + 0] = x[i + 0];
-        dst[i + 1] = x[i + 1];
-
+        *reinterpret_cast<sycl::vec<T, 2> *>(dst + i) = *reinterpret_cast<const sycl::vec<T, 2> *>(x + i);
         return;
     }
 
@@ -122,6 +116,63 @@ static void rope_neox(const T * x, T * dst, const int ne0, const int ne1, const
     dst[i + n_dims / 2] = x0 * sin_theta + x1 * cos_theta;
 }
 
+template <typename T, bool has_ff>
+static void rope_multi(const T * x, T * dst, const int ne0, const int ne1, const int ne2, const size_t s1,
+                        const size_t s2, const int n_dims, const int32_t * pos, const float freq_scale,
+                        const float ext_factor, const float attn_factor, const rope_corr_dims corr_dims,
+                        const float theta_scale, const float * freq_factors, const mrope_sections sections,
+                        const sycl::nd_item<3> & item_ct1) {
+    // get index pos
+    const int i0 = 2 * (item_ct1.get_group(1) * item_ct1.get_local_range(1) + item_ct1.get_local_id(1));
+    if (i0 >= ne0) {
+        return;
+    }
+    const int    row_dst   = (item_ct1.get_group(2) * item_ct1.get_local_range(2)) + item_ct1.get_local_id(2);
+
+    if (i0 >= n_dims) {
+        const int i = row_dst*ne0 + i0;
+        *reinterpret_cast<sycl::vec<T, 2> *>(dst + i) = *reinterpret_cast<const sycl::vec<T, 2> *>(x + i);
+        return;
+    }
+
+    const int    row_x     = row_dst % ne1;
+    const int    channel_x = row_dst / ne1;
+    const int    idst      = (row_dst * ne0) + (i0 / 2);
+    const size_t ix        = ((size_t) channel_x * s2) + ((size_t) row_x * s1) + (i0 / 2);
+
+    const int sect_dims = sections.v[0] + sections.v[1] + sections.v[2] + sections.v[3];
+    const int sec_w = sections.v[1] + sections.v[0];
+    const int sector = (i0 / 2) % sect_dims;
+
+
+    float theta_base = 0.0;
+    if (sector < sections.v[0]) {
+        theta_base = pos[channel_x]*sycl::pow(theta_scale, i0/2.0f);
+    }
+    else if (sector >= sections.v[0] && sector < sec_w) {
+        theta_base = pos[channel_x + ne2 * 1]*sycl::pow(theta_scale, i0/2.0f);
+    }
+    else if (sector >= sec_w && sector < sec_w + sections.v[2]) {
+        theta_base = pos[channel_x + ne2 * 2]*sycl::pow(theta_scale, i0/2.0f);
+    }
+    else if (sector >= sec_w + sections.v[2]) {
+        theta_base = pos[channel_x + ne2 * 3]*sycl::pow(theta_scale, i0/2.0f);
+    }
+
+    const float freq_factor = has_ff ? freq_factors[i0 / 2] : 1.0f;
+    float       cos_theta;
+    float       sin_theta;
+    rope_yarn(theta_base / freq_factor, freq_scale, corr_dims, i0, ext_factor, attn_factor, &cos_theta, &sin_theta);
+    const float x0 = x[ix + 0];
+    const float x1 = x[ix + n_dims/2];
+
+    // store results in dst
+    dst[idst + 0]      = x0 * cos_theta - x1 * sin_theta;
+    dst[idst + n_dims/2] = x0 * sin_theta + x1 * cos_theta;
+}
+
+
+
 template <typename T, bool has_ff>
 static void rope_vision(const T * x, T * dst, const int ne0, const int ne1, const int ne2, const size_t s1,
                         const size_t s2, const int n_dims, const int32_t * pos, const float freq_scale,
@@ -171,7 +222,7 @@ static void rope_norm_sycl(const T * x, T * dst, const int ne0, const int ne1, c
                            const float * freq_factors, queue_ptr stream) {
     GGML_ASSERT(ne0 % 2 == 0);
     const sycl::range<3> block_dims(1, SYCL_ROPE_BLOCK_SIZE, 1);
-    const int            num_blocks_x = (ne0 + 2 * SYCL_ROPE_BLOCK_SIZE - 1) / (2 * SYCL_ROPE_BLOCK_SIZE);
+    const int            num_blocks_x = ceil_div(ne0, (2 * SYCL_ROPE_BLOCK_SIZE));
     const sycl::range<3> block_nums(1, num_blocks_x, nr);
 
