node-llama-cpp 3.0.0-beta.9 → 3.0.0

package/llama/addon/addon.cpp

@@ -0,0 +1,223 @@
+#include "addonGlobals.h"
+#include "AddonModel.h"
+#include "AddonModelLora.h"
+#include "AddonGrammar.h"
+#include "AddonGrammarEvaluationState.h"
+#include "AddonSampler.h"
+#include "AddonContext.h"
+#include "globals/addonLog.h"
+#include "globals/addonProgress.h"
+#include "globals/getGpuInfo.h"
+
+bool backendInitialized = false;
+bool backendDisposed = false;
+
+Napi::Value systemInfo(const Napi::CallbackInfo& info) {
+    return Napi::String::From(info.Env(), llama_print_system_info());
+}
+
+Napi::Value addonGetSupportsGpuOffloading(const Napi::CallbackInfo& info) {
+    return Napi::Boolean::New(info.Env(), llama_supports_gpu_offload());
+}
+
+Napi::Value addonGetSupportsMmap(const Napi::CallbackInfo& info) {
+    return Napi::Boolean::New(info.Env(), llama_supports_mmap());
+}
+
+Napi::Value addonGetSupportsMlock(const Napi::CallbackInfo& info) {
+    return Napi::Boolean::New(info.Env(), llama_supports_mlock());
+}
+
+Napi::Value addonGetMathCores(const Napi::CallbackInfo& info) {
+    return Napi::Number::New(info.Env(), cpu_get_num_math());
+}
+
+Napi::Value addonGetBlockSizeForGgmlType(const Napi::CallbackInfo& info) {
+    const int ggmlType = info[0].As<Napi::Number>().Int32Value();
+
+    if (ggmlType < 0 || ggmlType > GGML_TYPE_COUNT) {
+        return info.Env().Undefined();
+    }
+
+    const auto blockSize = ggml_blck_size(static_cast<ggml_type>(ggmlType));
+
+    return Napi::Number::New(info.Env(), blockSize);
+}
+
+Napi::Value addonGetTypeSizeForGgmlType(const Napi::CallbackInfo& info) {
+    const int ggmlType = info[0].As<Napi::Number>().Int32Value();
+
+    if (ggmlType < 0 || ggmlType > GGML_TYPE_COUNT) {
+        return info.Env().Undefined();
+    }
+
+    const auto typeSize = ggml_type_size(static_cast<ggml_type>(ggmlType));
+
+    return Napi::Number::New(info.Env(), typeSize);
+}
+
+Napi::Value addonGetConsts(const Napi::CallbackInfo& info) {
+    Napi::Object consts = Napi::Object::New(info.Env());
+    consts.Set("ggmlMaxDims", Napi::Number::New(info.Env(), GGML_MAX_DIMS));
+    consts.Set("ggmlTypeF16Size", Napi::Number::New(info.Env(), ggml_type_size(GGML_TYPE_F16)));
+    consts.Set("ggmlTypeF32Size", Napi::Number::New(info.Env(), ggml_type_size(GGML_TYPE_F32)));
+    consts.Set("ggmlTensorOverhead", Napi::Number::New(info.Env(), ggml_tensor_overhead()));
+    consts.Set("llamaPosSize", Napi::Number::New(info.Env(), sizeof(llama_pos)));
+    consts.Set("llamaSeqIdSize", Napi::Number::New(info.Env(), sizeof(llama_seq_id)));
+
+    return consts;
+}
+
+class AddonBackendLoadWorker : public Napi::AsyncWorker {
+    public:
+        AddonBackendLoadWorker(const Napi::Env& env)
+            : Napi::AsyncWorker(env, "AddonBackendLoadWorker"),
+              deferred(Napi::Promise::Deferred::New(env)) {
+        }
+        ~AddonBackendLoadWorker() {
+        }
+
+        Napi::Promise GetPromise() {
+            return deferred.Promise();
+        }
+
+    protected:
+        Napi::Promise::Deferred deferred;
+
+        void Execute() {
+            try {
+                llama_backend_init();
+
+                try {
+                    if (backendDisposed) {
+                        llama_backend_free();
+                    } else {
+                        backendInitialized = true;
+                    }
+                } catch (const std::exception& e) {
+                    SetError(e.what());
+                } catch(...) {
+                    SetError("Unknown error when calling \"llama_backend_free\"");
+                }
+            } catch (const std::exception& e) {
+                SetError(e.what());
+            } catch(...) {
+                SetError("Unknown error when calling \"llama_backend_init\"");
+            }
+        }
+        void OnOK() {
+            deferred.Resolve(Env().Undefined());
+        }
+        void OnError(const Napi::Error& err) {
+            deferred.Reject(err.Value());
