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

#include "ggml-cuda.h"
#include "ggml-impl.h"
#include "ggml-backend-impl.h"

#include "ggml-cuda/common.cuh"
#include "ggml-cuda/acc.cuh"
#include "ggml-cuda/arange.cuh"
#include "ggml-cuda/argmax.cuh"
#include "ggml-cuda/argsort.cuh"
#include "ggml-cuda/binbcast.cuh"
#include "ggml-cuda/clamp.cuh"
#include "ggml-cuda/concat.cuh"
#include "ggml-cuda/conv-transpose-1d.cuh"
#include "ggml-cuda/convert.cuh"
#include "ggml-cuda/count-equal.cuh"
#include "ggml-cuda/cpy.cuh"
#include "ggml-cuda/cross-entropy-loss.cuh"
#include "ggml-cuda/diagmask.cuh"
#include "ggml-cuda/fattn.cuh"
#include "ggml-cuda/getrows.cuh"
#include "ggml-cuda/im2col.cuh"
#include "ggml-cuda/mmq.cuh"
#include "ggml-cuda/mmv.cuh"
#include "ggml-cuda/mmvq.cuh"
#include "ggml-cuda/norm.cuh"
#include "ggml-cuda/opt-step-adamw.cuh"
#include "ggml-cuda/out-prod.cuh"
#include "ggml-cuda/pad.cuh"
#include "ggml-cuda/pool2d.cuh"
#include "ggml-cuda/quantize.cuh"
#include "ggml-cuda/rope.cuh"
#include "ggml-cuda/scale.cuh"
#include "ggml-cuda/softmax.cuh"
#include "ggml-cuda/sum.cuh"
#include "ggml-cuda/sumrows.cuh"
#include "ggml-cuda/tsembd.cuh"
#include "ggml-cuda/unary.cuh"
#include "ggml-cuda/upscale.cuh"
#include "ggml-cuda/wkv6.cuh"
#include "ggml-cuda/gla.cuh"
#include "ggml.h"

#include <algorithm>
#include <array>
#include <atomic>
#include <charconv>
#include <cinttypes>
#include <cstddef>
#include <cstdint>
#include <float.h>
#include <limits>
#include <map>
#include <memory>
#include <mutex>
#include <stdint.h>
#include <stdio.h>
#include <stdarg.h>
#include <stdlib.h>
#include <string>
#include <vector>

static_assert(sizeof(half) == sizeof(ggml_fp16_t), "wrong fp16 size");

[[noreturn]]
void ggml_cuda_error(const char * stmt, const char * func, const char * file, int line, const char * msg) {
    int id = -1; // in case cudaGetDevice fails
    (void)cudaGetDevice(&id);

    GGML_LOG_ERROR(GGML_CUDA_NAME " error: %s\n", msg);
    GGML_LOG_ERROR("  current device: %d, in function %s at %s:%d\n", id, func, file, line);
    GGML_LOG_ERROR("  %s\n", stmt);
    // abort with GGML_ABORT to get a stack trace
    GGML_ABORT(GGML_CUDA_NAME " error");
}

// this is faster on Windows
// probably because the Windows CUDA libraries forget to make this check before invoking the drivers
void ggml_cuda_set_device(int device) {
    int current_device;
    CUDA_CHECK(cudaGetDevice(&current_device));

    if (device == current_device) {
        return;
    }

    CUDA_CHECK(cudaSetDevice(device));
}

int ggml_cuda_get_device() {
    int id;
    CUDA_CHECK(cudaGetDevice(&id));
    return id;
}

static cudaError_t ggml_cuda_device_malloc(void ** ptr, size_t size, int device) {
    ggml_cuda_set_device(device);
#if defined(GGML_USE_HIP) && defined(GGML_HIP_UMA)
    auto res = hipMallocManaged(ptr, size);
    if (res == hipSuccess) {
        // if error we "need" to know why...
        CUDA_CHECK(hipMemAdvise(*ptr, size, hipMemAdviseSetCoarseGrain, device));
    }
    return res;
#else

#if !defined(GGML_USE_HIP)
    cudaError_t err;
    if (getenv("GGML_CUDA_ENABLE_UNIFIED_MEMORY") != nullptr)
    {
        err = cudaMallocManaged(ptr, size);
    }
    else
    {
        err = cudaMalloc(ptr, size);
    }
    return err;
#else
    return cudaMalloc(ptr, size);
#endif // !defined(GGML_USE_HIP)

#endif
}

#if defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__)
static int ggml_cuda_parse_id(char devName[]) {
    // A list of possible Target IDs can be found under the rocclr/clr repo in device.cpp
    // these values are not stable so this is susceptible to breakage
    // https://github.com/ROCm/clr/blob/amd-staging/rocclr/device/device.cpp
    int archMajor = 0x0;
    int archMinor = 0x0;
    int archNum = GGML_CUDA_CC_OFFSET_AMD;
    int archLen = strlen(devName);
    char archName[archLen + 1];

    // strip leading 'gfx' while copying into our buffer
    if (archLen > 3) {
        strcpy(archName, &devName[3]);
        archLen -= 3;
    }

    // trim trailing :xnack- or :sramecc- statuses
    archLen = strcspn(archName, ":");
    archName[archLen] = '\0';

    // tease out the version information
    if (archLen > 8) {
        // versions labeled generic use '-' as delimiter
        // strip the trailing "-generic" then iterate through what remains
        if ((strstr(archName, "-generic"))) {
            archName[archLen - 8] = '\0';
            char * pch;
            if ((pch = strtok(archName, "-"))) {
                archMajor = (int)strtoul(pch, 0, 16);
                if ((pch = strtok(NULL, "-"))) {
                    archMinor = 0x10 * (int)strtoul(pch, 0, 16);
                }
            }
        }
    } else if (archLen >= 3) {
        // last two digits should be the minor * 0x10 + stepping
        archMinor = (int)strtoul(&archName[archLen - 2], 0, 16);
        archName[archLen - 2] = '\0';

        // only the major version remains
        archMajor = (int)strtoul(archName, 0, 16);
    }
    archNum += archMajor * 0x100;
    archNum += archMinor;
    return archNum;
}
#endif // defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__)

static ggml_cuda_device_info ggml_cuda_init() {
#ifdef __HIP_PLATFORM_AMD__
    // Workaround for a rocBLAS bug when using multiple graphics cards:
    // https://github.com/ROCmSoftwarePlatform/rocBLAS/issues/1346
    {
        int major_version = 0;
        size_t version_length = 0;
        if (rocblas_get_version_string_size(&version_length) == rocblas_status_success) {
            std::vector<char> version(version_length+1, '\0');
            if (rocblas_get_version_string(version.data(), version.size()) == rocblas_status_success) {
                version.resize(::strlen(version.data()));
                int parsed_value = 0;
                if (std::from_chars(version.data(), version.data() + version.size(), parsed_value).ec == std::errc()) {
                    major_version = parsed_value;
                }
            }
        }
        if (major_version < 4) {
            GGML_LOG_DEBUG(GGML_CUDA_NAME " calling rocblas_initialize as a workaround for a rocBLAS bug\n");
            rocblas_initialize();
            CUDA_CHECK(cudaDeviceSynchronize());
        }
    }
#endif

    ggml_cuda_device_info info = {};

    cudaError_t err = cudaGetDeviceCount(&info.device_count);
    if (err != cudaSuccess) {
        GGML_LOG_ERROR("%s: failed to initialize " GGML_CUDA_NAME ": %s\n", __func__, cudaGetErrorString(err));
        return info;
    }

    GGML_ASSERT(info.device_count <= GGML_CUDA_MAX_DEVICES);

    int64_t total_vram = 0;
#ifdef GGML_CUDA_FORCE_MMQ
    GGML_LOG_INFO("%s: GGML_CUDA_FORCE_MMQ:    yes\n", __func__);
#else
    GGML_LOG_INFO("%s: GGML_CUDA_FORCE_MMQ:    no\n", __func__);
#endif // GGML_CUDA_FORCE_MMQ
#ifdef GGML_CUDA_FORCE_CUBLAS
    GGML_LOG_INFO("%s: GGML_CUDA_FORCE_CUBLAS: yes\n", __func__);
#else
    GGML_LOG_INFO("%s: GGML_CUDA_FORCE_CUBLAS: no\n", __func__);
#endif // GGML_CUDA_FORCE_CUBLAS
    GGML_LOG_INFO("%s: found %d " GGML_CUDA_NAME " devices:\n", __func__, info.device_count);
    for (int id = 0; id < info.device_count; ++id) {
        int device_vmm = 0;

#if defined(GGML_USE_VMM)
        CUdevice device;
        CU_CHECK(cuDeviceGet(&device, id));
        CU_CHECK(cuDeviceGetAttribute(&device_vmm, CU_DEVICE_ATTRIBUTE_VIRTUAL_MEMORY_MANAGEMENT_SUPPORTED, device));

        if (device_vmm) {
            CUmemAllocationProp alloc_prop = {};
            alloc_prop.type = CU_MEM_ALLOCATION_TYPE_PINNED;
            alloc_prop.location.type = CU_MEM_LOCATION_TYPE_DEVICE;
            alloc_prop.location.id = id;
            CU_CHECK(cuMemGetAllocationGranularity(&info.devices[id].vmm_granularity, &alloc_prop, CU_MEM_ALLOC_GRANULARITY_RECOMMENDED));
        }
#endif // defined(GGML_USE_VMM)
        info.devices[id].vmm = !!device_vmm;

        cudaDeviceProp prop;
        CUDA_CHECK(cudaGetDeviceProperties(&prop, id));

        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;
#if defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__)
        info.devices[id].smpbo = prop.sharedMemPerBlock;

        info.devices[id].cc = ggml_cuda_parse_id(prop.gcnArchName);
        if ((info.devices[id].cc & 0xff00) == 0x0) {
            GGML_LOG_WARN("invalid architecture ID received for device %d %s: %s  cc %d.%d\n",
                            id, prop.name, prop.gcnArchName, prop.major, prop.minor);

            // Fallback to prop.major and prop.minor
            if (prop.major > 0) {
                info.devices[id].cc = GGML_CUDA_CC_OFFSET_AMD + prop.major * 0x100;
                info.devices[id].cc += prop.minor * 0x10;
            }
        }
        GGML_LOG_INFO("  Device %d: %s, %s (0x%x), VMM: %s, Wave Size: %d\n",
                      id, prop.name, prop.gcnArchName, info.devices[id].cc & 0xffff,
                      device_vmm ? "yes" : "no", prop.warpSize);
#else
        info.devices[id].smpbo = prop.sharedMemPerBlockOptin;
        info.devices[id].cc = 100*prop.major + 10*prop.minor;
        GGML_LOG_INFO("  Device %d: %s, compute capability %d.%d, VMM: %s\n",
                        id, prop.name, prop.major, prop.minor, device_vmm ? "yes" : "no");
#endif // defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__)
    }

    for (int id = 0; id < info.device_count; ++id) {
        info.default_tensor_split[id] /= total_vram;
    }

    // configure logging to stdout
    // CUBLAS_CHECK(cublasLoggerConfigure(1, 1, 0, nullptr));

    return info;
}

const ggml_cuda_device_info & ggml_cuda_info() {
    static ggml_cuda_device_info info = ggml_cuda_init();
    return info;
}

// #define DEBUG_CUDA_MALLOC

// buffer pool for cuda (legacy)
struct ggml_cuda_pool_leg : public ggml_cuda_pool {
    static const int MAX_BUFFERS = 256;

    int device;
    struct ggml_cuda_buffer {
        void * ptr = nullptr;
        size_t size = 0;
    };

    ggml_cuda_buffer buffer_pool[MAX_BUFFERS] = {};
    size_t pool_size = 0;

    explicit ggml_cuda_pool_leg(int device) :
        device(device) {
    }

    ~ggml_cuda_pool_leg() {
        ggml_cuda_set_device(device);
        for (int i = 0; i < MAX_BUFFERS; ++i) {
            ggml_cuda_buffer & b = buffer_pool[i];
            if (b.ptr != nullptr) {
                CUDA_CHECK(cudaFree(b.ptr));
                pool_size -= b.size;
            }
        }
        GGML_ASSERT(pool_size == 0);
    }

    void * alloc(size_t size, size_t * actual_size) override {
#ifdef DEBUG_CUDA_MALLOC
        int nnz = 0;
        size_t max_size = 0;
#endif
        size_t best_diff = 1ull << 36;
        int ibest = -1;
        for (int i = 0; i < MAX_BUFFERS; ++i) {
            ggml_cuda_buffer& b = buffer_pool[i];
            if (b.ptr != nullptr) {
#ifdef DEBUG_CUDA_MALLOC
                ++nnz;
                if (b.size > max_size) max_size = b.size;
#endif
                if (b.size >= size) {
                    size_t diff = b.size - size;
                    if (diff < best_diff) {
                        best_diff = diff;
                        ibest = i;
                        if (!best_diff) {
                            void * ptr = b.ptr;
                            *actual_size = b.size;
                            b.ptr = nullptr;
                            b.size = 0;
                            return ptr;
                        }
                    }
                }
            }
        }
        if (ibest >= 0) {
            ggml_cuda_buffer& b = buffer_pool[ibest];
            void * ptr = b.ptr;
            *actual_size = b.size;
            b.ptr = nullptr;
            b.size = 0;
            return ptr;
        }
        void * ptr;
        size_t look_ahead_size = (size_t) (1.05 * size);
        look_ahead_size = 256 * ((look_ahead_size + 255)/256);
        ggml_cuda_set_device(device);
        CUDA_CHECK(ggml_cuda_device_malloc(&ptr, look_ahead_size, device));
        *actual_size = look_ahead_size;
        pool_size += look_ahead_size;
#ifdef DEBUG_CUDA_MALLOC
        GGML_LOG_INFO("%s[%d]: %d buffers, max_size = %u MB, pool_size = %u MB, requested %u MB\n", __func__, device, nnz,
                           (uint32_t)(max_size / 1024 / 1024), (uint32_t)(pool_size / 1024 / 1024), (uint32_t)(size / 1024 / 1024));
#endif
        return ptr;
    }

    void free(void * ptr, size_t size) override {
        for (int i = 0; i < MAX_BUFFERS; ++i) {
            ggml_cuda_buffer& b = buffer_pool[i];
            if (b.ptr == nullptr) {
                b.ptr = ptr;
                b.size = size;
                return;
            }
        }
        GGML_LOG_DEBUG(GGML_CUDA_NAME " buffer pool full, increase MAX_CUDA_BUFFERS\n");
        ggml_cuda_set_device(device);
        CUDA_CHECK(cudaFree(ptr));
        pool_size -= size;
    }
};

// pool with virtual memory
#if defined(GGML_USE_VMM)
struct ggml_cuda_pool_vmm : public ggml_cuda_pool {
    static const size_t CUDA_POOL_VMM_MAX_SIZE = 1ull << 35; // 32 GB

    int device;
    CUdeviceptr pool_addr = 0;
    size_t pool_used = 0;
    size_t pool_size = 0;
    size_t granularity;
#if defined(GGML_USE_HIP)
    std::vector<std::pair<CUdeviceptr, size_t>> mappings;
#endif

    explicit ggml_cuda_pool_vmm(int device) :
        device(device),
        granularity(ggml_cuda_info().devices[device].vmm_granularity) {
    }

    ~ggml_cuda_pool_vmm() {
        if (pool_addr != 0) {
#if defined(GGML_USE_HIP)
            // Workaround for https://github.com/ROCm/ROCR-Runtime/issues/285
            for (std::pair<CUdeviceptr, size_t> & mapping : mappings) {
                CU_CHECK(cuMemUnmap(mapping.first, mapping.second));
            }
#else
            CU_CHECK(cuMemUnmap(pool_addr, pool_size));
#endif
            CU_CHECK(cuMemAddressFree(pool_addr, CUDA_POOL_VMM_MAX_SIZE));
        }
    }

    void * alloc(size_t size, size_t * actual_size) override {
        // round up the allocation size to the alignment to ensure that all allocations are aligned for all data types
        const size_t alignment = 128;
        size = alignment * ((size + alignment - 1) / alignment);

        size_t avail = pool_size - pool_used;

        if (size > avail) {
            // round up to the next multiple of the granularity
            size_t reserve_size = size - avail;
            reserve_size = granularity * ((reserve_size + granularity - 1) / granularity);

            GGML_ASSERT(pool_size + reserve_size <= CUDA_POOL_VMM_MAX_SIZE);

            // allocate more physical memory
            CUmemAllocationProp prop = {};
            prop.type = CU_MEM_ALLOCATION_TYPE_PINNED;
            prop.location.type = CU_MEM_LOCATION_TYPE_DEVICE;
            prop.location.id = device;
            CUmemGenericAllocationHandle handle;
            CU_CHECK(cuMemCreate(&handle, reserve_size, &prop, 0));

            // reserve virtual address space (if not already reserved)
            if (pool_addr == 0) {
                CU_CHECK(cuMemAddressReserve(&pool_addr, CUDA_POOL_VMM_MAX_SIZE, 0, 0, 0));
            }

            // map at the end of the pool
            CUdeviceptr start_ptr = (CUdeviceptr)((char *)(pool_addr) + pool_size);
            CU_CHECK(cuMemMap(start_ptr, reserve_size, 0, handle, 0));
#if defined(GGML_USE_HIP)
            mappings.push_back({start_ptr, reserve_size});
#endif

            // the memory allocation handle is no longer needed after mapping
            CU_CHECK(cuMemRelease(handle));

            // set access
            CUmemAccessDesc access = {};
            access.location.type = CU_MEM_LOCATION_TYPE_DEVICE;
            access.location.id = device;
            access.flags = CU_MEM_ACCESS_FLAGS_PROT_READWRITE;
            CU_CHECK(cuMemSetAccess((CUdeviceptr)((char *)(pool_addr) + pool_size), reserve_size, &access, 1));

            // add to the pool
            pool_size += reserve_size;

            //printf("cuda pool[%d]: size increased to %llu MB (reserved %llu MB)\n",
            //       device, (unsigned long long) (pool_size/1024/1024),
            //       (unsigned long long) (reserve_size/1024/1024));
        }

        GGML_ASSERT(pool_addr != 0);

        void * ptr = (void *) ((CUdeviceptr)((char *)(pool_addr) + pool_used));
        *actual_size = size;
        pool_used += size;

#ifdef DEBUG_CUDA_MALLOC
        printf("cuda pool[%d]: allocated %llu bytes at %llx\n", device, (unsigned long long) size, ptr);
#endif

        return ptr;
    }

    void free(void * ptr, size_t size) override {
#ifdef DEBUG_CUDA_MALLOC
        printf("cuda pool[%d]: freed %llu bytes at %llx\n", device, (unsigned long long) size, ptr);
#endif

        pool_used -= size;

        // all deallocations must be in reverse order of the allocations
        GGML_ASSERT(ptr == (void *) ((char *)(pool_addr) + pool_used));
    }
};
#endif // defined(GGML_USE_VMM)

std::unique_ptr<ggml_cuda_pool> ggml_backend_cuda_context::new_pool_for_device(int device) {
#if defined(GGML_USE_VMM)
    if (ggml_cuda_info().devices[device].vmm) {
        return std::unique_ptr<ggml_cuda_pool>(new ggml_cuda_pool_vmm(device));
    }
#endif // defined(GGML_USE_VMM)
    return std::unique_ptr<ggml_cuda_pool>(new ggml_cuda_pool_leg(device));
}