     const float theta_scale = powf(freq_base, -2.0f / n_dims);
@@ -208,7 +259,7 @@ static void rope_neox_sycl(const T * x, T * dst, const int ne0, const int ne1, c
                            const rope_corr_dims corr_dims, const float * freq_factors, queue_ptr stream) {
     GGML_ASSERT(ne0 % 2 == 0);
     const sycl::range<3> block_dims(1, SYCL_ROPE_BLOCK_SIZE, 1);
-    const int            num_blocks_x = (ne0 + 2 * SYCL_ROPE_BLOCK_SIZE - 1) / (2 * SYCL_ROPE_BLOCK_SIZE);
+    const int            num_blocks_x = ceil_div(ne0, (2 * SYCL_ROPE_BLOCK_SIZE));
     const sycl::range<3> block_nums(1, num_blocks_x, nr);
 
     const float theta_scale = powf(freq_base, -2.0f / n_dims);
@@ -228,6 +279,40 @@ static void rope_neox_sycl(const T * x, T * dst, const int ne0, const int ne1, c
     }
 }
 
+template <typename T>
+static void rope_multi_sycl(const T * x, T * dst, const int ne0, const int ne1, const int ne2, const size_t s1,
+                             const size_t s2, const int n_dims, const int nr, const int32_t * pos,
+                             const float freq_scale, const float freq_base, const float ext_factor,
+                             const float attn_factor, const rope_corr_dims corr_dims, const float * freq_factors,
+                             const mrope_sections sections, queue_ptr stream) {
+    GGML_ASSERT(ne0 % 2 == 0);
+    const sycl::range<3>    block_dims(1, SYCL_ROPE_BLOCK_SIZE, 1);
+    const int               n_blocks_y = ceil_div(ne0, (2 * SYCL_ROPE_BLOCK_SIZE));
+    const sycl::range<3>    grid_dims(1, n_blocks_y, nr);
+    const sycl::nd_range<3> nd_range(grid_dims * block_dims, block_dims);
+
+    const float theta_scale = std::pow(freq_base, -2.0f / n_dims);
+    // Add FP16 capability check if T could be sycl::half
+    if constexpr (std::is_same_v<T, sycl::half>) {
+        dpct::has_capability_or_fail(stream->get_device(), { sycl::aspect::fp16 });
+    }
+    // launch kernel
+    if (freq_factors == nullptr) {
+        stream->parallel_for(nd_range, [=](sycl::nd_item<3> item_ct1) {
+            rope_multi<T, false>(x, dst, ne0, ne1, ne2, s1, s2, n_dims, pos, freq_scale, ext_factor, attn_factor,
+                                  corr_dims, theta_scale, freq_factors, sections, item_ct1);
+        });
+    } else {
+        stream->parallel_for(nd_range, [=](sycl::nd_item<3> item_ct1) {
+            rope_multi<T, true>(x, dst, ne0, ne1, ne2, s1, s2, n_dims, pos, freq_scale, ext_factor, attn_factor,
+                                 corr_dims, theta_scale, freq_factors, sections, item_ct1);
+        });
+    }
+}
+
+
+
+
 // rope vision
 template <typename T>
 static void rope_vision_sycl(const T * x, T * dst, const int ne0, const int ne1, const int ne2, const size_t s1,
@@ -237,7 +322,7 @@ static void rope_vision_sycl(const T * x, T * dst, const int ne0, const int ne1,
                              const mrope_sections sections, queue_ptr stream) {
     GGML_ASSERT(ne0 % 2 == 0);
     const sycl::range<3>    block_dims(1, SYCL_ROPE_BLOCK_SIZE, 1);
-    const int               n_blocks_y = (ne0 + 2 * SYCL_ROPE_BLOCK_SIZE - 1) / (2 * SYCL_ROPE_BLOCK_SIZE);
+    const int               n_blocks_y = ceil_div(ne0, (2 * SYCL_ROPE_BLOCK_SIZE));
     const sycl::range<3>    grid_dims(1, n_blocks_y, nr);
     const sycl::nd_range<3> nd_range(grid_dims * block_dims, block_dims);
 