+        }
+};
+
+
+class AddonBackendUnloadWorker : public Napi::AsyncWorker {
+    public:
+        AddonBackendUnloadWorker(const Napi::Env& env)
+            : Napi::AsyncWorker(env, "AddonBackendUnloadWorker"),
+              deferred(Napi::Promise::Deferred::New(env)) {
+        }
+        ~AddonBackendUnloadWorker() {
+        }
+
+        Napi::Promise GetPromise() {
+            return deferred.Promise();
+        }
+
+    protected:
+        Napi::Promise::Deferred deferred;
+
+        void Execute() {
+            try {
+                if (backendInitialized) {
+                    backendInitialized = false;
+                    llama_backend_free();
+                }
+            } catch (const std::exception& e) {
+                SetError(e.what());
+            } catch(...) {
+                SetError("Unknown error when calling \"llama_backend_free\"");
+            }
+        }
+        void OnOK() {
+            deferred.Resolve(Env().Undefined());
+        }
+        void OnError(const Napi::Error& err) {
+            deferred.Reject(err.Value());
+        }
+};
+
+Napi::Value addonInit(const Napi::CallbackInfo& info) {
+    if (backendInitialized) {
+        Napi::Promise::Deferred deferred = Napi::Promise::Deferred::New(info.Env());
+        deferred.Resolve(info.Env().Undefined());
+        return deferred.Promise();
+    }
+
+    AddonBackendLoadWorker* worker = new AddonBackendLoadWorker(info.Env());
+    worker->Queue();
+    return worker->GetPromise();
+}
+
+Napi::Value addonDispose(const Napi::CallbackInfo& info) {
+    if (backendDisposed) {
+        Napi::Promise::Deferred deferred = Napi::Promise::Deferred::New(info.Env());
+        deferred.Resolve(info.Env().Undefined());
+        return deferred.Promise();
+    }
+
+    backendDisposed = true;
+
+    AddonBackendUnloadWorker* worker = new AddonBackendUnloadWorker(info.Env());
+    worker->Queue();
+    return worker->GetPromise();
+}
+
+static void addonFreeLlamaBackend(Napi::Env env, int* data) {
+    if (backendDisposed) {
+        return;
+    }
+
+    backendDisposed = true;
+    if (backendInitialized) {
+        backendInitialized = false;
+        llama_backend_free();
+    }
+}
+
+Napi::Object registerCallback(Napi::Env env, Napi::Object exports) {
+    exports.DefineProperties({
+        Napi::PropertyDescriptor::Function("systemInfo", systemInfo),
+        Napi::PropertyDescriptor::Function("getSupportsGpuOffloading", addonGetSupportsGpuOffloading),
+        Napi::PropertyDescriptor::Function("getSupportsMmap", addonGetSupportsMmap),
+        Napi::PropertyDescriptor::Function("getSupportsMlock", addonGetSupportsMlock),
+        Napi::PropertyDescriptor::Function("getMathCores", addonGetMathCores),
+        Napi::PropertyDescriptor::Function("getBlockSizeForGgmlType", addonGetBlockSizeForGgmlType),
+        Napi::PropertyDescriptor::Function("getTypeSizeForGgmlType", addonGetTypeSizeForGgmlType),
+        Napi::PropertyDescriptor::Function("getConsts", addonGetConsts),
+        Napi::PropertyDescriptor::Function("setLogger", setLogger),
+        Napi::PropertyDescriptor::Function("setLoggerLogLevel", setLoggerLogLevel),
+        Napi::PropertyDescriptor::Function("getGpuVramInfo", getGpuVramInfo),
+        Napi::PropertyDescriptor::Function("getGpuDeviceInfo", getGpuDeviceInfo),
+        Napi::PropertyDescriptor::Function("getGpuType", getGpuType),
+        Napi::PropertyDescriptor::Function("init", addonInit),
+        Napi::PropertyDescriptor::Function("dispose", addonDispose),
+    });
+    AddonModel::init(exports);
+    AddonModelLora::init(exports);
+    AddonGrammar::init(exports);
+    AddonGrammarEvaluationState::init(exports);
+    AddonContext::init(exports);
+    AddonSampler::init(exports);
+
+    llama_log_set(addonLlamaCppLogCallback, nullptr);
+
+    exports.AddFinalizer(addonFreeLlamaBackend, static_cast<int*>(nullptr));
+
+    return exports;
+}
+
+NODE_API_MODULE(NODE_GYP_MODULE_NAME, registerCallback)