// cuda buffer

struct ggml_backend_cuda_buffer_context {
    int device;
    void * dev_ptr = nullptr;
    std::string name;

    ggml_backend_cuda_buffer_context(int device, void * dev_ptr) :
        device(device), dev_ptr(dev_ptr),
        name(GGML_CUDA_NAME + std::to_string(device)) {
    }

    ~ggml_backend_cuda_buffer_context() {
        CUDA_CHECK(cudaFree(dev_ptr));
    }
};

static void ggml_backend_cuda_buffer_free_buffer(ggml_backend_buffer_t buffer) {
    ggml_backend_cuda_buffer_context * ctx = (ggml_backend_cuda_buffer_context *)buffer->context;
    delete ctx;
}

static bool ggml_backend_buffer_is_cuda(ggml_backend_buffer_t buffer) {
    return buffer->iface.free_buffer == ggml_backend_cuda_buffer_free_buffer;
}

static void * ggml_backend_cuda_buffer_get_base(ggml_backend_buffer_t buffer) {
    ggml_backend_cuda_buffer_context * ctx = (ggml_backend_cuda_buffer_context *)buffer->context;
    return ctx->dev_ptr;
}

static void ggml_backend_cuda_buffer_init_tensor(ggml_backend_buffer_t buffer, ggml_tensor * tensor) {
    ggml_backend_cuda_buffer_context * ctx = (ggml_backend_cuda_buffer_context *)buffer->context;

    if (tensor->view_src != NULL) {
        assert(tensor->view_src->buffer->buft == buffer->buft);
        return;
    }

    if (ggml_is_quantized(tensor->type) && tensor->view_src == nullptr && ggml_backend_buffer_get_usage(buffer) != GGML_BACKEND_BUFFER_USAGE_COMPUTE) {
        // initialize padding to 0 to avoid possible NaN values
        size_t original_size = ggml_nbytes(tensor);
        size_t padded_size = ggml_backend_buft_get_alloc_size(buffer->buft, tensor);

        if (padded_size > original_size) {
            ggml_cuda_set_device(ctx->device);
            CUDA_CHECK(cudaMemset((char *)tensor->data + original_size, 0, padded_size - original_size));
        }
    }
}

static void ggml_backend_cuda_buffer_memset_tensor(ggml_backend_buffer_t buffer, ggml_tensor * tensor, uint8_t value, size_t offset, size_t size) {
    ggml_backend_cuda_buffer_context * ctx = (ggml_backend_cuda_buffer_context *)buffer->context;

    ggml_cuda_set_device(ctx->device);
    CUDA_CHECK(cudaMemsetAsync((char *)tensor->data + offset, value, size, cudaStreamPerThread));
    CUDA_CHECK(cudaStreamSynchronize(cudaStreamPerThread));
}

static void ggml_backend_cuda_buffer_set_tensor(ggml_backend_buffer_t buffer, ggml_tensor * tensor, const void * data, size_t offset, size_t size) {
    ggml_backend_cuda_buffer_context * ctx = (ggml_backend_cuda_buffer_context *)buffer->context;

    ggml_cuda_set_device(ctx->device);
    CUDA_CHECK(cudaMemcpyAsync((char *)tensor->data + offset, data, size, cudaMemcpyHostToDevice, cudaStreamPerThread));
    CUDA_CHECK(cudaStreamSynchronize(cudaStreamPerThread));
}

static void ggml_backend_cuda_buffer_get_tensor(ggml_backend_buffer_t buffer, const ggml_tensor * tensor, void * data, size_t offset, size_t size) {
    ggml_backend_cuda_buffer_context * ctx = (ggml_backend_cuda_buffer_context *)buffer->context;

    ggml_cuda_set_device(ctx->device);
    CUDA_CHECK(cudaMemcpyAsync(data, (const char *)tensor->data + offset, size, cudaMemcpyDeviceToHost, cudaStreamPerThread));
    CUDA_CHECK(cudaStreamSynchronize(cudaStreamPerThread));
}

static bool ggml_backend_cuda_buffer_cpy_tensor(ggml_backend_buffer_t buffer, const ggml_tensor * src, ggml_tensor * dst) {
    if (ggml_backend_buffer_is_cuda(src->buffer)) {
        ggml_backend_cuda_buffer_context * src_ctx = (ggml_backend_cuda_buffer_context *)src->buffer->context;
        ggml_backend_cuda_buffer_context * dst_ctx = (ggml_backend_cuda_buffer_context *)dst->buffer->context;
        if (src_ctx->device == dst_ctx->device) {
            CUDA_CHECK(cudaMemcpyAsync(dst->data, src->data, ggml_nbytes(src), cudaMemcpyDeviceToDevice, cudaStreamPerThread));
        } else {
#ifdef GGML_CUDA_NO_PEER_COPY
            return false;
#else
            CUDA_CHECK(cudaMemcpyPeerAsync(dst->data, dst_ctx->device, src->data, src_ctx->device, ggml_nbytes(src), cudaStreamPerThread));
#endif
        }
        CUDA_CHECK(cudaStreamSynchronize(cudaStreamPerThread));
        return true;
    }
    return false;

    GGML_UNUSED(buffer);
}

static void ggml_backend_cuda_buffer_clear(ggml_backend_buffer_t buffer, uint8_t value) {
    ggml_backend_cuda_buffer_context * ctx = (ggml_backend_cuda_buffer_context *)buffer->context;

    ggml_cuda_set_device(ctx->device);
    CUDA_CHECK(cudaDeviceSynchronize());
    CUDA_CHECK(cudaMemset(ctx->dev_ptr, value, buffer->size));
    CUDA_CHECK(cudaDeviceSynchronize());
}

static const ggml_backend_buffer_i ggml_backend_cuda_buffer_interface = {
    /* .free_buffer     = */ ggml_backend_cuda_buffer_free_buffer,
    /* .get_base        = */ ggml_backend_cuda_buffer_get_base,
    /* .init_tensor     = */ ggml_backend_cuda_buffer_init_tensor,
    /* .memset_tensor   = */ ggml_backend_cuda_buffer_memset_tensor,
    /* .set_tensor      = */ ggml_backend_cuda_buffer_set_tensor,
    /* .get_tensor      = */ ggml_backend_cuda_buffer_get_tensor,
    /* .cpy_tensor      = */ ggml_backend_cuda_buffer_cpy_tensor,
    /* .clear           = */ ggml_backend_cuda_buffer_clear,
    /* .reset           = */ NULL,
};

// cuda buffer type
struct ggml_backend_cuda_buffer_type_context {
    int device;
    std::string name;
};

static const char * ggml_backend_cuda_buffer_type_get_name(ggml_backend_buffer_type_t buft) {
    ggml_backend_cuda_buffer_type_context * ctx = (ggml_backend_cuda_buffer_type_context *)buft->context;

    return ctx->name.c_str();
}

static bool ggml_backend_buft_is_cuda(ggml_backend_buffer_type_t buft) {
    return buft->iface.get_name == ggml_backend_cuda_buffer_type_get_name;
}

static ggml_backend_buffer_t ggml_backend_cuda_buffer_type_alloc_buffer(ggml_backend_buffer_type_t buft, size_t size) {
    ggml_backend_cuda_buffer_type_context * buft_ctx = (ggml_backend_cuda_buffer_type_context *)buft->context;

    ggml_cuda_set_device(buft_ctx->device);

    void * dev_ptr;
    cudaError_t err = ggml_cuda_device_malloc(&dev_ptr, size, buft_ctx->device);
    if (err != cudaSuccess) {
        // clear the error
        (void)cudaGetLastError();
        GGML_LOG_ERROR("%s: allocating %.2f MiB on device %d: cudaMalloc failed: %s\n", __func__, size / 1024.0 / 1024.0, buft_ctx->device, cudaGetErrorString(err));
        return nullptr;
    }

    ggml_backend_cuda_buffer_context * ctx = new ggml_backend_cuda_buffer_context(buft_ctx->device, dev_ptr);

    return ggml_backend_buffer_init(buft, ggml_backend_cuda_buffer_interface, ctx, size);
}

static size_t ggml_backend_cuda_buffer_type_get_alignment(ggml_backend_buffer_type_t buft) {
    return 128;

    GGML_UNUSED(buft);
}

static size_t ggml_backend_cuda_buffer_type_get_alloc_size(ggml_backend_buffer_type_t buft, const ggml_tensor * tensor) {
    size_t size = ggml_nbytes(tensor);
    int64_t ne0 = tensor->ne[0];

    if (ggml_is_quantized(tensor->type)) {
        if (ne0 % MATRIX_ROW_PADDING != 0) {
            size += ggml_row_size(tensor->type, MATRIX_ROW_PADDING - ne0 % MATRIX_ROW_PADDING);
        }
    }

    return size;

    GGML_UNUSED(buft);
}

static const ggml_backend_buffer_type_i ggml_backend_cuda_buffer_type_interface = {
    /* .get_name         = */ ggml_backend_cuda_buffer_type_get_name,
    /* .alloc_buffer     = */ ggml_backend_cuda_buffer_type_alloc_buffer,
    /* .get_alignment    = */ ggml_backend_cuda_buffer_type_get_alignment,
    /* .get_max_size     = */ NULL, // defaults to SIZE_MAX
    /* .get_alloc_size   = */ ggml_backend_cuda_buffer_type_get_alloc_size,
    /* .is_host          = */ NULL,
};

ggml_backend_buffer_type_t ggml_backend_cuda_buffer_type(int device) {
    static std::mutex mutex;
    std::lock_guard<std::mutex> lock(mutex);

    if (device >= ggml_backend_cuda_get_device_count()) {
        return nullptr;
    }

    static ggml_backend_buffer_type ggml_backend_cuda_buffer_types[GGML_CUDA_MAX_DEVICES];

    static bool ggml_backend_cuda_buffer_type_initialized = false;

    if (!ggml_backend_cuda_buffer_type_initialized) {
        for (int i = 0; i < ggml_backend_cuda_get_device_count(); i++) {
            ggml_backend_cuda_buffer_types[i] = {
                /* .iface    = */ ggml_backend_cuda_buffer_type_interface,
                /* .device   = */ ggml_backend_reg_dev_get(ggml_backend_cuda_reg(), i),
                /* .context  = */ new ggml_backend_cuda_buffer_type_context{i, GGML_CUDA_NAME + std::to_string(i)},
            };
        }
        ggml_backend_cuda_buffer_type_initialized = true;
    }

    return &ggml_backend_cuda_buffer_types[device];
}

// cuda split buffer

static int64_t get_row_rounding(const std::array<float, GGML_CUDA_MAX_DEVICES> & tensor_split) {
    int64_t row_rounding = 0;
    for (int id = 0; id < ggml_backend_cuda_get_device_count(); ++id) {
        if (tensor_split[id] >= (id + 1 < ggml_backend_cuda_get_device_count() ? tensor_split[id + 1] : 1.0f)) {
            continue;
        }

        const int cc = ggml_cuda_info().devices[id].cc;
        row_rounding = std::max(row_rounding, (int64_t)get_mmq_y_host(cc));
    }
    return row_rounding;
}

static void get_row_split(int64_t * row_low, int64_t * row_high, const ggml_tensor * tensor, const std::array<float, GGML_CUDA_MAX_DEVICES> & tensor_split, int id) {
    const int64_t nrows = ggml_nrows(tensor);
    const int64_t rounding = get_row_rounding(tensor_split);

    *row_low = id == 0 ? 0 : nrows*tensor_split[id];
    *row_low -= *row_low % rounding;

    if (id == ggml_backend_cuda_get_device_count() - 1) {
        *row_high = nrows;
    } else {
        *row_high = nrows*tensor_split[id + 1];
        *row_high -= *row_high % rounding;
    }
}

static size_t ggml_nbytes_split(const struct ggml_tensor * tensor, int nrows_split) {
    static_assert(GGML_MAX_DIMS == 4, "GGML_MAX_DIMS is not 4 - update this function");

    return nrows_split*ggml_row_size(tensor->type, tensor->ne[0]);
}

struct ggml_backend_cuda_split_buffer_type_context {
    int main_device;
    std::array<float, GGML_CUDA_MAX_DEVICES> tensor_split;
    std::string name;
};

struct ggml_backend_cuda_split_buffer_context {
    ~ggml_backend_cuda_split_buffer_context() {
        for (ggml_tensor_extra_gpu * extra : tensor_extras) {
            for (int id = 0; id < GGML_CUDA_MAX_DEVICES; ++id) {
                for (int64_t is = 0; is < GGML_CUDA_MAX_STREAMS; ++is) {
                    if (extra->events[id][is] != nullptr) {
                        CUDA_CHECK(cudaEventDestroy(extra->events[id][is]));
                    }
                }
                if (extra->data_device[id] != nullptr) {
                    CUDA_CHECK(cudaFree(extra->data_device[id]));
                }
            }
            delete extra;
        }
    }

    std::vector<ggml_tensor_extra_gpu *> tensor_extras;
};


static void ggml_backend_cuda_split_buffer_free_buffer(ggml_backend_buffer_t buffer) {
    ggml_backend_cuda_split_buffer_context * ctx = (ggml_backend_cuda_split_buffer_context *)buffer->context;
    delete ctx;
}

static void * ggml_backend_cuda_split_buffer_get_base(ggml_backend_buffer_t buffer) {
    // the pointers are stored in the tensor extras, this is just a dummy address and never dereferenced
    return (void *)0x1000;

    GGML_UNUSED(buffer);
}

static void ggml_backend_cuda_split_buffer_init_tensor(ggml_backend_buffer_t buffer, ggml_tensor * tensor) {
    GGML_ASSERT(tensor->view_src == nullptr); // views of split tensors are not supported

    ggml_backend_cuda_split_buffer_context * ctx = (ggml_backend_cuda_split_buffer_context *)buffer->context;
    ggml_backend_cuda_split_buffer_type_context * buft_ctx = (ggml_backend_cuda_split_buffer_type_context *)buffer->buft->context;

    const int64_t ne0 = tensor->ne[0];

    ggml_tensor_extra_gpu * extra = new ggml_tensor_extra_gpu{};
    ctx->tensor_extras.push_back(extra);

    for (int id = 0; id < ggml_backend_cuda_get_device_count(); ++id) {
        int64_t row_low, row_high;
        get_row_split(&row_low, &row_high, tensor, buft_ctx->tensor_split, id);

        int64_t nrows_split = row_high - row_low;
        if (nrows_split == 0) {
            continue;
        }

        size_t size = ggml_nbytes_split(tensor, nrows_split);
        const size_t original_size = size;

        // pad last row to a multiple of 512 elements to avoid out-of-bounds memory accesses
        if (ne0 % MATRIX_ROW_PADDING != 0) {
            size += ggml_row_size(tensor->type, MATRIX_ROW_PADDING - ne0 % MATRIX_ROW_PADDING);
        }

        // FIXME: do not crash if cudaMalloc fails
        // currently, init_tensor cannot fail, it needs to be fixed in ggml-backend first
        ggml_cuda_set_device(id);
        char * buf;
        CUDA_CHECK(ggml_cuda_device_malloc((void**)&buf, size, id));

        // set padding to 0 to avoid possible NaN values
        if (size > original_size) {
            CUDA_CHECK(cudaMemset(buf + original_size, 0, size - original_size));
        }

        extra->data_device[id] = buf;

        for (int64_t is = 0; is < GGML_CUDA_MAX_STREAMS; ++is) {
            CUDA_CHECK(cudaEventCreateWithFlags(&extra->events[id][is], cudaEventDisableTiming));
        }
    }
    tensor->extra = extra;
}

static void ggml_backend_cuda_split_buffer_set_tensor(ggml_backend_buffer_t buffer, ggml_tensor * tensor, const void * data, size_t offset, size_t size) {
    // split tensors must always be set in their entirety at once
    GGML_ASSERT(offset == 0);
    GGML_ASSERT(size == ggml_nbytes(tensor));

    ggml_backend_cuda_split_buffer_type_context * buft_ctx = (ggml_backend_cuda_split_buffer_type_context *)buffer->buft->context;

    const int64_t ne0 = tensor->ne[0];
    const size_t nb1 = tensor->nb[1];
    ggml_tensor_extra_gpu * extra = (ggml_tensor_extra_gpu *)tensor->extra;

    for (int id = 0; id < ggml_backend_cuda_get_device_count(); ++id) {
        int64_t row_low, row_high;
        get_row_split(&row_low, &row_high, tensor, buft_ctx->tensor_split, id);

        int64_t nrows_split = row_high - row_low;
        if (nrows_split == 0) {
            continue;
        }

        const size_t offset_split = row_low*nb1;
        size_t size = ggml_nbytes_split(tensor, nrows_split);
        const size_t original_size = size;

        // pad last row to a multiple of 512 elements to avoid out-of-bounds memory accesses
        if (ne0 % MATRIX_ROW_PADDING != 0) {
            size += ggml_row_size(tensor->type, MATRIX_ROW_PADDING - ne0 % MATRIX_ROW_PADDING);
        }

        const char * buf_host = (const char *)data + offset_split;
        CUDA_CHECK(cudaMemcpyAsync(extra->data_device[id], buf_host, original_size, cudaMemcpyHostToDevice, cudaStreamPerThread));
    }

    for (int id = 0; id < ggml_backend_cuda_get_device_count(); ++id) {
        CUDA_CHECK(cudaStreamSynchronize(cudaStreamPerThread));
    }
}

static void ggml_backend_cuda_split_buffer_get_tensor(ggml_backend_buffer_t buffer, const ggml_tensor * tensor, void * data, size_t offset, size_t size) {
    // split tensors must always be set in their entirety at once
    GGML_ASSERT(offset == 0);
    GGML_ASSERT(size == ggml_nbytes(tensor));

    ggml_backend_cuda_split_buffer_type_context * buft_ctx = (ggml_backend_cuda_split_buffer_type_context *)buffer->buft->context;

    const int64_t ne0 = tensor->ne[0];
    const size_t nb1 = tensor->nb[1];
    ggml_tensor_extra_gpu * extra = (ggml_tensor_extra_gpu *)tensor->extra;

    for (int id = 0; id < ggml_backend_cuda_get_device_count(); ++id) {
        int64_t row_low, row_high;
        get_row_split(&row_low, &row_high, tensor, buft_ctx->tensor_split, id);

        int64_t nrows_split = row_high - row_low;
        if (nrows_split == 0) {
            continue;
        }

        const size_t offset_split = row_low*nb1;
        size_t size = ggml_nbytes_split(tensor, nrows_split);
        const size_t original_size = size;