@@ -298,8 +383,17 @@ inline void ggml_sycl_op_rope(ggml_backend_sycl_context & ctx, ggml_tensor *dst)
     memcpy(&sections.v,  (int32_t *) dst->op_params + 11, sizeof(int)*4);
 
     const bool is_neox = mode & GGML_ROPE_TYPE_NEOX;
+    const bool is_mrope = mode & GGML_ROPE_TYPE_MROPE;
     const bool is_vision = mode == GGML_ROPE_TYPE_VISION;
 
+    if (is_mrope) {
+        GGML_ASSERT(sections.v[0] > 0 || sections.v[1] > 0 || sections.v[2] > 0);
+    }
+
+    if (is_vision) {
+        GGML_ASSERT(n_dims == ne00/2);
+    }
+
     const int32_t * pos = (const int32_t *) dst->src[1]->data;
 
     const float * freq_factors = nullptr;
@@ -326,6 +420,19 @@ inline void ggml_sycl_op_rope(ggml_backend_sycl_context & ctx, ggml_tensor *dst)
         } else {
             GGML_ABORT("fatal error");
         }
+    } else if (is_mrope && !is_vision) {
+        GGML_SYCL_DEBUG("%s: mrope path\n", __func__);
+        if (dst->src[0]->type == GGML_TYPE_F16) {
+            rope_multi_sycl((const sycl::half *)dst->src[0]->data, (sycl::half *)dst->data, ne00, ne01, ne02, s01,
+                s02, n_dims, nr, pos, freq_scale, freq_base, ext_factor, attn_factor, corr_dims,
+                freq_factors, sections, main_stream);
+        } else if (dst->src[0]->type == GGML_TYPE_F32) {
+            rope_multi_sycl((const float *) dst->src[0]->data, (float *) dst->data, ne00, ne01, ne02, s01, s02, n_dims,
+                             nr, pos, freq_scale, freq_base, ext_factor, attn_factor, corr_dims, freq_factors, sections,
+                             main_stream);
+        } else {
+            GGML_ABORT("Fatal error: Tensor type unsupported!");
+        }
     } else if (is_vision) {
         GGML_SYCL_DEBUG("%s: vision path\n", __func__);
         if (dst->src[0]->type == GGML_TYPE_F16) {

package/cpp/llama.cpp/ggml/src/ggml-vulkan/ggml-vulkan.cpp

@@ -1652,7 +1652,7 @@ static std::array<uint32_t, 2> fa_rows_cols(FaCodePath path, uint32_t D, uint32_
         return {64, 32};
     }
     return {64, 64};
-};
+}
 
 static bool ggml_vk_matmul_shmem_support(const vk_device& device, const std::vector<uint32_t>& warptile, bool mul_mat_id, ggml_type src0_type) {
 

package/cpp/llama.cpp/ggml/src/ggml.c

@@ -133,7 +133,7 @@ static void ggml_print_backtrace_symbols(void) {
 }
 #endif
 
-static void ggml_print_backtrace(void) {
+void ggml_print_backtrace(void) {
     const char * GGML_NO_BACKTRACE = getenv("GGML_NO_BACKTRACE");
     if (GGML_NO_BACKTRACE) {
         return;
@@ -160,6 +160,10 @@ static void ggml_print_backtrace(void) {
     const int parent_pid = getpid();
     const int child_pid = fork();
     if (child_pid < 0) { // error
+#if defined(__linux__)
+        close(lock[1]);
+        close(lock[0]);
+#endif
         return;
     } else if (child_pid == 0) { // child
         char attach[32];
@@ -167,6 +171,7 @@ static void ggml_print_backtrace(void) {
 #if defined(__linux__)
         close(lock[1]);
         (void) !read(lock[0], lock, 1);
+        close(lock[0]);
 #endif
         // try gdb
         execlp("gdb", "gdb", "--batch",
@@ -195,7 +200,7 @@ static void ggml_print_backtrace(void) {
     }
 }
 #else
-static void ggml_print_backtrace(void) {
+void ggml_print_backtrace(void) {
     // platform not supported
 }
 #endif
@@ -216,6 +221,8 @@ void ggml_abort(const char * file, int line, const char * fmt, ...) {
     abort();
 }
 