        // pad last row to a multiple of 512 elements to avoid out-of-bounds memory accesses
        if (ne0 % MATRIX_ROW_PADDING != 0) {
            size += ggml_row_size(tensor->type, MATRIX_ROW_PADDING - ne0 % MATRIX_ROW_PADDING);
        }

        char * buf_host = (char *)data + offset_split;
        CUDA_CHECK(cudaMemcpyAsync(buf_host, extra->data_device[id], original_size, cudaMemcpyDeviceToHost, cudaStreamPerThread));
    }

    for (int id = 0; id < ggml_backend_cuda_get_device_count(); ++id) {
        CUDA_CHECK(cudaStreamSynchronize(cudaStreamPerThread));
    }
}

static void ggml_backend_cuda_split_buffer_clear(ggml_backend_buffer_t buffer, uint8_t value) {
    GGML_UNUSED(buffer);
    GGML_UNUSED(value);
}

static const ggml_backend_buffer_i ggml_backend_cuda_split_buffer_interface = {
    /* .free_buffer     = */ ggml_backend_cuda_split_buffer_free_buffer,
    /* .get_base        = */ ggml_backend_cuda_split_buffer_get_base,
    /* .init_tensor     = */ ggml_backend_cuda_split_buffer_init_tensor,
    /* .memset_tensor   = */ NULL,
    /* .set_tensor      = */ ggml_backend_cuda_split_buffer_set_tensor,
    /* .get_tensor      = */ ggml_backend_cuda_split_buffer_get_tensor,
    /* .cpy_tensor      = */ NULL,
    /* .clear           = */ ggml_backend_cuda_split_buffer_clear,
    /* .reset           = */ NULL,
};

// cuda split buffer type

static const char * ggml_backend_cuda_split_buffer_type_get_name(ggml_backend_buffer_type_t buft) {
    ggml_backend_cuda_split_buffer_type_context * ctx = (ggml_backend_cuda_split_buffer_type_context *)buft->context;

    return ctx->name.c_str();
}

static bool ggml_backend_buft_is_cuda_split(ggml_backend_buffer_type_t buft) {
    return buft->iface.get_name == ggml_backend_cuda_split_buffer_type_get_name;
}

static ggml_backend_buffer_t ggml_backend_cuda_split_buffer_type_alloc_buffer(ggml_backend_buffer_type_t buft, size_t size) {
    // since we don't know the exact split after rounding, we cannot allocate the device buffers at this point
    // instead, we allocate them for each tensor separately in init_tensor
    // however, the size still represents the maximum cumulative size of all the device buffers after the tensors are allocated,
    // as returned by get_alloc_size. this limit is enforced during tensor allocation by ggml-alloc, so it must be correct.
    ggml_backend_cuda_split_buffer_context * ctx = new ggml_backend_cuda_split_buffer_context();

    return ggml_backend_buffer_init(buft, ggml_backend_cuda_split_buffer_interface, ctx, size);
}

static size_t ggml_backend_cuda_split_buffer_type_get_alignment(ggml_backend_buffer_type_t buft) {
    return 128;

    GGML_UNUSED(buft);
}

static size_t ggml_backend_cuda_split_buffer_type_get_alloc_size(ggml_backend_buffer_type_t buft, const ggml_tensor * tensor) {
    ggml_backend_cuda_split_buffer_type_context * ctx = (ggml_backend_cuda_split_buffer_type_context *)buft->context;

    size_t total_size = 0;

    const int64_t ne0 = tensor->ne[0];

    for (int id = 0; id < ggml_backend_cuda_get_device_count(); ++id) {
        int64_t row_low, row_high;
        get_row_split(&row_low, &row_high, tensor, ctx->tensor_split, id);

        int64_t nrows_split = row_high - row_low;
        if (nrows_split == 0) {
            continue;
        }

        total_size += ggml_nbytes_split(tensor, nrows_split);

        // pad last row to a multiple of 512 elements to avoid out-of-bounds memory accesses
        if (ne0 % MATRIX_ROW_PADDING != 0) {
            total_size += ggml_row_size(tensor->type, MATRIX_ROW_PADDING - ne0 % MATRIX_ROW_PADDING);
        }
    }

    return total_size;
}

static bool ggml_backend_cuda_split_buffer_type_is_host(ggml_backend_buffer_type_t buft) {
    return false;

    GGML_UNUSED(buft);
}

static const ggml_backend_buffer_type_i ggml_backend_cuda_split_buffer_type_interface = {
    /* .get_name         = */ ggml_backend_cuda_split_buffer_type_get_name,
    /* .alloc_buffer     = */ ggml_backend_cuda_split_buffer_type_alloc_buffer,
    /* .get_alignment    = */ ggml_backend_cuda_split_buffer_type_get_alignment,
    /* .get_max_size     = */ NULL, // defaults to SIZE_MAX
    /* .get_alloc_size   = */ ggml_backend_cuda_split_buffer_type_get_alloc_size,
    /* .is_host          = */ ggml_backend_cuda_split_buffer_type_is_host,
};

ggml_backend_buffer_type_t ggml_backend_cuda_split_buffer_type(int main_device, const float * tensor_split) {
    static std::mutex mutex;
    std::lock_guard<std::mutex> lock(mutex);

    static std::map<std::pair<int, std::array<float, GGML_CUDA_MAX_DEVICES>>, struct ggml_backend_buffer_type> buft_map;

    std::array<float, GGML_CUDA_MAX_DEVICES> tensor_split_arr = {};

    bool all_zero = tensor_split == nullptr || std::all_of(tensor_split, tensor_split + GGML_CUDA_MAX_DEVICES, [](float x) { return x == 0.0f; });
    if (all_zero) {
        tensor_split_arr = ggml_cuda_info().default_tensor_split;
    } else {
        float split_sum = 0.0f;
        for (int i = 0; i < ggml_backend_cuda_get_device_count(); ++i) {
            tensor_split_arr[i] = split_sum;
            split_sum += tensor_split[i];
        }
        for (int i = 0; i < ggml_backend_cuda_get_device_count(); ++i) {
            tensor_split_arr[i] /= split_sum;
        }
    }

    auto it = buft_map.find({main_device, tensor_split_arr});
    if (it != buft_map.end()) {
        return &it->second;
    }
    auto * ctx = new ggml_backend_cuda_split_buffer_type_context{
        main_device,
        tensor_split_arr,
        GGML_CUDA_NAME + std::to_string(main_device) + "_Split",
    };

    struct ggml_backend_buffer_type buft {
        /* .iface   = */ ggml_backend_cuda_split_buffer_type_interface,
        /* .device  = */ ggml_backend_reg_dev_get(ggml_backend_cuda_reg(), main_device),
        /* .context = */ ctx,
    };

    auto result = buft_map.emplace(std::make_pair(main_device, tensor_split_arr), buft);
    return &result.first->second;
}

// host buffer type

static const char * ggml_backend_cuda_host_buffer_type_name(ggml_backend_buffer_type_t buft) {
    return GGML_CUDA_NAME "_Host";

    GGML_UNUSED(buft);
}

static void ggml_backend_cuda_host_buffer_free_buffer(ggml_backend_buffer_t buffer) {
    CUDA_CHECK(cudaFreeHost(buffer->context));
}

static void * ggml_cuda_host_malloc(size_t size) {
    if (getenv("GGML_CUDA_NO_PINNED") != nullptr) {
        return nullptr;
    }

    void * ptr = nullptr;
    cudaError_t err = cudaMallocHost((void **) &ptr, size);
    if (err != cudaSuccess) {
        // clear the error
        (void)cudaGetLastError();
        GGML_LOG_DEBUG("%s: failed to allocate %.2f MiB of pinned memory: %s\n", __func__,
                           size / 1024.0 / 1024.0, cudaGetErrorString(err));
        return nullptr;
    }

    return ptr;
}

static ggml_backend_buffer_t ggml_backend_cuda_host_buffer_type_alloc_buffer(ggml_backend_buffer_type_t buft, size_t size) {
    void * ptr = ggml_cuda_host_malloc(size);

    if (ptr == nullptr) {
        // fallback to cpu buffer
        return ggml_backend_buft_alloc_buffer(ggml_backend_cpu_buffer_type(), size);
    }

    ggml_backend_buffer_t buffer = ggml_backend_cpu_buffer_from_ptr(ptr, size);
    buffer->buft = buft;
    buffer->iface.free_buffer = ggml_backend_cuda_host_buffer_free_buffer;

    return buffer;
}

ggml_backend_buffer_type_t ggml_backend_cuda_host_buffer_type() {
    static struct ggml_backend_buffer_type ggml_backend_cuda_buffer_type_host = {
        /* .iface    = */ {
            /* .get_name         = */ ggml_backend_cuda_host_buffer_type_name,
            /* .alloc_buffer     = */ ggml_backend_cuda_host_buffer_type_alloc_buffer,
            /* .get_alignment    = */ ggml_backend_cpu_buffer_type()->iface.get_alignment,
            /* .get_max_size     = */ NULL, // defaults to SIZE_MAX
            /* .get_alloc_size   = */ ggml_backend_cpu_buffer_type()->iface.get_alloc_size,
            /* .is_host          = */ ggml_backend_cpu_buffer_type()->iface.is_host,
        },
        /* .device   = */ ggml_backend_reg_dev_get(ggml_backend_cuda_reg(), 0),
        /* .context  = */ nullptr,
    };

    return &ggml_backend_cuda_buffer_type_host;
}

//static bool ggml_backend_buffer_is_cuda_host(ggml_backend_buffer_t buffer) {
//    return buffer->buft->iface.get_name == ggml_backend_cuda_host_buffer_type_name;
//}

/// kernels

typedef void (*ggml_cuda_op_mul_mat_t)(
    ggml_backend_cuda_context & ctx,
    const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst, const char * src0_dd_i, const float * src1_ddf_i,
    const char * src1_ddq_i, float * dst_dd_i, const int64_t row_low, const int64_t row_high, const int64_t src1_ncols,
    const int64_t src1_padded_row_size, cudaStream_t stream);

#ifndef GGML_CUDA_PEER_MAX_BATCH_SIZE
#define GGML_CUDA_PEER_MAX_BATCH_SIZE 128
#endif // GGML_CUDA_PEER_MAX_BATCH_SIZE

#define MUL_MAT_SRC1_COL_STRIDE 128

static cudaError_t ggml_cuda_cpy_tensor_2d(
    void * dst, const struct ggml_tensor * src, int64_t i3, int64_t i2, int64_t i1_low, int64_t i1_high, cudaStream_t stream) {

    GGML_ASSERT(ggml_backend_buffer_is_cuda(src->buffer));
    const char * src_ptr = (const char *) src->data;
    char       * dst_ptr = (char       *) dst;

    const int64_t ne0 = src->ne[0];
    const int64_t nb0 = src->nb[0];
    const int64_t nb1 = src->nb[1];
    const int64_t nb2 = src->nb[2];
    const int64_t nb3 = src->nb[3];
    const enum ggml_type type = src->type;
    const int64_t ts = ggml_type_size(type);
    const int64_t bs = ggml_blck_size(type);
    const int64_t i1_diff = i1_high - i1_low;

    const char * x = src_ptr + i1_low*nb1 + i2*nb2 + i3*nb3;
    if (nb0 == ts && nb1 == ts*ne0/bs) {
        return cudaMemcpyAsync(dst_ptr, x, i1_diff*nb1, cudaMemcpyDeviceToDevice, stream);
    } else if (nb0 == ts) {
        return cudaMemcpy2DAsync(dst_ptr, ts*ne0/bs, x, nb1, ts*ne0/bs, i1_diff, cudaMemcpyDeviceToDevice, stream);
    } else {
        for (int64_t i1 = 0; i1 < i1_diff; i1++) {
            const void * rx = (const void *) ((const char *) x + i1*nb1);
            void * rd = (void *) (dst_ptr + i1*ts*ne0/bs);
            // pretend the row is a matrix with cols=1
            cudaError_t r = cudaMemcpy2DAsync(rd, ts/bs, rx, nb0, ts/bs, ne0, cudaMemcpyDeviceToDevice, stream);
            if (r != cudaSuccess) {
                return r;
            }
        }
        return cudaSuccess;
    }
}

static void ggml_cuda_op_mul_mat_cublas(
    ggml_backend_cuda_context & ctx,
    const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst, const char * src0_dd_i, const float * src1_ddf_i,
    const char * src1_ddq_i, float * dst_dd_i, const int64_t row_low, const int64_t row_high, const int64_t src1_ncols,
    const int64_t src1_padded_row_size, cudaStream_t stream) {

    GGML_ASSERT(src0_dd_i  != nullptr);
    GGML_ASSERT(src1_ddf_i != nullptr);
    GGML_ASSERT(dst_dd_i   != nullptr);

    const int64_t ne00 = src0->ne[0];
    const int64_t ne10 = src1->ne[0];

    const int64_t ne0 = dst->ne[0];

    const int64_t row_diff = row_high - row_low;

    int id = ggml_cuda_get_device();

    // the main device has a larger memory buffer to hold the results from all GPUs
    // ldc == nrows of the matrix that cuBLAS writes into
    int64_t ldc = id == ctx.device ? ne0 : row_diff;

    const int compute_capability = ggml_cuda_info().devices[id].cc;

    const bool use_fp16 = (src0->type == GGML_TYPE_F16 || ggml_is_quantized(src0->type)) && ggml_is_contiguous(src0) && row_diff == src0->ne[1] && dst->op_params[0] == GGML_PREC_DEFAULT;

    if (compute_capability >= GGML_CUDA_CC_VOLTA && use_fp16) {
        // convert src0 and src1 to fp16, multiply as fp16, convert dst to fp32
        ggml_cuda_pool_alloc<half> src0_as_f16(ctx.pool(id));
        if (src0->type != GGML_TYPE_F16) {
            const to_fp16_cuda_t to_fp16_cuda = ggml_get_to_fp16_cuda(src0->type);
            GGML_ASSERT(to_fp16_cuda != nullptr);
            size_t ne = row_diff*ne00;
            src0_as_f16.alloc(ne);
            to_fp16_cuda(src0_dd_i, src0_as_f16.get(), ne, stream);
        }
        const half * src0_ptr = src0->type == GGML_TYPE_F16 ? (const half *) src0_dd_i : src0_as_f16.get();

        ggml_cuda_pool_alloc<half> src1_as_f16(ctx.pool(id));
        if (src1->type != GGML_TYPE_F16) {
            const to_fp16_cuda_t to_fp16_cuda = ggml_get_to_fp16_cuda(src1->type);
            GGML_ASSERT(to_fp16_cuda != nullptr);
            size_t ne = src1_ncols*ne10;
            src1_as_f16.alloc(ne);
            to_fp16_cuda(src1_ddf_i, src1_as_f16.get(), ne, stream);
        }
        const half * src1_ptr = src1->type == GGML_TYPE_F16 ? (const half *) src1_ddf_i : src1_as_f16.get();

        CUBLAS_CHECK(cublasSetStream(ctx.cublas_handle(id), stream));

        if (GGML_CUDA_CC_IS_CDNA(compute_capability)) {
            const float alpha = 1.0f;
            const float beta = 0.0f;
            CUBLAS_CHECK(
                cublasGemmEx(ctx.cublas_handle(id), CUBLAS_OP_T, CUBLAS_OP_N,
                        row_diff, src1_ncols, ne10,
                        &alpha, src0_ptr,  CUDA_R_16F, ne00,
                                src1_ptr,  CUDA_R_16F, ne10,
                        &beta,   dst_dd_i, CUDA_R_32F, ldc,
                        CUBLAS_COMPUTE_32F,
                        CUBLAS_GEMM_DEFAULT_TENSOR_OP));
        } else {
            ggml_cuda_pool_alloc<half> dst_f16(ctx.pool(id), row_diff*src1_ncols);

            const half alpha_f16 = 1.0f;
            const half beta_f16 = 0.0f;

            CUBLAS_CHECK(
                cublasGemmEx(ctx.cublas_handle(id), CUBLAS_OP_T, CUBLAS_OP_N,
                        row_diff, src1_ncols, ne10,
                        &alpha_f16, src0_ptr,      CUDA_R_16F, ne00,
                                    src1_ptr,      CUDA_R_16F, ne10,
                        &beta_f16,  dst_f16.get(), CUDA_R_16F, ldc,
                        CUBLAS_COMPUTE_16F,
                        CUBLAS_GEMM_DEFAULT_TENSOR_OP));

            const to_fp32_cuda_t to_fp32_cuda = ggml_get_to_fp32_cuda(GGML_TYPE_F16);
            to_fp32_cuda(dst_f16.get(), dst_dd_i, row_diff*src1_ncols, stream);
        }
    } else {
        ggml_cuda_pool_alloc<float> src0_ddq_as_f32(ctx.pool(id));
        ggml_cuda_pool_alloc<float> src1_ddq_as_f32(ctx.pool(id));

        if (src0->type != GGML_TYPE_F32) {
            const to_fp32_cuda_t to_fp32_cuda = ggml_get_to_fp32_cuda(src0->type);
            GGML_ASSERT(to_fp32_cuda != nullptr);
            src0_ddq_as_f32.alloc(row_diff*ne00);
            to_fp32_cuda(src0_dd_i, src0_ddq_as_f32.get(), row_diff*ne00, stream);
        }
        if (src1->type != GGML_TYPE_F32) {
            const to_fp32_cuda_t to_fp32_cuda = ggml_get_to_fp32_cuda(src1->type);
            GGML_ASSERT(to_fp32_cuda != nullptr);
            src1_ddq_as_f32.alloc(src1_ncols*ne10);
            to_fp32_cuda(src1_ddf_i, src1_ddq_as_f32.get(), src1_ncols*ne10, stream);
        }

        const float * src0_ddf_i = src0->type == GGML_TYPE_F32 ? (const float *) src0_dd_i : src0_ddq_as_f32.get();
        const float * src1_ddf1_i = src1->type == GGML_TYPE_F32 ? (const float *) src1_ddf_i : src1_ddq_as_f32.get();

        const float alpha = 1.0f;
        const float beta = 0.0f;

        CUBLAS_CHECK(cublasSetStream(ctx.cublas_handle(id), stream));
        CUBLAS_CHECK(
            cublasSgemm(ctx.cublas_handle(id), CUBLAS_OP_T, CUBLAS_OP_N,
                    row_diff, src1_ncols, ne10,
                    &alpha, src0_ddf_i,  ne00,
                            src1_ddf1_i, ne10,
                    &beta,  dst_dd_i,    ldc));
    }

    GGML_UNUSED(dst);
    GGML_UNUSED(src1_ddq_i);
    GGML_UNUSED(src1_padded_row_size);
}

static void ggml_cuda_set_peer_access(const int n_tokens, int main_device) {
    static bool peer_access_enabled = false;

    const bool enable_peer_access = n_tokens <= GGML_CUDA_PEER_MAX_BATCH_SIZE;

    if (peer_access_enabled == enable_peer_access) {
        return;
    }

#ifdef NDEBUG
    for (int id = 0; id < ggml_backend_cuda_get_device_count(); ++id) {
        ggml_cuda_set_device(id);
        CUDA_CHECK(cudaDeviceSynchronize());
    }

    for (int id = 0; id < ggml_backend_cuda_get_device_count(); ++id) {
        ggml_cuda_set_device(id);

        for (int id_other = 0; id_other < ggml_backend_cuda_get_device_count(); ++id_other) {
            if (id == id_other) {
                continue;
            }
            if (id != main_device && id_other != main_device) {
                continue;
            }

            int can_access_peer;
            CUDA_CHECK(cudaDeviceCanAccessPeer(&can_access_peer, id, id_other));
            if (can_access_peer) {
                if (enable_peer_access) {
                    cudaError_t err = cudaDeviceEnablePeerAccess(id_other, 0);
                    if (err != cudaErrorPeerAccessAlreadyEnabled) {
                        CUDA_CHECK(err);
                    } else {
                        // reset the error
                        (void)cudaGetLastError();
                    }
                } else {
                    cudaError_t err = cudaDeviceDisablePeerAccess(id_other);
                    if (err != cudaErrorPeerAccessNotEnabled) {
                        CUDA_CHECK(err);
                    } else {
                        // reset the error
                        (void)cudaGetLastError();
                    }
                }
            }
        }
    }

    ggml_cuda_set_device(main_device);
#endif // NDEBUG

    peer_access_enabled = enable_peer_access;