+// ggml_print_backtrace is registered with std::set_terminate by ggml.cpp
+
 //
 // logging
 //

package/cpp/llama.cpp/ggml/src/ggml.cpp

@@ -0,0 +1,26 @@
+#include "ggml-impl.h"
+
+#include <cstdlib>
+#include <exception>
+
+static std::terminate_handler previous_terminate_handler;
+
+GGML_NORETURN static void ggml_uncaught_exception() {
+    ggml_print_backtrace();
+    if (previous_terminate_handler) {
+        previous_terminate_handler();
+    }
+    abort(); // unreachable unless previous_terminate_handler was nullptr
+}
+
+static bool ggml_uncaught_exception_init = []{
+    const char * GGML_NO_BACKTRACE = getenv("GGML_NO_BACKTRACE");
+    if (GGML_NO_BACKTRACE) {
+        return false;
+    }
+    const auto prev{std::get_terminate()};
+    GGML_ASSERT(prev != ggml_uncaught_exception);
+    previous_terminate_handler = prev;
+    std::set_terminate(ggml_uncaught_exception);
+    return true;
+}();

package/cpp/llama.cpp/ggml/src/gguf.cpp

@@ -347,11 +347,28 @@ struct gguf_context * gguf_init_from_file_impl(FILE * file, struct gguf_init_par
     int64_t n_tensors = 0;
 
     if (ok && gr.read(ctx->version)) {
-        if (ctx->version == 1) {
+        if (ok && ctx->version == 0) {
+            GGML_LOG_ERROR("%s: bad GGUF version: %" PRIu32 "\n", __func__, ctx->version);
+            ok = false;
+        }
+
+        /*
+         * bit layout is different when reading non-native endian models.
+         * assuming that the GGUF version is 3, the non-native endian model
+         * would read it as 0x30000000. we can use the AND operation against
+         * the last 4 hexadecimal digits to check if the model is the same
+         * endianness as the host system.
+        */
+        if (ok && (ctx->version & 0x0000FFFF) == 0x00000000) {
+            GGML_LOG_ERROR("%s: failed to load model: this GGUF file version %" PRIu32 " is extremely large, is there a mismatch between the host and model endianness?\n", __func__, ctx->version);
+            ok = false;
+        }
+
+        if (ok && ctx->version == 1) {
             GGML_LOG_ERROR("%s: GGUFv1 is no longer supported, please use a more up-to-date version\n", __func__);
             ok = false;
         }
-        if (ctx->version > GGUF_VERSION) {
+        if (ok && ctx->version > GGUF_VERSION) {
             GGML_LOG_ERROR("%s: this GGUF file is version %" PRIu32 " but this software only supports up to version %d\n",
                 __func__, ctx->version, GGUF_VERSION);
             ok = false;

package/cpp/llama.cpp/include/llama.h

@@ -259,9 +259,9 @@ extern "C" {
         llama_token  *  token;
         float        *  embd;
         llama_pos    *  pos;
-        int32_t      *  n_seq_id;
-        llama_seq_id ** seq_id;
-        int8_t       *  logits; // TODO: rename this to "output"
+        int32_t      *  n_seq_id; // TODO: remove, should belong to only 1 sequence
+        llama_seq_id ** seq_id;   // TODO: become llama_seq_id * seq_id;
+        int8_t       *  logits;   // TODO: rename this to "output"
     } llama_batch;
 
     enum llama_model_kv_override_type {
@@ -366,6 +366,8 @@ extern "C" {
         bool no_perf;     // measure performance timings
         bool op_offload;  // offload host tensor operations to device
         bool swa_full;    // use full-size SWA cache (https://github.com/ggml-org/llama.cpp/pull/13194#issuecomment-2868343055)
+                          // NOTE: setting to false when n_seq_max > 1 can cause bad performance in some cases
+                          //       ref: https://github.com/ggml-org/llama.cpp/pull/13845#issuecomment-2924800573
     };
 