    GGML_UNUSED(main_device);
}

static cudaError_t ggml_cuda_Memcpy2DPeerAsync(
    void * dst, int dstDevice, size_t dpitch, void * src, int srcDevice, size_t spitch, size_t width, size_t height, cudaStream_t stream) {

#if !defined(GGML_USE_HIP) && !defined(GGML_USE_MUSA)
    // cudaMemcpy2DAsync may fail with copies between vmm pools of different devices
    cudaMemcpy3DPeerParms p = {};
    p.dstDevice = dstDevice;
    p.dstPtr = make_cudaPitchedPtr(dst, dpitch, dpitch, height);
    p.srcDevice = srcDevice;
    p.srcPtr = make_cudaPitchedPtr(src, spitch, spitch, height);
    p.extent = make_cudaExtent(width, height, 1);
    return cudaMemcpy3DPeerAsync(&p, stream);
#else
    // HIP does not support cudaMemcpy3DPeerAsync or vmm pools
    GGML_UNUSED(dstDevice);
    GGML_UNUSED(srcDevice);
    return cudaMemcpy2DAsync(dst, dpitch, src, spitch, width, height, cudaMemcpyDeviceToDevice, stream);
#endif // !defined(GGML_USE_HIP) && !defined(GGML_USE_MUSA)
}

static void ggml_cuda_op_mul_mat(
    ggml_backend_cuda_context & ctx,
    const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst, ggml_cuda_op_mul_mat_t op,
    quantize_cuda_t quantize_src1) {

    const int64_t ne00 = src0->ne[0];
    const int64_t ne01 = src0->ne[1];
    const int64_t ne02 = src0->ne[2];
    const int64_t ne03 = src0->ne[3];

    const int64_t ne10 = src1->ne[0];
    const int64_t ne11 = src1->ne[1];
    const int64_t ne12 = src1->ne[2];
    const int64_t ne13 = src1->ne[3];
    const int64_t nrows1 = ggml_nrows(src1);

    const int64_t ne0 = dst->ne[0];
    const int64_t ne1 = dst->ne[1];

    const int64_t nb2 = dst->nb[2];
    const int64_t nb3 = dst->nb[3];

    GGML_ASSERT(ggml_backend_buffer_is_cuda(dst->buffer));
    GGML_ASSERT(ggml_backend_buffer_is_cuda(src1->buffer));
    ggml_backend_cuda_buffer_context * src1_ctx = (ggml_backend_cuda_buffer_context *) src1->buffer->context;
    ggml_backend_cuda_buffer_context * dst_ctx  = (ggml_backend_cuda_buffer_context *) dst->buffer->context;

    GGML_ASSERT(src1->type == GGML_TYPE_F32 || (src1->ne[2] == 1 && src1->ne[3] == 1));

    GGML_ASSERT(ne12 % ne02 == 0);
    GGML_ASSERT(ne13 % ne03 == 0);

    const int64_t i02_divisor = ne12 / ne02;
    const int64_t i03_divisor = ne13 / ne03;

    const size_t src0_ts = ggml_type_size(src0->type);
    const size_t src0_bs = ggml_blck_size(src0->type);
    const size_t q8_1_ts = sizeof(block_q8_1);
    const size_t q8_1_bs = QK8_1;

    const bool src0_is_contiguous = ggml_is_contiguous(src0);
    const bool src1_is_contiguous = ggml_is_contiguous(src1);

    const int64_t src1_padded_col_size = GGML_PAD(ne10, MATRIX_ROW_PADDING);

    const bool split = ggml_backend_buft_is_cuda_split(src0->buffer->buft);
    GGML_ASSERT(!(split && ne02 > 1));
    GGML_ASSERT(!(split && ne03 > 1));
    GGML_ASSERT(!(split && ne02 < ne12));
    GGML_ASSERT(!(split && ne03 < ne13));

    ggml_tensor_extra_gpu * src0_extra = split ? (ggml_tensor_extra_gpu *) src0->extra : nullptr;


    std::array<float, GGML_CUDA_MAX_DEVICES> tensor_split;
    if (split) {
        ggml_backend_cuda_split_buffer_type_context * buft_ctx = (ggml_backend_cuda_split_buffer_type_context *) src0->buffer->buft->context;
        tensor_split = buft_ctx->tensor_split;
    }

    struct dev_data {
        int cc;

        ggml_cuda_pool_alloc<char>   src0_dd_alloc;
        ggml_cuda_pool_alloc<float> src1_ddf_alloc;
        ggml_cuda_pool_alloc<char>  src1_ddq_alloc;
        ggml_cuda_pool_alloc<float>   dst_dd_alloc;

        char  *  src0_dd = nullptr;
        float * src1_ddf = nullptr; // float
        char  * src1_ddq = nullptr; // q8_1
        float *   dst_dd = nullptr;

        int64_t  row_low;
        int64_t row_high;
    };

    dev_data dev[GGML_CUDA_MAX_DEVICES];

    int used_devices = 0;

    for (int id = 0; id < ggml_backend_cuda_get_device_count(); ++id) {
        dev[id].cc = ggml_cuda_info().devices[id].cc;

        // by default, use all rows
        dev[id].row_low  = 0;
        dev[id].row_high = ne01;

        // for multi GPU, get the row boundaries from tensor split
        // and round to mul_mat_q tile sizes
        if (split) {
            const int64_t rounding = get_row_rounding(tensor_split);

            if (id != 0) {
                dev[id].row_low  = ne01*tensor_split[id];
                if (dev[id].row_low < ne01) {
                    dev[id].row_low -= dev[id].row_low % rounding;
                }
            }

            if (id != ggml_backend_cuda_get_device_count() - 1) {
                dev[id].row_high  = ne01*tensor_split[id + 1];
                if (dev[id].row_high < ne01) {
                    dev[id].row_high -= dev[id].row_high % rounding;
                }
            }
        }
    }

    for (int id = 0; id < ggml_backend_cuda_get_device_count(); ++id) {
        if ((!split && id != ctx.device) || dev[id].row_low == dev[id].row_high) {
            continue;
        }

        used_devices++;

        const bool src1_on_device = id == src1_ctx->device;
        const bool  dst_on_device = id == dst_ctx->device;

        ggml_cuda_set_device(id);
        cudaStream_t stream = ctx.stream(id, 0);

        if (src0_is_contiguous) {
            dev[id].src0_dd = split ? (char *) src0_extra->data_device[id] : (char *) src0->data;
        } else {
            // If src0 is not contiguous it will be copied to a temporary buffer.
            // This buffer needs to be cleared entirely because multiple regions will function as padding.
            const size_t nbytes_data    = ggml_nbytes(src0);
            const size_t nbytes_padding = ggml_row_size(src0->type, MATRIX_ROW_PADDING - ne00 % MATRIX_ROW_PADDING);
            dev[id].src0_dd = dev[id].src0_dd_alloc.alloc(ctx.pool(id), nbytes_data + nbytes_padding);
        // TODO: remove this for MUSA once the Guilty Lockup issue is resolved
#ifndef GGML_USE_MUSA
            CUDA_CHECK(cudaMemsetAsync(dev[id].src0_dd, 0, nbytes_data + nbytes_padding, stream));
#else // GGML_USE_MUSA
            CUDA_CHECK(cudaMemsetAsync(dev[id].src0_dd + nbytes_data, 0, nbytes_padding, stream));
#endif // !GGML_USE_MUSA
        }

        // If src0 is on a temporary compute buffer (partial offloading) there may be some padding that needs to be cleared:
        if (ne00 % MATRIX_ROW_PADDING != 0 && ggml_is_quantized(src0->type) && ggml_backend_buffer_get_usage(src0->buffer) == GGML_BACKEND_BUFFER_USAGE_COMPUTE && src0->view_src == nullptr) {
            const size_t nbytes_data    = ggml_row_size(src0->type, (dev[id].row_high - dev[id].row_low)*ne00);
            const size_t nbytes_padding = ggml_row_size(src0->type, MATRIX_ROW_PADDING - ne00 % MATRIX_ROW_PADDING);
            CUDA_CHECK(cudaMemsetAsync(dev[id].src0_dd + nbytes_data, 0, nbytes_padding, stream));
        }

        if (src1_on_device && src1_is_contiguous) {
            dev[id].src1_ddf = (float *) src1->data;
        } else {
            dev[id].src1_ddf = dev[id].src1_ddf_alloc.alloc(ctx.pool(id), ggml_nelements(src1));
        }

        if (quantize_src1) {
            size_t src_1_ddq_size = nrows1*src1_padded_col_size*q8_1_ts/q8_1_bs;
            if (quantize_src1 == quantize_mmq_q8_1_cuda) {
                src_1_ddq_size += get_mmq_x_max_host(dev[id].cc)*sizeof(block_q8_1_mmq);
            }
            dev[id].src1_ddq = dev[id].src1_ddq_alloc.alloc(ctx.pool(id), src_1_ddq_size);

            if (src1_on_device && src1_is_contiguous) {
                quantize_src1(dev[id].src1_ddf, dev[id].src1_ddq, ne10, ne11, ne12*ne13, src1_padded_col_size, src0->type, stream);
                CUDA_CHECK(cudaGetLastError());
            }
        }

        if (dst_on_device) {
            dev[id].dst_dd = (float *) dst->data;
        } else {
            const size_t size_dst_ddf = split ? (dev[id].row_high - dev[id].row_low)*ne1 : ggml_nelements(dst);
            dev[id].dst_dd = dev[id].dst_dd_alloc.alloc(ctx.pool(id), size_dst_ddf);
        }
    }

    // if multiple devices are used they need to wait for the main device
    // here an event is recorded that signals that the main device has finished calculating the input data
    if (split && used_devices > 1) {
        ggml_cuda_set_device(ctx.device);
        CUDA_CHECK(cudaEventRecord(src0_extra->events[ctx.device][0], ctx.stream()));
    }

    const int64_t src1_col_stride = split && used_devices > 1 ? MUL_MAT_SRC1_COL_STRIDE : ne11;
    for (int64_t src1_col_0 = 0; src1_col_0 < ne11; src1_col_0 += src1_col_stride) {
        const int64_t is = split ? (src1_col_0/src1_col_stride) % GGML_CUDA_MAX_STREAMS : 0;
        const int64_t src1_ncols = src1_col_0 + src1_col_stride > ne11 ? ne11 - src1_col_0 : src1_col_stride;

        for (int id = 0; id < ggml_backend_cuda_get_device_count(); ++id) {
            if ((!split && id != ctx.device) || dev[id].row_low == dev[id].row_high) {
                continue;
            }

            const bool src1_on_device = id == src1_ctx->device;
            const bool  dst_on_device = id == dst_ctx->device;
            const int64_t row_diff = dev[id].row_high - dev[id].row_low;

            ggml_cuda_set_device(id);
            cudaStream_t stream = ctx.stream(id, is);

            // wait for main GPU data if necessary
            if (split && (id != ctx.device || is != 0)) {
                CUDA_CHECK(cudaStreamWaitEvent(stream, src0_extra->events[ctx.device][0], 0));
            }

            for (int64_t i0 = 0; i0 < ne13*ne12; ++i0) {
                const int64_t i03 = i0 / ne12;
                const int64_t i02 = i0 % ne12;

                size_t src1_ddq_i_offset = i0*ne11 * src1_padded_col_size*q8_1_ts/q8_1_bs;
                if (quantize_src1 == quantize_mmq_q8_1_cuda) {
                    src1_ddq_i_offset += src1_col_0 * sizeof(block_q8_1_mmq);
                } else {
                    src1_ddq_i_offset += src1_col_0 * src1_padded_col_size*q8_1_ts/q8_1_bs;
                }

                // for split tensors the data begins at i0 == i0_offset_low
                const size_t nbytes_src0_matrix = ne01*ne00*src0_ts / src0_bs;
                char  *  src0_dd_i =  dev[id].src0_dd + ((i03/i03_divisor)*ne02 + (i02/i02_divisor)) * nbytes_src0_matrix;
                float * src1_ddf_i = dev[id].src1_ddf + (i0*ne11 + src1_col_0) * ne10;
                char  * src1_ddq_i = dev[id].src1_ddq +  src1_ddq_i_offset;
                float *   dst_dd_i =   dev[id].dst_dd + (i0*ne1  + src1_col_0) * (dst_on_device ? ne0 : row_diff);

                // the main device memory buffer can be on VRAM scratch, with space for all partial results
                // in that case an offset on dst_ddf_i is needed
                if (id == ctx.device) {
                    dst_dd_i += dev[id].row_low; // offset is 0 if no tensor split
                }

                // copy src0, src1 to device if necessary
                if (src1_is_contiguous) {
                    if (id != ctx.device) {
                        if (quantize_src1) {
                            char * src1_ddq_i_source = dev[ctx.device].src1_ddq + src1_ddq_i_offset;
                            if (quantize_src1 == quantize_mmq_q8_1_cuda) {
                                const size_t pitch = ne11*sizeof(block_q8_1_mmq);
                                const size_t width = src1_ncols*sizeof(block_q8_1_mmq);
                                const size_t height = src1_padded_col_size/(4*QK8_1);
                                CUDA_CHECK(ggml_cuda_Memcpy2DPeerAsync(src1_ddq_i, id, pitch, src1_ddq_i_source, ctx.device, pitch, width, height, stream));
                            } else {
                                CUDA_CHECK(cudaMemcpyPeerAsync(
                                    src1_ddq_i, id, src1_ddq_i_source, ctx.device, src1_ncols*src1_padded_col_size*q8_1_ts/q8_1_bs, stream));
                            }
                        } else {
                            float * src1_ddf_i_source = (float *) src1->data;
                            src1_ddf_i_source += (i0*ne11 + src1_col_0) * ne10;
                            CUDA_CHECK(cudaMemcpyPeerAsync(src1_ddf_i, id, src1_ddf_i_source, ctx.device,
                                                            src1_ncols*ne10*sizeof(float), stream));
                        }
                    }
                } else if (src1_on_device && !src1_is_contiguous) {
                    CUDA_CHECK(ggml_cuda_cpy_tensor_2d(
                                src1_ddf_i, src1, i03, i02, src1_col_0, src1_col_0+src1_ncols, stream));
                } else {
                    GGML_ABORT("fatal error");
                }

                if (quantize_src1 && !src1_is_contiguous) {
                    quantize_src1(src1_ddf_i, src1_ddq_i, ne10, src1_ncols, 1, src1_padded_col_size, src0->type, stream);
                    CUDA_CHECK(cudaGetLastError());
                }

                if (src1_col_0 == 0 && !src0_is_contiguous && i03 % i03_divisor == 0 && i02 % i02_divisor == 0) {
                    CUDA_CHECK(ggml_cuda_cpy_tensor_2d(
                        src0_dd_i, src0, i03/i03_divisor, i02/i02_divisor, dev[id].row_low, dev[id].row_high, stream));
                }

                // do the computation
                op(ctx, src0, src1, dst, src0_dd_i, src1_ddf_i, src1_ddq_i, dst_dd_i,
                    dev[id].row_low, dev[id].row_high, src1_ncols, src1_padded_col_size, stream);
                CUDA_CHECK(cudaGetLastError());

                // copy dst to host or other device if necessary
                if (!dst_on_device) {
                    void * dst_off_device = dst->data;
                    if (split) {
                        // src0 = weight matrix is saved as a transposed matrix for better memory layout.
                        // dst is NOT transposed.
                        // The outputs of matrix matrix multiplications can therefore NOT simply be concatenated for >1 GPU.
                        // Instead they need to be copied to the correct slice in ne0 = dst row index.
                        // If dst is a vector with ne0 == 1 then you don't have to do this but it still produces correct results.
                        float * dhf_dst_i = (float *) ((char *) dst_off_device + i02*nb2 + i03*nb3);
                        GGML_ASSERT(dst->nb[1] == ne0*sizeof(float));
                        dhf_dst_i += src1_col_0*ne0 + dev[id].row_low;
                        CUDA_CHECK(ggml_cuda_Memcpy2DPeerAsync(
                            dhf_dst_i, ctx.device, ne0*sizeof(float), dst_dd_i, id, row_diff*sizeof(float), row_diff*sizeof(float), src1_ncols, stream));
                    } else {
                        float * dhf_dst_i = (float *) ((char *) dst_off_device + i02*nb2 + i03*nb3);
                        GGML_ASSERT(dst->nb[1] == ne0*sizeof(float));
                        dhf_dst_i += src1_col_0*ne0;
                        CUDA_CHECK(cudaMemcpyAsync(dhf_dst_i, dst_dd_i, src1_ncols*ne0*sizeof(float), cudaMemcpyDeviceToDevice, stream));
                    }
                }

                // add event for the main device to wait on until other device is done
                if (split && (id != ctx.device || is != 0)) {
                    CUDA_CHECK(cudaEventRecord(src0_extra->events[id][is], stream));
                }
            }
        }
    }

    // main device waits for all other devices to be finished
    if (split && ggml_backend_cuda_get_device_count() > 1) {
        int64_t is_max = (ne11 + MUL_MAT_SRC1_COL_STRIDE - 1) / MUL_MAT_SRC1_COL_STRIDE;
        is_max = is_max <= GGML_CUDA_MAX_STREAMS ? is_max : GGML_CUDA_MAX_STREAMS;

        ggml_cuda_set_device(ctx.device);
        for (int id = 0; id < ggml_backend_cuda_get_device_count(); ++id) {
            if (dev[id].row_low == dev[id].row_high) {
                continue;
            }
            for (int64_t is = 0; is < is_max; ++is) {
                CUDA_CHECK(cudaStreamWaitEvent(ctx.stream(), src0_extra->events[id][is], 0));
            }
        }
    }
}

static __global__ void k_compute_batched_ptrs(
        const half * src0_as_f16, const half * src1_as_f16, char * dst,
        const void ** ptrs_src, void ** ptrs_dst,
        int64_t ne12, int64_t ne13,
        int64_t ne23,
        size_t  nb02, size_t  nb03,
        size_t  nb12, size_t  nb13,
        size_t  nbd2, size_t  nbd3,
        int64_t r2,   int64_t r3) {
    int64_t i13 = blockIdx.x * blockDim.x + threadIdx.x;
    int64_t i12 = blockIdx.y * blockDim.y + threadIdx.y;

    if (i13 >= ne13 || i12 >= ne12) {
        return;
    }

    int64_t i03 = i13 / r3;
    int64_t i02 = i12 / r2;

    ptrs_src[0*ne23 + i12 + i13*ne12] = (const char *) src0_as_f16 + i02*nb02 + i03*nb03;
    ptrs_src[1*ne23 + i12 + i13*ne12] = (const char *) src1_as_f16 + i12*nb12 + i13*nb13;
    ptrs_dst[0*ne23 + i12 + i13*ne12] = (      char *)         dst + i12*nbd2 + i13*nbd3;
}

static void ggml_cuda_mul_mat_batched_cublas(ggml_backend_cuda_context & ctx, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
    GGML_ASSERT(!ggml_is_transposed(src0));
    GGML_ASSERT(!ggml_is_transposed(src1));

    GGML_ASSERT(ggml_backend_buffer_is_cuda(src0->buffer));
    GGML_ASSERT(src0->type == GGML_TYPE_F16);

    GGML_TENSOR_BINARY_OP_LOCALS

    const int64_t ne_dst = ggml_nelements(dst);

    cudaStream_t main_stream = ctx.stream();

    CUBLAS_CHECK(cublasSetStream(ctx.cublas_handle(), main_stream));

    void * src0_ddq = src0->data;
    half * src0_f16 = (half *) src0_ddq;
    float * src1_ddf = (float *) src1->data;
    float * dst_ddf  = (float *) dst->data;

    // convert src1 to fp16
    ggml_cuda_pool_alloc<half> src1_f16_alloc(ctx.pool());
    if (src1->type != GGML_TYPE_F16) {
        const to_fp16_cuda_t to_fp16_cuda = ggml_get_to_fp16_cuda(src1->type);
        const int64_t ne_src1 = ggml_nelements(src1);
        src1_f16_alloc.alloc(ne_src1);
        GGML_ASSERT(to_fp16_cuda != nullptr);
        to_fp16_cuda(src1_ddf, src1_f16_alloc.get(), ne_src1, main_stream);
    }
    half * src1_f16 = src1->type == GGML_TYPE_F16 ? (half *) src1_ddf : src1_f16_alloc.get();

    ggml_cuda_pool_alloc<half> dst_f16(ctx.pool());
    char * dst_t;

    cublasComputeType_t cu_compute_type = CUBLAS_COMPUTE_16F;
    cudaDataType_t      cu_data_type    = CUDA_R_16F;

    // dst strides
    size_t nbd2 = dst->nb[2];
    size_t nbd3 = dst->nb[3];

    const half  alpha_f16 = 1.0f;
    const half  beta_f16  = 0.0f;

    const float alpha_f32 = 1.0f;
    const float beta_f32  = 0.0f;

    const void * alpha = &alpha_f16;
    const void * beta  = &beta_f16;

    if (dst->op_params[0] == GGML_PREC_DEFAULT) {
        dst_t = (char *) dst_f16.alloc(ne_dst);

        nbd2 /= sizeof(float) / sizeof(half);
        nbd3 /= sizeof(float) / sizeof(half);
    } else {
        dst_t = (char *) dst_ddf;

        cu_compute_type = CUBLAS_COMPUTE_32F;
        cu_data_type    = CUDA_R_32F;

        alpha = &alpha_f32;
        beta  = &beta_f32;
    }

    if (GGML_CUDA_CC_IS_CDNA(ggml_cuda_info().devices[ctx.device].cc)) {
        cu_compute_type = CUBLAS_COMPUTE_32F;
        alpha = &alpha_f32;
        beta  = &beta_f32;
    }

    GGML_ASSERT(ne12 % ne02 == 0);
    GGML_ASSERT(ne13 % ne03 == 0);

    // broadcast factors
    const int64_t r2 = ne12/ne02;
    const int64_t r3 = ne13/ne03;

#if 0
    // use cublasGemmEx
    {
        for (int i13 = 0; i13 < ne13; ++i13) {
            for (int i12 = 0; i12 < ne12; ++i12) {
                int i03 = i13 / r3;
                int i02 = i12 / r2;

                CUBLAS_CHECK(
                        cublasGemmEx(g_cublas_handles[g_main_device], CUBLAS_OP_T, CUBLAS_OP_N,
                            ne01, ne11, ne10,
                            alpha, (const char *) src0_as_f16 + i02*src0->nb[2]   + i03*src0->nb[3]  , CUDA_R_16F,   nb01/sizeof(half),
                                   (const char *) src1_as_f16 + i12*src1->nb[2]/2 + i13*src1->nb[3]/2, CUDA_R_16F,   nb11/sizeof(float),
                            beta,  (      char *)       dst_t + i12*nbd2          + i13*nbd3,          cu_data_type, ne01,
                            cu_compute_type,
                            CUBLAS_GEMM_DEFAULT_TENSOR_OP));
            }
        }
    }
#else
#ifdef GGML_USE_MUSA
    GGML_ASSERT(false);
#else // !GGML_USE_MUSA
    if (r2 == 1 && r3 == 1 && ggml_is_contiguous_2(src0) && ggml_is_contiguous_2(src1)) {
        // there is no broadcast and src0, src1 are contiguous across dims 2, 3
        // use cublasGemmStridedBatchedEx
        CUBLAS_CHECK(
        cublasGemmStridedBatchedEx(ctx.cublas_handle(), CUBLAS_OP_T, CUBLAS_OP_N,
                ne01, ne11, ne10,
                alpha, (const char *) src0_f16, CUDA_R_16F,   nb01/nb00, nb02/nb00,  // strideA
                       (const char *) src1_f16, CUDA_R_16F,   nb11/nb10, nb12/nb10,  // strideB
                beta,  (      char *)    dst_t, cu_data_type, ne01,       nb2/nb0,   // strideC
                ne12*ne13,
                cu_compute_type,
                CUBLAS_GEMM_DEFAULT_TENSOR_OP));
    } else {
        // use cublasGemmBatchedEx
        const int ne23 = ne12*ne13;

        ggml_cuda_pool_alloc<const void *> ptrs_src(ctx.pool(), 2*ne23);
        ggml_cuda_pool_alloc<      void *> ptrs_dst(ctx.pool(), 1*ne23);

        dim3 block_dims(ne13, ne12);
        k_compute_batched_ptrs<<<1, block_dims, 0, main_stream>>>(
                src0_f16, src1_f16, dst_t,
                ptrs_src.get(), ptrs_dst.get(),
                ne12, ne13,
                ne23,
                nb02, nb03,
                src1->type == GGML_TYPE_F16 ? nb12 : nb12/2,
                src1->type == GGML_TYPE_F16 ? nb13 : nb13/2,
                nbd2, nbd3,
                r2, r3);
        CUDA_CHECK(cudaGetLastError());

        CUBLAS_CHECK(
        cublasGemmBatchedEx(ctx.cublas_handle(), CUBLAS_OP_T, CUBLAS_OP_N,
                ne01, ne11, ne10,
                alpha, (const void **) (ptrs_src.get() + 0*ne23), CUDA_R_16F,   nb01/nb00,
                       (const void **) (ptrs_src.get() + 1*ne23), CUDA_R_16F,   nb11/nb10,
                beta,  (      void **) (ptrs_dst.get() + 0*ne23), cu_data_type, ne01,
                ne23,
                cu_compute_type,
                CUBLAS_GEMM_DEFAULT_TENSOR_OP));
    }
#endif // GGML_USE_MUSA
#endif

    if (dst->op_params[0] == GGML_PREC_DEFAULT) {
        const to_fp32_cuda_t to_fp32_cuda = ggml_get_to_fp32_cuda(GGML_TYPE_F16);
        to_fp32_cuda(dst_f16.get(), dst_ddf, ne_dst, main_stream);
    }
}

static void ggml_cuda_mul_mat(ggml_backend_cuda_context & ctx, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
    const bool split = ggml_backend_buft_is_cuda_split(src0->buffer->buft);

    bool use_mul_mat_vec   = (src0->type == GGML_TYPE_F16 || src0->type == GGML_TYPE_BF16)
        && src1->type == GGML_TYPE_F32 && dst->type == GGML_TYPE_F32
        && src0->ne[0] % 2 == 0 && src1->ne[1] == 1;
    bool use_mul_mat_vec_q = ggml_is_quantized(src0->type)
        && src1->type == GGML_TYPE_F32 && dst->type == GGML_TYPE_F32
        && src1->ne[1] <= MMVQ_MAX_BATCH_SIZE;
    bool use_mul_mat_q     = ggml_is_quantized(src0->type)
        && src1->type == GGML_TYPE_F32 && dst->type == GGML_TYPE_F32;

    bool any_gpus_with_slow_fp16   = false;
    bool any_gpus_without_fp16_mma = false;

    if (split) {
        ggml_backend_cuda_split_buffer_type_context * buft_ctx = (ggml_backend_cuda_split_buffer_type_context *) src0->buffer->buft->context;
        auto & tensor_split = buft_ctx->tensor_split;
        for (int id = 0; id < ggml_backend_cuda_get_device_count(); ++id) {
            // skip devices that are not going to do any work:
            if (tensor_split[id] >= (id + 1 < ggml_backend_cuda_get_device_count() ? tensor_split[id + 1] : 1.0f)) {
                continue;
            }

            const int cc              = ggml_cuda_info().devices[id].cc;
            use_mul_mat_q             = use_mul_mat_q             && ggml_cuda_should_use_mmq(src0->type, cc, src1->ne[1]);
            any_gpus_with_slow_fp16   = any_gpus_with_slow_fp16   || !fast_fp16_hardware_available(cc);
            any_gpus_without_fp16_mma = any_gpus_without_fp16_mma || !fp16_mma_hardware_available(cc);
        }
    } else {
        const int cc              = ggml_cuda_info().devices[ctx.device].cc;
        use_mul_mat_q             = use_mul_mat_q             && ggml_cuda_should_use_mmq(src0->type, cc, src1->ne[1]);
        any_gpus_with_slow_fp16   = any_gpus_with_slow_fp16   || !fast_fp16_hardware_available(cc);
        any_gpus_without_fp16_mma = any_gpus_without_fp16_mma || !fp16_mma_hardware_available(cc);
    }

    // debug helpers
    //printf("src0: %8d %8d %8d %8d\n", src0->ne[0], src0->ne[1], src0->ne[2], src0->ne[3]);
    //printf("      %8d %8d %8d %8d\n", src0->nb[0], src0->nb[1], src0->nb[2], src0->nb[3]);
    //printf("src1: %8d %8d %8d %8d\n", src1->ne[0], src1->ne[1], src1->ne[2], src1->ne[3]);
    //printf("      %8d %8d %8d %8d\n", src1->nb[0], src1->nb[1], src1->nb[2], src1->nb[3]);
    //printf("src0 is contiguous %d, transposed %d, type = %s, name = %s\n", ggml_is_contiguous(src0), ggml_is_transposed(src0), ggml_type_name(src0->type), src0->name);
    //printf("src1 is contiguous %d, transposed %d, type = %s, name = %s\n", ggml_is_contiguous(src1), ggml_is_transposed(src1), ggml_type_name(src1->type), src1->name);

    if (!split && use_mul_mat_vec && (src0->ne[1] < MMV_MAX_ROWS || any_gpus_without_fp16_mma)) {
        // the custom F16 vector kernel can be used over batched cuBLAS GEMM
        // but this is only faster for GPUs without tensor cores or with a thin src0 matrix (particularly KQV in attention)
        ggml_cuda_mul_mat_vec(ctx, src0, src1, dst);
    } else if (!split && src0->type == GGML_TYPE_F16 && (src1->type == GGML_TYPE_F16 || !any_gpus_with_slow_fp16)
               && !ggml_is_transposed(src0) && !ggml_is_transposed(src1) && src1->ne[2]*src1->ne[3] > 1) {
        // general KQ + KQV multi-batch without FlashAttention
        ggml_cuda_mul_mat_batched_cublas(ctx, src0, src1, dst);
    } else if (use_mul_mat_vec) {
        ggml_cuda_op_mul_mat(ctx, src0, src1, dst, ggml_cuda_op_mul_mat_vec, nullptr);
    } else if (use_mul_mat_vec_q) {
        ggml_cuda_op_mul_mat(ctx, src0, src1, dst, ggml_cuda_op_mul_mat_vec_q, quantize_row_q8_1_cuda);
    } else if (use_mul_mat_q) {
        ggml_cuda_op_mul_mat(ctx, src0, src1, dst, ggml_cuda_op_mul_mat_q, quantize_mmq_q8_1_cuda);
    } else {
        ggml_cuda_op_mul_mat(ctx, src0, src1, dst, ggml_cuda_op_mul_mat_cublas, nullptr);
    }
}

struct mmid_row_mapping {
    int32_t i1;
    int32_t i2;
};

static __global__ void k_copy_src1_to_contiguous(const char * __restrict__ src1_original, char * __restrict__ src1_contiguous,
                                                 int * __restrict__ cur_src1_row, mmid_row_mapping * __restrict__ row_mapping,
                                                 const char * __restrict ids, int64_t i02, size_t ids_nb1, size_t ids_nb0,
                                                 int64_t ne11, int64_t ne10,
                                                 size_t nb11, size_t nb12) {
    int32_t iid1 = blockIdx.x;
    int32_t id = blockIdx.y;

    const int32_t row_id_i = *(const int32_t *) (ids + iid1*ids_nb1 + id*ids_nb0);

    if (row_id_i != i02) {
        return;
    }

    const int64_t i11 = id % ne11;
    const int64_t i12 = iid1;

    __shared__ int src1_row;
    if (threadIdx.x == 0) {
        src1_row = atomicAdd(cur_src1_row, 1);
        row_mapping[src1_row] = {id, iid1};
    }
    __syncthreads();

    const float * src1_row_original = (const float *)(src1_original + i11*nb11 + i12*nb12);
    float * src1_row_contiguous = (float *)(src1_contiguous + src1_row*nb11);

    for (int i = threadIdx.x; i < ne10; i += blockDim.x) {
        src1_row_contiguous[i] = src1_row_original[i];
    }
}

static __global__ void k_copy_dst_from_contiguous(char * __restrict__ dst_original, const char * __restrict__ dst_contiguous,
                                                  const mmid_row_mapping * __restrict__ row_mapping,
                                                  int64_t ne0,
                                                  size_t nb1, size_t nb2) {
    int32_t i = blockIdx.x;

    const int32_t i1 = row_mapping[i].i1;
    const int32_t i2 = row_mapping[i].i2;

    const float * dst_row_contiguous = (const float *)(dst_contiguous + i*nb1);
    float * dst_row_original = (float *)(dst_original + i1*nb1 + i2*nb2);

    for (int j = threadIdx.x; j < ne0; j += blockDim.x) {
        dst_row_original[j] = dst_row_contiguous[j];
    }
}

static void ggml_cuda_mul_mat_id(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
    const ggml_tensor * src0 = dst->src[0];
    const ggml_tensor * src1 = dst->src[1];
    const ggml_tensor * ids  = dst->src[2];

    GGML_TENSOR_BINARY_OP_LOCALS

    GGML_ASSERT(!ggml_backend_buft_is_cuda_split(src0->buffer->buft) && "mul_mat_id does not support split buffers");

    cudaStream_t stream = ctx.stream();

    const int64_t n_as = ne02;
    const int64_t n_ids = ids->ne[0];

    std::vector<char> ids_host(ggml_nbytes(ids));
    const char * ids_dev = (const char *) ids->data;
    CUDA_CHECK(cudaMemcpyAsync(ids_host.data(), ids_dev, ggml_nbytes(ids), cudaMemcpyDeviceToHost, stream));
    CUDA_CHECK(cudaStreamSynchronize(stream));

    ggml_tensor src0_row = *src0;
    ggml_tensor src1_row = *src1;
    ggml_tensor dst_row  = *dst;

    char * src0_original = (char *) src0->data;
    char * src1_original = (char *) src1->data;
    char * dst_original  = (char *)  dst->data;

    src0_row.ne[2] = 1;
    src0_row.ne[3] = 1;
    src0_row.nb[3] = nb02;

    src1_row.ne[1] = 1;
    src1_row.ne[2] = 1;
    src1_row.ne[3] = 1;
    src1_row.nb[2] = nb11;
    src1_row.nb[3] = nb11;

    dst_row.ne[1] = 1;
    dst_row.ne[2] = 1;
    dst_row.ne[3] = 1;
    dst_row.nb[2] = nb1;
    dst_row.nb[3] = nb1;

    if (ne12 == 1) {
        for (int64_t iid1 = 0; iid1 < ids->ne[1]; iid1++) {
            for (int64_t id = 0; id < n_ids; id++) {
                const int32_t i02 = *(const int32_t *) (ids_host.data() + iid1*ids->nb[1] + id*ids->nb[0]);

                GGML_ASSERT(i02 >= 0 && i02 < n_as);

                const int64_t i11 = id % ne11;
                const int64_t i12 = iid1;

                const int64_t i1 = id;
                const int64_t i2 = i12;

                src0_row.data = src0_original + i02*nb02;
                src1_row.data = src1_original + i11*nb11 + i12*nb12;
                dst_row.data  =  dst_original + i1*nb1   + i2*nb2;

                ggml_cuda_mul_mat(ctx, &src0_row, &src1_row, &dst_row);
            }
        }
    } else {
        ggml_cuda_pool_alloc<char> src1_contiguous(ctx.pool(), sizeof(float)*ggml_nelements(src1));
        ggml_cuda_pool_alloc<char>  dst_contiguous(ctx.pool(), sizeof(float)*ggml_nelements(dst));

        src1_row.data = src1_contiguous.get();
        dst_row.data  =  dst_contiguous.get();

        for (int64_t i02 = 0; i02 < n_as; i02++) {
            int64_t num_src1_rows = 0;

            for (int64_t iid1 = 0; iid1 < ids->ne[1]; iid1++) {
                for (int64_t id = 0; id < n_ids; id++) {
                    const int32_t row_id_i = *(const int32_t *) (ids_host.data() + iid1*ids->nb[1] + id*ids->nb[0]);

                    GGML_ASSERT(row_id_i >= 0 && row_id_i < n_as);

                    if (row_id_i != i02) {
                        continue;
                    }

                    num_src1_rows++;
                }
            }

            if (num_src1_rows == 0) {
                continue;
            }

            ggml_cuda_pool_alloc<int> dev_cur_src1_row(ctx.pool(), 1);
            ggml_cuda_pool_alloc<mmid_row_mapping> dev_row_mapping(ctx.pool(), num_src1_rows);
            CUDA_CHECK(cudaMemsetAsync(dev_cur_src1_row.get(), 0, sizeof(int), stream));

            {
                dim3 block_dims(std::min((unsigned int)ne10, 768u));
                dim3 grid_dims(ids->ne[1], n_ids);
                k_copy_src1_to_contiguous<<<grid_dims, block_dims, 0, stream>>>(
                        src1_original, src1_contiguous.get(),
                        dev_cur_src1_row.get(), dev_row_mapping.get(),
                        ids_dev, i02, ids->nb[1], ids->nb[0],
                        ne11, ne10,
                        nb11, nb12);
                CUDA_CHECK(cudaGetLastError());
            }

            src0_row.data = src0_original + i02*nb02;