     // model quantization parameters
@@ -502,6 +504,7 @@ extern "C" {
     LLAMA_API int32_t llama_model_n_layer    (const struct llama_model * model);
     LLAMA_API int32_t llama_model_n_head     (const struct llama_model * model);
     LLAMA_API int32_t llama_model_n_head_kv  (const struct llama_model * model);
+    LLAMA_API int32_t llama_model_n_swa      (const struct llama_model * model);
 
     // Get the model's RoPE frequency scaling factor
     LLAMA_API float llama_model_rope_freq_scale_train(const struct llama_model * model);
@@ -652,7 +655,6 @@ extern "C" {
     // Adds relative position "delta" to all tokens that belong to the specified sequence and have positions in [p0, p1)
     // If the KV cache is RoPEd, the KV data is updated accordingly:
     //   - lazily on next llama_decode()
-    //   - explicitly with llama_kv_self_update()
     // p0 < 0 : [0,  p1]
     // p1 < 0 : [p0, inf)
     LLAMA_API void llama_kv_self_seq_add(
@@ -665,7 +667,6 @@ extern "C" {
     // Integer division of the positions by factor of `d > 1`
     // If the KV cache is RoPEd, the KV data is updated accordingly:
     //   - lazily on next llama_decode()
-    //   - explicitly with llama_kv_self_update()
     // p0 < 0 : [0,  p1]
     // p1 < 0 : [p0, inf)
     LLAMA_API void llama_kv_self_seq_div(
@@ -677,12 +678,14 @@ extern "C" {
 
     // Returns the smallest position present in the KV cache for the specified sequence
     // This is typically non-zero only for SWA caches
+    // Note that all positions in the range [pos_min, pos_max] are guaranteed to be present in the KV cache
     // Return -1 if the sequence is empty
     LLAMA_API llama_pos llama_kv_self_seq_pos_min(
             struct llama_context * ctx,
                     llama_seq_id   seq_id);
 
     // Returns the largest position present in the KV cache for the specified sequence
+    // Note that all positions in the range [pos_min, pos_max] are guaranteed to be present in the KV cache
     // Return -1 if the sequence is empty
     LLAMA_API llama_pos llama_kv_self_seq_pos_max(
             struct llama_context * ctx,
@@ -691,14 +694,15 @@ extern "C" {
     // Defragment the KV cache
     // This will be applied:
     //   - lazily on next llama_decode()
-    //   - explicitly with llama_kv_self_update()
-    LLAMA_API void llama_kv_self_defrag(struct llama_context * ctx);
+    LLAMA_API DEPRECATED(void llama_kv_self_defrag(struct llama_context * ctx),
+            "simply remove this call, the context will automatically decide when to do a defragmentation based on 'defrag_thold'");
 
     // Check if the context supports KV cache shifting
     LLAMA_API bool llama_kv_self_can_shift(const struct llama_context * ctx);
 
     // Apply the KV cache updates (such as K-shifts, defragmentation, etc.)
-    LLAMA_API void llama_kv_self_update(struct llama_context * ctx);
+    LLAMA_API DEPRECATED(void llama_kv_self_update(struct llama_context * ctx),
+            "simply remove this call, updates are applied lazily on the next llama_decode()");
 
     //
     // State / sessions

package/cpp/llama.cpp/src/CMakeLists.txt

@@ -21,6 +21,9 @@ add_library(llama
             llama-impl.cpp
             llama-io.cpp
             llama-kv-cache.cpp
+            llama-kv-cache-unified.cpp
+            llama-kv-cache-unified-iswa.cpp
+            llama-kv-cache-recurrent.cpp
             llama-memory.cpp
             llama-mmap.cpp
             llama-model-loader.cpp