            GGML_ASSERT(nb11 == sizeof(float)*ne10);
            GGML_ASSERT(nb1 == sizeof(float)*ne0);

            src1_row.ne[1] = num_src1_rows;
            src1_row.nb[1] = nb11;
            src1_row.nb[2] = num_src1_rows*nb11;
            src1_row.nb[3] = num_src1_rows*nb11;

            dst_row.ne[1] = num_src1_rows;
            dst_row.nb[1] = nb1;
            dst_row.nb[2] = num_src1_rows*nb1;
            dst_row.nb[3] = num_src1_rows*nb1;

            ggml_cuda_mul_mat(ctx, &src0_row, &src1_row, &dst_row);

            {
                dim3 block_dims(std::min((unsigned int)ne0, 768u));
                dim3 grid_dims(num_src1_rows);
                k_copy_dst_from_contiguous<<<grid_dims, block_dims, 0, stream>>>(
                        dst_original, dst_contiguous.get(),
                        dev_row_mapping.get(),
                        ne0,
                        nb1, nb2);
                CUDA_CHECK(cudaGetLastError());
            }
        }
    }
}

static bool ggml_cuda_compute_forward(ggml_backend_cuda_context & ctx, struct ggml_tensor * dst) {
    // why is this here instead of mul_mat?
    if (dst->src[0] != nullptr && ggml_backend_buft_is_cuda_split(dst->src[0]->buffer->buft)) {
        ggml_cuda_set_peer_access(dst->src[1]->ne[1], ctx.device);
    }

    switch (dst->op) {
        case GGML_OP_ARGMAX:
            ggml_cuda_argmax(ctx, dst);
            break;
        case GGML_OP_COUNT_EQUAL:
            ggml_cuda_count_equal(ctx, dst);
            break;
        case GGML_OP_REPEAT:
            ggml_cuda_op_repeat(ctx, dst);
            break;
        case GGML_OP_REPEAT_BACK:
            ggml_cuda_op_repeat_back(ctx, dst);
            break;
        case GGML_OP_GET_ROWS:
            ggml_cuda_op_get_rows(ctx, dst);
            break;
        case GGML_OP_GET_ROWS_BACK:
            ggml_cuda_op_get_rows_back(ctx, dst);
            break;
        case GGML_OP_DUP:
            ggml_cuda_dup(ctx, dst);
            break;
        case GGML_OP_CPY:
            ggml_cuda_cpy(ctx, dst->src[0], dst->src[1]);
            break;
        case GGML_OP_CONT:
            ggml_cuda_dup(ctx, dst);
            break;
        case GGML_OP_ADD:
        case GGML_OP_ADD1: // TODO: more efficient implementation
            ggml_cuda_op_add(ctx, dst);
            break;
        case GGML_OP_SUB:
            ggml_cuda_op_sub(ctx, dst);
            break;
        case GGML_OP_ACC:
            ggml_cuda_op_acc(ctx, dst);
            break;
        case GGML_OP_MUL:
            ggml_cuda_op_mul(ctx, dst);
            break;
        case GGML_OP_DIV:
            ggml_cuda_op_div(ctx, dst);
            break;
        case GGML_OP_UNARY:
            switch (ggml_get_unary_op(dst)) {
                case GGML_UNARY_OP_NEG:
                    ggml_cuda_op_neg(ctx, dst);
                    break;
                case GGML_UNARY_OP_STEP:
                    ggml_cuda_op_step(ctx, dst);
                    break;
                case GGML_UNARY_OP_GELU:
                    ggml_cuda_op_gelu(ctx, dst);
                    break;
                case GGML_UNARY_OP_SILU:
                    ggml_cuda_op_silu(ctx, dst);
                    break;
                case GGML_UNARY_OP_GELU_QUICK:
                    ggml_cuda_op_gelu_quick(ctx, dst);
                    break;
                case GGML_UNARY_OP_TANH:
                    ggml_cuda_op_tanh(ctx, dst);
                    break;
                case GGML_UNARY_OP_RELU:
                    ggml_cuda_op_relu(ctx, dst);
                    break;
                case GGML_UNARY_OP_SIGMOID:
                    ggml_cuda_op_sigmoid(ctx, dst);
                    break;
                case GGML_UNARY_OP_HARDSIGMOID:
                    ggml_cuda_op_hardsigmoid(ctx, dst);
                    break;
                case GGML_UNARY_OP_HARDSWISH:
                    ggml_cuda_op_hardswish(ctx, dst);
                    break;
                case GGML_UNARY_OP_EXP:
                    ggml_cuda_op_exp(ctx, dst);
                    break;
                default:
                    return false;
            }
            break;
        case GGML_OP_NORM:
            ggml_cuda_op_norm(ctx, dst);
            break;
        case GGML_OP_GROUP_NORM:
            ggml_cuda_op_group_norm(ctx, dst);
            break;
        case GGML_OP_CONCAT:
            ggml_cuda_op_concat(ctx, dst);
            break;
        case GGML_OP_UPSCALE:
            ggml_cuda_op_upscale(ctx, dst);
            break;
        case GGML_OP_PAD:
            ggml_cuda_op_pad(ctx, dst);
            break;
        case GGML_OP_ARANGE:
            ggml_cuda_op_arange(ctx, dst);
            break;
        case GGML_OP_TIMESTEP_EMBEDDING:
            ggml_cuda_op_timestep_embedding(ctx, dst);
            break;
        case GGML_OP_LEAKY_RELU:
            ggml_cuda_op_leaky_relu(ctx, dst);
            break;
        case GGML_OP_SILU_BACK:
            ggml_cuda_op_silu_back(ctx, dst);
            break;
        case GGML_OP_RMS_NORM:
            ggml_cuda_op_rms_norm(ctx, dst);
            break;
        case GGML_OP_RMS_NORM_BACK:
            ggml_cuda_op_rms_norm_back(ctx, dst);
            break;
        case GGML_OP_MUL_MAT:
            ggml_cuda_mul_mat(ctx, dst->src[0], dst->src[1], dst);
            break;
        case GGML_OP_MUL_MAT_ID:
            ggml_cuda_mul_mat_id(ctx, dst);
            break;
        case GGML_OP_OUT_PROD:
            ggml_cuda_out_prod(ctx, dst);
            break;
        case GGML_OP_SCALE:
            ggml_cuda_op_scale(ctx, dst);
            break;
        case GGML_OP_SQR:
            ggml_cuda_op_sqr(ctx, dst);
            break;
        case GGML_OP_SQRT:
            ggml_cuda_op_sqrt(ctx, dst);
            break;
        case GGML_OP_SIN:
            ggml_cuda_op_sin(ctx, dst);
            break;
        case GGML_OP_COS:
            ggml_cuda_op_cos(ctx, dst);
            break;
        case GGML_OP_CLAMP:
            ggml_cuda_op_clamp(ctx, dst);
            break;
        case GGML_OP_NONE:
        case GGML_OP_RESHAPE:
        case GGML_OP_VIEW:
        case GGML_OP_PERMUTE:
        case GGML_OP_TRANSPOSE:
                break;
        case GGML_OP_DIAG_MASK_INF:
            ggml_cuda_op_diag_mask_inf(ctx, dst);
            break;
        case GGML_OP_SOFT_MAX:
            ggml_cuda_op_soft_max(ctx, dst);
            break;
        case GGML_OP_SOFT_MAX_BACK:
            ggml_cuda_op_soft_max_back(ctx, dst);
            break;
        case GGML_OP_ROPE:
            ggml_cuda_op_rope(ctx, dst);
            break;
        case GGML_OP_ROPE_BACK:
            ggml_cuda_op_rope_back(ctx, dst);
            break;
        case GGML_OP_IM2COL:
            ggml_cuda_op_im2col(ctx, dst);
            break;
        case GGML_OP_CONV_TRANSPOSE_1D:
            ggml_cuda_op_conv_transpose_1d(ctx,dst);
            break;
        case GGML_OP_POOL_2D:
            ggml_cuda_op_pool2d(ctx, dst);
            break;
        case GGML_OP_SUM:
            ggml_cuda_op_sum(ctx, dst);
            break;
        case GGML_OP_SUM_ROWS:
            ggml_cuda_op_sum_rows(ctx, dst);
            break;
        case GGML_OP_ARGSORT:
            ggml_cuda_op_argsort(ctx, dst);
            break;
        case GGML_OP_FLASH_ATTN_EXT:
            ggml_cuda_flash_attn_ext(ctx, dst);
            break;
        case GGML_OP_CROSS_ENTROPY_LOSS:
            ggml_cuda_cross_entropy_loss(ctx, dst);
            break;
        case GGML_OP_RWKV_WKV6:
            ggml_cuda_op_rwkv_wkv6(ctx, dst);
            break;
        case GGML_OP_GATED_LINEAR_ATTN:
            ggml_cuda_op_gated_linear_attn(ctx, dst);
            break;
        case GGML_OP_CROSS_ENTROPY_LOSS_BACK:
            ggml_cuda_cross_entropy_loss_back(ctx, dst);
            break;
        case GGML_OP_OPT_STEP_ADAMW:
            ggml_cuda_opt_step_adamw(ctx, dst);
            break;
        default:
            return false;
    }

    cudaError_t err = cudaGetLastError();
    if (err != cudaSuccess) {
        GGML_LOG_ERROR("%s: %s failed\n", __func__, ggml_op_desc(dst));
        CUDA_CHECK(err);
    }

    return true;
}

////////////////////////////////////////////////////////////////////////////////

// backend

static const char * ggml_backend_cuda_get_name(ggml_backend_t backend) {
    ggml_backend_cuda_context * cuda_ctx = (ggml_backend_cuda_context *)backend->context;

    return cuda_ctx->name.c_str();
}

static void ggml_backend_cuda_free(ggml_backend_t backend) {
    ggml_backend_cuda_context * cuda_ctx = (ggml_backend_cuda_context *)backend->context;

    delete cuda_ctx;
    delete backend;
}

static void ggml_backend_cuda_set_tensor_async(ggml_backend_t backend, ggml_tensor * tensor, const void * data, size_t offset, size_t size) {
    ggml_backend_cuda_context * cuda_ctx = (ggml_backend_cuda_context *)backend->context;
    ggml_backend_buffer_t buf = tensor->view_src ? tensor->view_src->buffer : tensor->buffer;

    GGML_ASSERT(buf->buft == ggml_backend_cuda_buffer_type(cuda_ctx->device) && "unsupported buffer type");

    CUDA_CHECK(cudaMemcpyAsync((char *)tensor->data + offset, data, size, cudaMemcpyHostToDevice, cuda_ctx->stream()));
}

static void ggml_backend_cuda_get_tensor_async(ggml_backend_t backend, const ggml_tensor * tensor, void * data, size_t offset, size_t size) {
    ggml_backend_cuda_context * cuda_ctx = (ggml_backend_cuda_context *)backend->context;
    ggml_backend_buffer_t buf = tensor->view_src ? tensor->view_src->buffer : tensor->buffer;

    GGML_ASSERT(buf->buft == ggml_backend_cuda_buffer_type(cuda_ctx->device) && "unsupported buffer type");

    CUDA_CHECK(cudaMemcpyAsync(data, (const char *)tensor->data + offset, size, cudaMemcpyDeviceToHost, cuda_ctx->stream()));
}

static bool ggml_backend_cuda_cpy_tensor_async(ggml_backend_t backend_src, ggml_backend_t backend_dst, const ggml_tensor * src, ggml_tensor * dst) {
    ggml_backend_buffer_t buf_src = src->view_src ? src->view_src->buffer : src->buffer;
    ggml_backend_buffer_t buf_dst = dst->view_src ? dst->view_src->buffer : dst->buffer;

    if (!ggml_backend_is_cuda(backend_src) || !ggml_backend_is_cuda(backend_dst)) {
        return false;
    }

    if (!ggml_backend_buffer_is_cuda(src->buffer) || !ggml_backend_buffer_is_cuda(dst->buffer)) {
        return false;
    }

    // device -> device copy
    ggml_backend_cuda_context * cuda_ctx_src = (ggml_backend_cuda_context *)backend_src->context;
    ggml_backend_cuda_context * cuda_ctx_dst = (ggml_backend_cuda_context *)backend_dst->context;

    ggml_backend_cuda_buffer_context * buf_ctx_src = (ggml_backend_cuda_buffer_context *)buf_src->context;
    ggml_backend_cuda_buffer_context * buf_ctx_dst = (ggml_backend_cuda_buffer_context *)buf_dst->context;

    if (cuda_ctx_src->device != buf_ctx_src->device || cuda_ctx_dst->device != buf_ctx_dst->device) {
#ifndef NDEBUG
        GGML_LOG_DEBUG("%s: backend and buffer devices do not match\n", __func__);
#endif
        return false;
    }

    if (backend_src != backend_dst) {
        // copy on src stream
        if (cuda_ctx_src->device == cuda_ctx_dst->device) {
            CUDA_CHECK(cudaMemcpyAsync(dst->data, src->data, ggml_nbytes(dst), cudaMemcpyDeviceToDevice, cuda_ctx_src->stream()));
        } else {
#ifdef GGML_CUDA_NO_PEER_COPY
            return false;
#else
            CUDA_CHECK(cudaMemcpyPeerAsync(dst->data, cuda_ctx_dst->device, src->data, cuda_ctx_src->device, ggml_nbytes(dst), cuda_ctx_src->stream()));
#endif
        }

        // record event on src stream after the copy
        if (!cuda_ctx_src->copy_event) {
            ggml_cuda_set_device(cuda_ctx_src->device);
            CUDA_CHECK(cudaEventCreateWithFlags(&cuda_ctx_src->copy_event, cudaEventDisableTiming));
        }

        CUDA_CHECK(cudaEventRecord(cuda_ctx_src->copy_event, cuda_ctx_src->stream()));

        // wait on dst stream for the copy to complete
        CUDA_CHECK(cudaStreamWaitEvent(cuda_ctx_dst->stream(), cuda_ctx_src->copy_event, 0));
    } else {
        // src and dst are on the same backend
        CUDA_CHECK(cudaMemcpyAsync(dst->data, src->data, ggml_nbytes(dst), cudaMemcpyDeviceToDevice, cuda_ctx_src->stream()));
    }
    return true;
}

static void ggml_backend_cuda_synchronize(ggml_backend_t backend) {
    ggml_backend_cuda_context * cuda_ctx = (ggml_backend_cuda_context *)backend->context;

    CUDA_CHECK(cudaStreamSynchronize(cuda_ctx->stream()));

    GGML_UNUSED(backend);
}

#ifdef USE_CUDA_GRAPH
static bool check_node_graph_compatibility_and_refresh_copy_ops(ggml_backend_cuda_context * cuda_ctx, ggml_cgraph * cgraph,
    std::vector<void *> & ggml_cuda_cpy_fn_ptrs, bool use_cuda_graph) {

    // Loop over nodes in GGML graph to obtain info needed for CUDA graph
    cuda_ctx->cuda_graph->updated_kernel_arg.clear();
    for (int i = 0; i < cgraph->n_nodes; i++) {
        ggml_tensor * node = cgraph->nodes[i];

        if (ggml_is_empty(node) || node->op == GGML_OP_RESHAPE || node->op == GGML_OP_TRANSPOSE || node->op == GGML_OP_VIEW || node->op == GGML_OP_PERMUTE || node->op == GGML_OP_NONE) {
            continue;
        }

        if (node->src[0] && node->src[0]->buffer && ggml_backend_buft_is_cuda_split(node->src[0]->buffer->buft)) {
            use_cuda_graph = false; // Split buffers are not supported by CUDA graph capture
#ifndef NDEBUG
            GGML_LOG_DEBUG("%s: disabling CUDA graphs due to split buffer\n", __func__);
#endif
        }

        if (node->op == GGML_OP_MUL_MAT_ID) {
            use_cuda_graph = false; // This node type is not supported by CUDA graph capture
#ifndef NDEBUG
            GGML_LOG_DEBUG("%s: disabling CUDA graphs due to mul_mat_id\n", __func__);
#endif
        }

        if (node->op == GGML_OP_ADD && node->src[1] && node->src[1]->ne[1] > 1) {
            // disable CUDA graphs for batch size > 1 for now.
            // Changes in batch size or context size can cause changes to the grid size of some kernels.
            use_cuda_graph = false;
#ifndef NDEBUG
            GGML_LOG_DEBUG("%s: disabling CUDA graphs due to batch size > 1 [%s] [%ld %ld %ld %ld]\n", __func__, node->name, node->ne[0], node->ne[1], node->ne[2], node->ne[3]);
#endif
        }

        if (node->op == GGML_OP_CPY) {
            // store the copy op parameter which changes with each token.
            cuda_ctx->cuda_graph->updated_kernel_arg.push_back((char **) &(node->src[1]->data));
            // store a pointer to each copy op CUDA kernel to identify it later
            void * ptr = ggml_cuda_cpy_fn(node->src[0], node->src[1]);
            if (!ptr) {
                use_cuda_graph = false;
#ifndef NDEBUG
                GGML_LOG_DEBUG("%s: disabling CUDA graphs due to unsupported copy op\n", __func__);
#endif
            } else {
                if (std::find(ggml_cuda_cpy_fn_ptrs.begin(), ggml_cuda_cpy_fn_ptrs.end(), ptr) == ggml_cuda_cpy_fn_ptrs.end()) {
                    ggml_cuda_cpy_fn_ptrs.push_back(ptr);
                }
            }
        }

        if (!use_cuda_graph) {
            break;
        }
    }

    return use_cuda_graph;
}

static void set_ggml_graph_node_properties(ggml_tensor * node, ggml_graph_node_properties * graph_node_properties) {
    graph_node_properties->node_address = node->data;
    graph_node_properties->node_op = node->op;
    for (int i = 0; i < GGML_MAX_DIMS; i++) {
        graph_node_properties->ne[i] = node->ne[i];
        graph_node_properties->nb[i] = node->nb[i];
    }
    for (int i = 0; i < GGML_MAX_SRC; i++) {
        graph_node_properties->src_address[i] = node->src[i] ? node->src[i]->data : nullptr;
    }
    memcpy(graph_node_properties->op_params, node->op_params, GGML_MAX_OP_PARAMS);
}

static bool ggml_graph_node_has_matching_properties(ggml_tensor * node, ggml_graph_node_properties * graph_node_properties) {
    if (node->data != graph_node_properties->node_address &&
          node->op != GGML_OP_CPY &&
          node->op != GGML_OP_VIEW) {
        return false;
    }

    if (node->op != graph_node_properties->node_op) {
        return false;
    }

    for (int i = 0; i < GGML_MAX_DIMS; i++) {
        if (node->ne[i] != graph_node_properties->ne[i]) {
            return false;
        }
        if (node->nb[i] != graph_node_properties->nb[i]) {
            return false;
        }
    }

    for (int i = 0; i < GGML_MAX_SRC; i++) {
        if (node->src[i] &&
            node->src[i]->data != graph_node_properties->src_address[i] &&
            node->op != GGML_OP_CPY &&
            node->op != GGML_OP_VIEW
        ) {
            return false;
        }
    }

    if (node->op == GGML_OP_SCALE &&
        memcmp(graph_node_properties->op_params, node->op_params, GGML_MAX_OP_PARAMS) != 0) {
        return false;
    }

    return true;
}

static void maintain_cuda_graph(ggml_backend_cuda_context * cuda_ctx, std::vector<void *> & ggml_cuda_cpy_fn_ptrs, bool cuda_graph_update_required) {

    if (cuda_graph_update_required) {
        // Extract nodes from graph
        // First call with null argument gets number of nodes in graph
        CUDA_CHECK(cudaGraphGetNodes(cuda_ctx->cuda_graph->graph, nullptr, &cuda_ctx->cuda_graph->num_nodes));
        // Subsequent call with non-null argument gets nodes
        cuda_ctx->cuda_graph->nodes.clear();
        cuda_ctx->cuda_graph->nodes.resize(cuda_ctx->cuda_graph->num_nodes);
        cuda_ctx->cuda_graph->params.clear();
        cuda_ctx->cuda_graph->params.resize(cuda_ctx->cuda_graph->num_nodes);
        if (cuda_ctx->cuda_graph->num_nodes > 0) {
            CUDA_CHECK(cudaGraphGetNodes(cuda_ctx->cuda_graph->graph, cuda_ctx->cuda_graph->nodes.data(), &cuda_ctx->cuda_graph->num_nodes));

            // Loop over nodes, and extract kernel parameters from each node
            for (size_t i = 0; i < cuda_ctx->cuda_graph->num_nodes; i++) {
                cudaGraphNodeType node_type;
                CUDA_CHECK(cudaGraphNodeGetType(cuda_ctx->cuda_graph->nodes[i], &node_type));
                if (node_type == cudaGraphNodeTypeKernel) {
                    cudaError_t stat = cudaGraphKernelNodeGetParams(cuda_ctx->cuda_graph->nodes[i], &cuda_ctx->cuda_graph->params[i]); // Get params using runtime
                    if (stat == cudaErrorInvalidDeviceFunction) {
                        // Fails due to incorrect handling by CUDA runtime of CUDA BLAS node.
                        // We don't need to update blas nodes, so clear error and move on.
                        (void)cudaGetLastError();
                    } else {
                        GGML_ASSERT(stat == cudaSuccess);
                    }
                }
            }
        }
    } else {
        // One of the arguments to the copy kernel is updated for each token, hence we need to
        // replace that argument with the updated value in the CUDA graph
        // on update steps, the live parameters will already be captured
        int k = 0;
        for (size_t i = 0; i < cuda_ctx->cuda_graph->num_nodes; i++) {
            if(count(ggml_cuda_cpy_fn_ptrs.begin(), ggml_cuda_cpy_fn_ptrs.end(), cuda_ctx->cuda_graph->params[i].func) > 0) {
                char ** updated_kernel_arg_ptr = cuda_ctx->cuda_graph->updated_kernel_arg.at(k++);
                cuda_ctx->cuda_graph->params[i].kernelParams[1] = updated_kernel_arg_ptr;
                CUDA_CHECK(cudaGraphKernelNodeSetParams(cuda_ctx->cuda_graph->nodes[i], &cuda_ctx->cuda_graph->params[i]));
            }
        }
    }
}

static bool is_cuda_graph_update_required(ggml_backend_cuda_context * cuda_ctx, ggml_cgraph * cgraph) {

    bool cuda_graph_update_required = false;

    if (cuda_ctx->cuda_graph->instance == nullptr) {
        cuda_graph_update_required = true;
    }

    // Check if the graph size has changed
    if (cuda_ctx->cuda_graph->ggml_graph_properties.size() != (size_t)cgraph->n_nodes) {
        cuda_graph_update_required = true;
        cuda_ctx->cuda_graph->ggml_graph_properties.resize(cgraph->n_nodes);
    }

    // Loop over nodes in GGML graph to determine if CUDA graph update is required
    // and store properties to allow this comparison for the next token
    for (int i = 0; i < cgraph->n_nodes; i++) {
        bool has_matching_properties = true;
        if (!cuda_graph_update_required) {
            has_matching_properties = ggml_graph_node_has_matching_properties(cgraph->nodes[i], &cuda_ctx->cuda_graph->ggml_graph_properties[i]);
        }
        if (!has_matching_properties) {
            cuda_graph_update_required = true;
        }
        set_ggml_graph_node_properties(cgraph->nodes[i], &cuda_ctx->cuda_graph->ggml_graph_properties[i]);
    }

    return cuda_graph_update_required;
}

static void update_cuda_graph_executable(ggml_backend_cuda_context * cuda_ctx) {

    cudaGraphExecUpdateResultInfo result_info;
#ifdef __HIP_PLATFORM_AMD__
    hipGraphNode_t errorNode;
    hipError_t stat = hipGraphExecUpdate(cuda_ctx->cuda_graph->instance, cuda_ctx->cuda_graph->graph, &errorNode, &result_info);
#else
    cudaError_t stat = cudaGraphExecUpdate(cuda_ctx->cuda_graph->instance, cuda_ctx->cuda_graph->graph, &result_info);
#endif
    if (stat == cudaErrorGraphExecUpdateFailure) {
#ifndef NDEBUG
        GGML_LOG_DEBUG("%s: CUDA graph update failed\n", __func__);
#endif

        // The pre-existing graph exec cannot be updated due to violated constraints
        // so instead clear error and re-instantiate
        (void)cudaGetLastError();
        CUDA_CHECK(cudaGraphExecDestroy(cuda_ctx->cuda_graph->instance));
        cuda_ctx->cuda_graph->instance = nullptr;
        CUDA_CHECK(cudaGraphInstantiate(&cuda_ctx->cuda_graph->instance, cuda_ctx->cuda_graph->graph, NULL, NULL, 0));
    } else {
        GGML_ASSERT(stat == cudaSuccess);
    }
}
#endif

static void evaluate_and_capture_cuda_graph(ggml_backend_cuda_context * cuda_ctx, ggml_cgraph * cgraph,
   [[maybe_unused]] std::vector<void *> & ggml_cuda_cpy_fn_ptrs,  bool & graph_evaluated_or_captured, bool & use_cuda_graph,
    bool & cuda_graph_update_required) {

    while (!graph_evaluated_or_captured) {
        // 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];

                if (ggml_is_empty(node) || node->op == GGML_OP_RESHAPE || node->op == GGML_OP_TRANSPOSE || node->op == GGML_OP_VIEW || node->op == GGML_OP_PERMUTE || node->op == GGML_OP_NONE) {
                    continue;
                }

#ifndef NDEBUG
                assert(node->buffer->buft == ggml_backend_cuda_buffer_type(cuda_ctx->device));
                for (int j = 0; j < GGML_MAX_SRC; j++) {
                    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));
                    }
                }
#endif

                bool ok = ggml_cuda_compute_forward(*cuda_ctx, node);
                if (!ok) {
                    GGML_LOG_ERROR("%s: op not supported %s (%s)\n", __func__, node->name, ggml_op_name(node->op));
                }
                GGML_ASSERT(ok);
            }
        }

#ifdef USE_CUDA_GRAPH
        if (use_cuda_graph && cuda_graph_update_required) { // End CUDA graph capture
            if (cuda_ctx->cuda_graph->graph != nullptr) {
                CUDA_CHECK(cudaGraphDestroy(cuda_ctx->cuda_graph->graph));
                cuda_ctx->cuda_graph->graph = nullptr;
            }

            CUDA_CHECK(cudaStreamEndCapture(cuda_ctx->stream(), &cuda_ctx->cuda_graph->graph));
            graph_evaluated_or_captured = true; // CUDA graph has been captured
        } else {
            graph_evaluated_or_captured = true; // ggml graph has been directly evaluated
        }
    }

    if (use_cuda_graph) {
        if (cuda_ctx->cuda_graph->instance == nullptr) { // Create executable graph from captured graph.
            CUDA_CHECK(cudaGraphInstantiate(&cuda_ctx->cuda_graph->instance, cuda_ctx->cuda_graph->graph, NULL, NULL, 0));
        }

        // Perform update to graph (if required for this token), and change copy parameter (required for every token)
        maintain_cuda_graph(cuda_ctx, ggml_cuda_cpy_fn_ptrs, cuda_graph_update_required);

        // Update graph executable
        update_cuda_graph_executable(cuda_ctx);

        // Launch graph
        CUDA_CHECK(cudaGraphLaunch(cuda_ctx->cuda_graph->instance, cuda_ctx->stream()));
#else
        graph_evaluated_or_captured = true;
#endif  // USE_CUDA_GRAPH
    }
}

static enum ggml_status ggml_backend_cuda_graph_compute(ggml_backend_t backend, ggml_cgraph * cgraph) {
    ggml_backend_cuda_context * cuda_ctx = (ggml_backend_cuda_context *)backend->context;

    ggml_cuda_set_device(cuda_ctx->device);

    // vector of pointers to CUDA cpy kernels, which are required to identify
    // kernel parameters which need updated in the graph for each token
    std::vector<void *> ggml_cuda_cpy_fn_ptrs;

#ifdef USE_CUDA_GRAPH
    static const bool disable_cuda_graphs_due_to_env = (getenv("GGML_CUDA_DISABLE_GRAPHS") != nullptr);

    // Objects required for CUDA Graph
    if (cuda_ctx->cuda_graph == nullptr) {
        cuda_ctx->cuda_graph.reset(new ggml_cuda_graph());
    }

    bool use_cuda_graph = true;
    bool cuda_graph_update_required = false;

    if (cuda_ctx->cuda_graph->graph == nullptr) {
        if (ggml_cuda_info().devices[cuda_ctx->device].cc < GGML_CUDA_CC_AMPERE) {
            cuda_ctx->cuda_graph->disable_due_to_gpu_arch = true;
#ifndef NDEBUG
            GGML_LOG_DEBUG("%s: disabling CUDA graphs due to GPU architecture\n", __func__);
#endif
        }
    }

    // Disable CUDA graphs in presence of env var, old GPU, use-case which is changing too rapidly,
    // or previous graph capture failure.
    // Also disable for multi-gpu for now. TO DO investigate
    if (disable_cuda_graphs_due_to_env
        || cuda_ctx->cuda_graph->disable_due_to_gpu_arch
        || cuda_ctx->cuda_graph->disable_due_to_too_many_updates
        || cuda_ctx->cuda_graph->disable_due_to_failed_graph_capture) {
        use_cuda_graph = false;
    }

    if (use_cuda_graph) {
        cuda_graph_update_required = is_cuda_graph_update_required(cuda_ctx, cgraph);

        use_cuda_graph = check_node_graph_compatibility_and_refresh_copy_ops(cuda_ctx, cgraph,
                             ggml_cuda_cpy_fn_ptrs, use_cuda_graph);

        // Disable CUDA graphs (from the next token) if the use-case is demanding too many consecutive graph updates.
        if (use_cuda_graph && cuda_graph_update_required) {
            cuda_ctx->cuda_graph->number_consecutive_updates++;
        } else {
            cuda_ctx->cuda_graph->number_consecutive_updates = 0;
        }

        if (cuda_ctx->cuda_graph->number_consecutive_updates >= 4) {
            cuda_ctx->cuda_graph->disable_due_to_too_many_updates = true;
#ifndef NDEBUG
            GGML_LOG_DEBUG("%s: disabling CUDA graphs due to too many consecutive updates\n", __func__);
#endif
        }
    }

    if (use_cuda_graph && cuda_graph_update_required) { // Start CUDA graph capture
        CUDA_CHECK(cudaStreamBeginCapture(cuda_ctx->stream(), cudaStreamCaptureModeRelaxed));
    }

#else
    bool use_cuda_graph = false;
    bool cuda_graph_update_required = false;
#endif // USE_CUDA_GRAPH

    bool graph_evaluated_or_captured = false;

    evaluate_and_capture_cuda_graph(cuda_ctx, cgraph, ggml_cuda_cpy_fn_ptrs, graph_evaluated_or_captured, use_cuda_graph, cuda_graph_update_required);

    return GGML_STATUS_SUCCESS;
}

static void ggml_backend_cuda_event_record(ggml_backend_t backend, ggml_backend_event_t event) {
    ggml_backend_cuda_context * cuda_ctx = (ggml_backend_cuda_context *)backend->context;

    CUDA_CHECK(cudaEventRecord((cudaEvent_t)event->context, cuda_ctx->stream()));
}

static void ggml_backend_cuda_event_wait(ggml_backend_t backend, ggml_backend_event_t event) {
    ggml_backend_cuda_context * cuda_ctx = (ggml_backend_cuda_context *)backend->context;

    if (ggml_backend_is_cuda(backend)) {
        CUDA_CHECK(cudaStreamWaitEvent(cuda_ctx->stream(), (cudaEvent_t)event->context, 0));
    } else {
#if 0
        // untested
        auto wait_fn = [](void * user_data) {
            ggml_backend_event_t event = (ggml_backend_event_t)user_data;
            ggml_backend_event_synchronize(event);
        };

        CUDA_CHECK(cudaLaunchHostFunc(cuda_ctx->stream(), wait_fn, event));
#endif
        GGML_ABORT("fatal error");
    }
}

static const ggml_backend_i ggml_backend_cuda_interface = {
    /* .get_name                = */ ggml_backend_cuda_get_name,
    /* .free                    = */ ggml_backend_cuda_free,
    /* .set_tensor_async        = */ ggml_backend_cuda_set_tensor_async,
    /* .get_tensor_async        = */ ggml_backend_cuda_get_tensor_async,
    /* .cpy_tensor_async        = */ ggml_backend_cuda_cpy_tensor_async,
    /* .synchronize             = */ ggml_backend_cuda_synchronize,
    /* .graph_plan_create       = */ NULL,
    /* .graph_plan_free         = */ NULL,
    /* .graph_plan_update       = */ NULL,
    /* .graph_plan_compute      = */ NULL,
    /* .graph_compute           = */ ggml_backend_cuda_graph_compute,
    /* .event_record            = */ ggml_backend_cuda_event_record,
    /* .event_wait              = */ ggml_backend_cuda_event_wait,
};

static ggml_guid_t ggml_backend_cuda_guid() {
    static ggml_guid guid = { 0x2c, 0xdd, 0xe8, 0x1c, 0x65, 0xb3, 0x65, 0x73, 0x6a, 0x12, 0x88, 0x61, 0x1c, 0xc9, 0xdc, 0x25 };
    return &guid;
}

bool ggml_backend_is_cuda(ggml_backend_t backend) {
    return backend != NULL && ggml_guid_matches(backend->guid, ggml_backend_cuda_guid());
}

int ggml_backend_cuda_get_device_count() {
    return ggml_cuda_info().device_count;
}

void ggml_backend_cuda_get_device_description(int device, char * description, size_t description_size) {
    cudaDeviceProp prop;
    CUDA_CHECK(cudaGetDeviceProperties(&prop, device));
    snprintf(description, description_size, "%s", prop.name);
}

void ggml_backend_cuda_get_device_memory(int device, size_t * free, size_t * total) {
    ggml_cuda_set_device(device);

    CUDA_CHECK(cudaMemGetInfo(free, total));
}

bool ggml_backend_cuda_register_host_buffer(void * buffer, size_t size) {
    if (getenv("GGML_CUDA_REGISTER_HOST") == nullptr) {
        return false;
    }

#if CUDART_VERSION >= 11010 || defined(GGML_USE_MUSA)
    cudaError_t err = cudaHostRegister(buffer, size, cudaHostRegisterPortable | cudaHostRegisterReadOnly);
    if (err != cudaSuccess) {
        // clear the error
        (void)cudaGetLastError();

        GGML_LOG_DEBUG("%s: failed to register %.2f MiB of pinned memory: %s\n", __func__,
                           size / 1024.0 / 1024.0, cudaGetErrorString(err));
        return false;
    }
    return true;
#else
    GGML_UNUSED(buffer);
    GGML_UNUSED(size);
    return false;
#endif // CUDART_VERSION >= 11010 || defined(GGML_USE_MUSA)
}

void ggml_backend_cuda_unregister_host_buffer(void * buffer) {
    if (getenv("GGML_CUDA_REGISTER_HOST") == nullptr) {
        return;
    }

    cudaError_t err = cudaHostUnregister(buffer);
    if (err != cudaSuccess) {
        // clear the error
        (void)cudaGetLastError();
    }
}


// backend device

struct ggml_backend_cuda_device_context {
    int device;
    std::string name;
    std::string description;
};

static const char * ggml_backend_cuda_device_get_name(ggml_backend_dev_t dev) {
    ggml_backend_cuda_device_context * ctx = (ggml_backend_cuda_device_context *)dev->context;
    return ctx->name.c_str();
}

static const char * ggml_backend_cuda_device_get_description(ggml_backend_dev_t dev) {
    ggml_backend_cuda_device_context * ctx = (ggml_backend_cuda_device_context *)dev->context;
    return ctx->description.c_str();
}

static void ggml_backend_cuda_device_get_memory(ggml_backend_dev_t dev, size_t * free, size_t * total) {
    ggml_backend_cuda_device_context * ctx = (ggml_backend_cuda_device_context *)dev->context;
    ggml_cuda_set_device(ctx->device);
    CUDA_CHECK(cudaMemGetInfo(free, total));
}

static enum ggml_backend_dev_type ggml_backend_cuda_device_get_type(ggml_backend_dev_t dev) {
    GGML_UNUSED(dev);
    return GGML_BACKEND_DEVICE_TYPE_GPU;
}

static void ggml_backend_cuda_device_get_props(ggml_backend_dev_t dev, ggml_backend_dev_props * props) {
    props->name        = ggml_backend_cuda_device_get_name(dev);
    props->description = ggml_backend_cuda_device_get_description(dev);
    props->type        = ggml_backend_cuda_device_get_type(dev);
    ggml_backend_cuda_device_get_memory(dev, &props->memory_free, &props->memory_total);

    bool host_buffer = getenv("GGML_CUDA_NO_PINNED") == nullptr;
#ifdef GGML_CUDA_NO_PEER_COPY
    bool events = false;
#else
    bool events = true;
#endif

    props->caps = {
        /* .async                 = */ true,
        /* .host_buffer           = */ host_buffer,
        /* .buffer_from_host_ptr  = */ false,
        /* .events                = */ events,
    };
}

static ggml_backend_t ggml_backend_cuda_device_init_backend(ggml_backend_dev_t dev, const char * params) {
    GGML_UNUSED(params);
    ggml_backend_cuda_device_context * ctx = (ggml_backend_cuda_device_context *)dev->context;
    return ggml_backend_cuda_init(ctx->device);
}

static ggml_backend_buffer_type_t ggml_backend_cuda_device_get_buffer_type(ggml_backend_dev_t dev) {
    ggml_backend_cuda_device_context * ctx = (ggml_backend_cuda_device_context *)dev->context;
    return ggml_backend_cuda_buffer_type(ctx->device);
}

static ggml_backend_buffer_type_t ggml_backend_cuda_device_get_host_buffer_type(ggml_backend_dev_t dev) {
    GGML_UNUSED(dev);
    return ggml_backend_cuda_host_buffer_type();
}

// TODO: move these functions here
static bool ggml_backend_cuda_device_supports_op(ggml_backend_dev_t dev, const ggml_tensor * op) {
    ggml_backend_cuda_device_context * dev_ctx = (ggml_backend_cuda_device_context *) dev->context;

    // split buffers can only be used with GGML_OP_MUL_MAT
    if (op->op != GGML_OP_MUL_MAT) {
        for (int i = 0; i < GGML_MAX_SRC; i++) {
            if (op->src[i] && op->src[i]->buffer && ggml_backend_buft_is_cuda_split(op->src[i]->buffer->buft)) {
                return false;
            }
        }
    }

    // check if all the sources are allocated on this device
    for (int i = 0; i < GGML_MAX_SRC; i++) {
        if (op->src[i] && op->src[i]->buffer && ggml_backend_buft_is_cuda(op->src[i]->buffer->buft)) {
            ggml_backend_cuda_buffer_type_context * buft_ctx = (ggml_backend_cuda_buffer_type_context *)op->src[i]->buffer->buft->context;
            if (buft_ctx->device != dev_ctx->device) {
                return false;
            }
        }
    }

    switch (op->op) {
        case GGML_OP_UNARY:
            switch (ggml_get_unary_op(op)) {
                case GGML_UNARY_OP_NEG:
                case GGML_UNARY_OP_STEP:
                case GGML_UNARY_OP_GELU:
                case GGML_UNARY_OP_SILU:
                case GGML_UNARY_OP_RELU:
                case GGML_UNARY_OP_SIGMOID:
                case GGML_UNARY_OP_HARDSIGMOID:
                case GGML_UNARY_OP_HARDSWISH:
                case GGML_UNARY_OP_GELU_QUICK:
                case GGML_UNARY_OP_TANH:
                case GGML_UNARY_OP_EXP:
                    return ggml_is_contiguous(op->src[0]);
                default:
                    return false;
            }
            break;
        case GGML_OP_MUL_MAT:
        case GGML_OP_MUL_MAT_ID:
            {
                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)) {
                    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;
                    get_row_split(&row_low, &row_high, a, buft_ctx->tensor_split, dev_ctx->device);
                    if (row_low == row_high) {
                        return false;
                    }
                }
                if (b->type == GGML_TYPE_F16 && a->type != GGML_TYPE_F16) {
                    return false;
                }
#ifdef GGML_USE_MUSA
                if (b->type == GGML_TYPE_F16 && b->ne[2]*b->ne[3] > 1 &&
                    !ggml_is_transposed(a) && !ggml_is_transposed(b)) {
                    return false;
                }
#endif // GGML_USE_MUSA
                switch (a->type) {
                    case GGML_TYPE_F32:
                    case GGML_TYPE_F16:
                    case GGML_TYPE_Q4_0:
                    case GGML_TYPE_Q4_1:
                    case GGML_TYPE_Q5_0:
                    case GGML_TYPE_Q5_1:
                    case GGML_TYPE_Q8_0:
                    case GGML_TYPE_Q2_K:
                    case GGML_TYPE_Q3_K:
                    case GGML_TYPE_Q4_K:
                    case GGML_TYPE_Q5_K:
                    case GGML_TYPE_Q6_K:
                    case GGML_TYPE_Q8_K:
                    case GGML_TYPE_IQ1_M:
                    case GGML_TYPE_IQ1_S:
                    case GGML_TYPE_IQ2_S:
                    case GGML_TYPE_IQ2_XS:
                    case GGML_TYPE_IQ2_XXS:
                    case GGML_TYPE_IQ3_S:
                    case GGML_TYPE_IQ3_XXS:
                    case GGML_TYPE_IQ4_NL:
                    case GGML_TYPE_IQ4_XS:
                    case GGML_TYPE_BF16:
#ifdef GGML_USE_MUSA
                        if (a->type == GGML_TYPE_Q3_K) {
                            return false;
                        }
#endif // GGML_USE_MUSA
                        return true;
                    default:
                        return false;
                }
            } break;
        case GGML_OP_OUT_PROD:
            return op->type == GGML_TYPE_F32 && op->src[0]->type == GGML_TYPE_F32 && op->src[1]->type == GGML_TYPE_F32;
        case GGML_OP_GET_ROWS:
            {
                switch (op->src[0]->type) {
                    case GGML_TYPE_F16:
                    case GGML_TYPE_F32:
                    case GGML_TYPE_Q4_0:
                    case GGML_TYPE_Q4_1:
                    case GGML_TYPE_Q5_0:
                    case GGML_TYPE_Q5_1:
                    case GGML_TYPE_Q8_0:
                        return true;
                    default:
                        return false;
                }
            } break;
        case GGML_OP_GET_ROWS_BACK:
            {
                return op->type == GGML_TYPE_F32 && op->src[0]->type == GGML_TYPE_F32 && op->ne[2] == 1 && op->ne[3] == 1;
            } break;
        case GGML_OP_CPY:
            {
                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_F16) {
                    return true;
                }
                if (src0_type == GGML_TYPE_F32 && src1_type == GGML_TYPE_Q8_0) {
                    return true;
                }
                if (src0_type == GGML_TYPE_Q8_0 && src1_type == GGML_TYPE_F32) {
                    return true;
                }
                if (src0_type == GGML_TYPE_F32 && src1_type == GGML_TYPE_Q4_0) {
                    return true;
                }
                if (src0_type == GGML_TYPE_F32 && src1_type == GGML_TYPE_Q4_1) {
                    return true;
                }
                if (src0_type == GGML_TYPE_F32 && src1_type == GGML_TYPE_Q5_0) {
                    return true;
                }
                if (src0_type == GGML_TYPE_F32 && src1_type == GGML_TYPE_Q5_1) {
                    return true;
                }
                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;
                }
                return false;
            } break;
        case GGML_OP_DUP:
            {
                ggml_type src0_type = op->src[0]->type;
                return src0_type != GGML_TYPE_I32 && src0_type != GGML_TYPE_I16;
            } break;
        case GGML_OP_ARGMAX:
        case GGML_OP_COUNT_EQUAL:
            {
                return true;
            } break;
        case GGML_OP_REPEAT:
            {
                ggml_type src0_type = op->src[0]->type;
                return src0_type != GGML_TYPE_I32 && src0_type != GGML_TYPE_I16;
            } break;
        case GGML_OP_REPEAT_BACK:
                return op->type == GGML_TYPE_F32 && (op->src[0]->ne[2]*op->src[0]->ne[3]) <= (1 << 15);
        case GGML_OP_CONCAT:
            {
                ggml_type src0_type = op->src[0]->type;
                return src0_type != GGML_TYPE_I32 && src0_type != GGML_TYPE_I16;
            } break;
        case GGML_OP_CONV_TRANSPOSE_1D:
            {
                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;
                }
                return false;
            } break;
        case GGML_OP_SILU_BACK:
            return ggml_is_contiguous(op->src[0]);
            break;
        case GGML_OP_NORM:
        case GGML_OP_RMS_NORM:
            return true;
        case GGML_OP_RMS_NORM_BACK:
            return ggml_is_contiguous(op->src[0]) && op->ne[0] % WARP_SIZE == 0;
            break;
        case GGML_OP_NONE:
        case GGML_OP_RESHAPE:
        case GGML_OP_VIEW:
        case GGML_OP_PERMUTE:
        case GGML_OP_TRANSPOSE:
        case GGML_OP_ADD:
        case GGML_OP_ADD1:
        case GGML_OP_SUB:
        case GGML_OP_MUL:
        case GGML_OP_DIV:
        case GGML_OP_SCALE:
        case GGML_OP_SQR:
        case GGML_OP_SQRT:
        case GGML_OP_SIN:
        case GGML_OP_COS:
        case GGML_OP_CLAMP:
            return true;
        case GGML_OP_CONT:
            return op->src[0]->type != GGML_TYPE_BF16;
        case GGML_OP_DIAG_MASK_INF:
        case GGML_OP_SOFT_MAX:
            return true;
        case GGML_OP_SOFT_MAX_BACK: {
            float max_bias = 0.0f;
            memcpy(&max_bias, (const float *) op->op_params + 1, sizeof(float));
            return max_bias == 0.0f;
        }
        case GGML_OP_ROPE:
        case GGML_OP_ROPE_BACK: {
            const size_t ts = ggml_type_size(op->src[0]->type);
            const int64_t ne0_012 = op->src[0]->ne[0] * op->src[0]->ne[1] * op->src[0]->ne[2];
            return op->src[0]->nb[0] == ts && op->src[0]->nb[3] == ne0_012*ts;
        }
        case GGML_OP_IM2COL:
        case GGML_OP_POOL_2D:
        case GGML_OP_SUM:
        case GGML_OP_SUM_ROWS:
        case GGML_OP_ARGSORT:
        case GGML_OP_ACC:
            return true;
        case GGML_OP_GROUP_NORM:
            return ggml_is_contiguous(op->src[0]);
        case GGML_OP_UPSCALE:
        case GGML_OP_PAD:
        case GGML_OP_ARANGE:
        case GGML_OP_TIMESTEP_EMBEDDING:
        case GGML_OP_LEAKY_RELU:
        case GGML_OP_RWKV_WKV6:
        case GGML_OP_GATED_LINEAR_ATTN:
            return true;
        case GGML_OP_FLASH_ATTN_EXT: {
#ifndef FLASH_ATTN_AVAILABLE
            return false;
#endif
            if (op->src[1]->type == GGML_TYPE_BF16 || op->src[2]->type == GGML_TYPE_BF16) {
                return false;
            }
            if (op->src[0]->ne[0] ==  64 && op->src[1]->type == GGML_TYPE_F16) {
                return true;
            }
            if (op->src[0]->ne[0] == 128) {
                return true;
            }
            if (op->src[0]->ne[0] == 256 && op->src[1]->type == GGML_TYPE_F16 && op->src[2]->type == GGML_TYPE_F16) {
                return true;
            }
            return fp16_mma_available(ggml_cuda_info().devices[dev_ctx->device].cc) &&
                op->src[1]->type == GGML_TYPE_F16 && op->src[2]->type == GGML_TYPE_F16;
        }
        case GGML_OP_CROSS_ENTROPY_LOSS:
        case GGML_OP_CROSS_ENTROPY_LOSS_BACK:
        case GGML_OP_OPT_STEP_ADAMW:
            return true;
        default:
            return false;
    }
}

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;
}

static int64_t get_op_batch_size(const ggml_tensor * op) {
    switch (op->op) {
        case GGML_OP_GET_ROWS:
            return 0;
        case GGML_OP_MUL_MAT:
            return op->ne[1];
        case GGML_OP_MUL_MAT_ID:
        case GGML_OP_ROPE:
        case GGML_OP_ROPE_BACK:
            return op->ne[2];
        default:
            return ggml_nrows(op);
    }
}

static bool ggml_backend_cuda_device_offload_op(ggml_backend_dev_t dev, const ggml_tensor * op) {
    const int min_batch_size = 32;

    return get_op_batch_size(op) >= min_batch_size;

    GGML_UNUSED(dev);
}

static ggml_backend_event_t ggml_backend_cuda_device_event_new(ggml_backend_dev_t dev) {
#ifdef GGML_CUDA_NO_PEER_COPY
    return nullptr;
#else
    ggml_backend_cuda_device_context * dev_ctx = (ggml_backend_cuda_device_context *)dev->context;

    ggml_cuda_set_device(dev_ctx->device);

    cudaEvent_t event;
    CUDA_CHECK(cudaEventCreateWithFlags(&event, cudaEventDisableTiming));

    return new ggml_backend_event {
        /* .device  = */ dev,
        /* .context = */ event,
    };
#endif
}

static void ggml_backend_cuda_device_event_free(ggml_backend_dev_t dev, ggml_backend_event_t event) {
    GGML_UNUSED(dev);

    CUDA_CHECK(cudaEventDestroy((cudaEvent_t)event->context));
    delete event;
}

static void ggml_backend_cuda_device_event_synchronize(ggml_backend_dev_t dev, ggml_backend_event_t event) {
    GGML_UNUSED(dev);
    CUDA_CHECK(cudaEventSynchronize((cudaEvent_t)event->context));
}

static const ggml_backend_device_i ggml_backend_cuda_device_interface = {
    /* .get_name                = */ ggml_backend_cuda_device_get_name,
    /* .get_description         = */ ggml_backend_cuda_device_get_description,
    /* .get_memory              = */ ggml_backend_cuda_device_get_memory,
    /* .get_type                = */ ggml_backend_cuda_device_get_type,
    /* .get_props               = */ ggml_backend_cuda_device_get_props,
    /* .init_backend            = */ ggml_backend_cuda_device_init_backend,
    /* .get_buffer_type         = */ ggml_backend_cuda_device_get_buffer_type,
    /* .get_host_buffer_type    = */ ggml_backend_cuda_device_get_host_buffer_type,
    /* .buffer_from_host_ptr    = */ NULL,
    /* .supports_op             = */ ggml_backend_cuda_device_supports_op,
    /* .supports_buft           = */ ggml_backend_cuda_device_supports_buft,
    /* .offload_op              = */ ggml_backend_cuda_device_offload_op,
    /* .event_new               = */ ggml_backend_cuda_device_event_new,
    /* .event_free              = */ ggml_backend_cuda_device_event_free,
    /* .event_synchronize       = */ ggml_backend_cuda_device_event_synchronize,
};

// backend reg

struct ggml_backend_cuda_reg_context {
    std::vector<ggml_backend_dev_t> devices;
};

static const char * ggml_backend_cuda_reg_get_name(ggml_backend_reg_t reg) {
    GGML_UNUSED(reg);
    return GGML_CUDA_NAME;
}

static size_t ggml_backend_cuda_reg_get_device_count(ggml_backend_reg_t reg) {
    ggml_backend_cuda_reg_context * ctx = (ggml_backend_cuda_reg_context *)reg->context;
    return ctx->devices.size();
}

static ggml_backend_dev_t ggml_backend_cuda_reg_get_device(ggml_backend_reg_t reg, size_t index) {
    ggml_backend_cuda_reg_context * ctx = (ggml_backend_cuda_reg_context *)reg->context;
    GGML_ASSERT(index < ctx->devices.size());
    return ctx->devices[index];
}

static ggml_backend_feature * ggml_backend_cuda_get_features(ggml_backend_reg_t reg) {
    static std::vector<ggml_backend_feature> features = []() {
        std::vector<ggml_backend_feature> features;
    #define _STRINGIFY(...) #__VA_ARGS__
    #define STRINGIFY(...) _STRINGIFY(__VA_ARGS__)

    #ifdef __CUDA_ARCH_LIST__
        features.push_back({ "ARCHS", STRINGIFY(__CUDA_ARCH_LIST__) });
    #endif

    #ifdef GGML_CUDA_FORCE_MMQ
        features.push_back({ "FORCE_MMQ", "1" });
    #endif

    #ifdef GGML_CUDA_FORCE_CUBLAS
        features.push_back({ "FORCE_CUBLAS", "1" });
    #endif

    #ifndef GGML_USE_VMM
        features.push_back({ "NO_VMM", "1" });
    #endif

    #ifdef GGML_CUDA_NO_PEER_COPY
        features.push_back({ "NO_PEER_COPY", "1" });
    #endif

    #ifdef GGML_CUDA_F16
        features.push_back({ "F16", "1" });
    #endif

    #ifdef GGML_CUDA_USE_GRAPHS
        features.push_back({ "USE_GRAPHS", "1" });
    #endif

    #ifdef GGML_CUDA_PEER_MAX_BATCH_SIZE
        features.push_back({ "PEER_MAX_BATCH_SIZE", STRINGIFY(GGML_CUDA_PEER_MAX_BATCH_SIZE) });
    #endif

    #ifdef GGML_CUDA_FA_ALL_QUANTS
        features.push_back({ "FA_ALL_QUANTS", "1" });
    #endif

    #undef _STRINGIFY
    #undef STRINGIFY

        features.push_back({ nullptr, nullptr });

        return features;
    }();

    return features.data();

    GGML_UNUSED(reg);
}

static void * ggml_backend_cuda_reg_get_proc_address(ggml_backend_reg_t reg, const char * name) {
    GGML_UNUSED(reg);
    if (strcmp(name, "ggml_backend_split_buffer_type") == 0) {
        return (void *)ggml_backend_cuda_split_buffer_type;
    }
    if (strcmp(name, "ggml_backend_register_host_buffer") == 0) {
        return (void *)ggml_backend_cuda_register_host_buffer;
    }
    if (strcmp(name, "ggml_backend_unregister_host_buffer") == 0) {
        return (void *)ggml_backend_cuda_unregister_host_buffer;
    }
    if (strcmp(name, "ggml_backend_get_features") == 0) {
        return (void *)ggml_backend_cuda_get_features;
    }
    return nullptr;
}

static const ggml_backend_reg_i ggml_backend_cuda_reg_interface = {
    /* .get_name          = */ ggml_backend_cuda_reg_get_name,
    /* .get_device_count  = */ ggml_backend_cuda_reg_get_device_count,
    /* .get_device        = */ ggml_backend_cuda_reg_get_device,
    /* .get_proc_address  = */ ggml_backend_cuda_reg_get_proc_address,
};

// backend registry
ggml_backend_reg_t ggml_backend_cuda_reg() {
    static ggml_backend_reg reg;
    static bool initialized = false;

    {
        static std::mutex mutex;
        std::lock_guard<std::mutex> lock(mutex);
        if (!initialized) {
            ggml_backend_cuda_reg_context * ctx = new ggml_backend_cuda_reg_context;

            for (int i = 0; i < ggml_cuda_info().device_count; i++) {
                ggml_backend_cuda_device_context * dev_ctx = new ggml_backend_cuda_device_context;
                dev_ctx->device = i;
                dev_ctx->name = GGML_CUDA_NAME + std::to_string(i);

                ggml_cuda_set_device(i);
                cudaDeviceProp prop;
                CUDA_CHECK(cudaGetDeviceProperties(&prop, i));
                dev_ctx->description = prop.name;

                ggml_backend_dev_t dev = new ggml_backend_device {
                    /* .iface   = */ ggml_backend_cuda_device_interface,
                    /* .reg     = */ &reg,
                    /* .context = */ dev_ctx
                };
                ctx->devices.push_back(dev);
            }

            reg = ggml_backend_reg {
                /* .api_version = */ GGML_BACKEND_API_VERSION,
                /* .iface       = */ ggml_backend_cuda_reg_interface,
                /* .context     = */ ctx
            };
        }

        initialized = true;
    }

    return &reg;
}

ggml_backend_t ggml_backend_cuda_init(int device) {
    if (device < 0 || device >= ggml_backend_cuda_get_device_count()) {
        GGML_LOG_ERROR("%s: invalid device %d\n", __func__, device);
        return nullptr;
    }

    ggml_backend_cuda_context * ctx = new ggml_backend_cuda_context(device);
    if (ctx == nullptr) {
        GGML_LOG_ERROR("%s: failed to allocate context\n", __func__);
        return nullptr;
    }

    ggml_backend_t cuda_backend = new ggml_backend {
        /* .guid      = */ ggml_backend_cuda_guid(),
        /* .interface = */ ggml_backend_cuda_interface,
        /* .device    = */ ggml_backend_reg_dev_get(ggml_backend_cuda_reg(), device),
        /* .context   = */ ctx,
    };

    return cuda_backend;
}

GGML_BACKEND_DL_IMPL(ggml_backend_cuda_reg)