package/llama/addon/addonGlobals.cpp

@@ -0,0 +1,22 @@
+#include <sstream>
+#include <vector>
+#include "addonGlobals.h"
+#include "napi.h"
+
+void adjustNapiExternalMemoryAdd(Napi::Env env, uint64_t size) {
+    const uint64_t chunkSize = std::numeric_limits<int64_t>::max();
+    while (size > 0) {
+        int64_t adjustSize = std::min(size, chunkSize);
+        Napi::MemoryManagement::AdjustExternalMemory(env, adjustSize);
+        size -= adjustSize;
+    }
+}
+
+void adjustNapiExternalMemorySubtract(Napi::Env env, uint64_t size) {
+    const uint64_t chunkSize = std::numeric_limits<int64_t>::max();
+    while (size > 0) {
+        int64_t adjustSize = std::min(size, chunkSize);
+        Napi::MemoryManagement::AdjustExternalMemory(env, -adjustSize);
+        size -= adjustSize;
+    }
+}

package/llama/addon/addonGlobals.h

@@ -0,0 +1,12 @@
+#pragma once
+#include "napi.h"
+
+class AddonModel;
+class AddonModelLora;
+class AddonModelData;
+class AddonContext;
+class AddonGrammar;
+class AddonGrammarEvaluationState;
+
+void adjustNapiExternalMemoryAdd(Napi::Env env, uint64_t size);
+void adjustNapiExternalMemorySubtract(Napi::Env env, uint64_t size);

package/llama/addon/globals/addonLog.cpp

@@ -0,0 +1,136 @@
+#include <sstream>
+
+#include "addonLog.h"
+
+AddonThreadSafeLogCallbackFunction addonThreadSafeLoggerCallback;
+bool addonJsLoggerCallbackSet = false;
+int addonLoggerLogLevel = 5;
+
+static int addonGetGgmlLogLevelNumber(ggml_log_level level) {
+    switch (level) {
+        case GGML_LOG_LEVEL_ERROR: return 2;
+        case GGML_LOG_LEVEL_WARN: return 3;
+        case GGML_LOG_LEVEL_INFO: return 4;
+        case GGML_LOG_LEVEL_NONE: return 5;
+        case GGML_LOG_LEVEL_DEBUG: return 6;
+    }
+
+    return 1;
+}
+
+void addonCallJsLogCallback(
+    Napi::Env env, Napi::Function callback, AddonThreadSafeLogCallbackFunctionContext* context, addon_logger_log* data
+) {
+    bool called = false;
+
+    if (env != nullptr && callback != nullptr && addonJsLoggerCallbackSet) {
+        try {
+            callback.Call({
+                Napi::Number::New(env, data->logLevelNumber),
+                Napi::String::New(env, data->stringStream->str()),
+            });
+            called = true;
+        } catch (const Napi::Error& e) {
+            called = false;
+        }
+    }
+
+    if (!called && data != nullptr) {
+        if (data->logLevelNumber == 2) {
+            fputs(data->stringStream->str().c_str(), stderr);
+            fflush(stderr);
+        } else {
+            fputs(data->stringStream->str().c_str(), stdout);
+            fflush(stdout);
+        }
+    }
+
+    if (data != nullptr) {
+        delete data->stringStream;
+        delete data;
+    }
+}
+
+void addonLlamaCppLogCallback(ggml_log_level level, const char* text, void* user_data) {
+    int logLevelNumber = addonGetGgmlLogLevelNumber(level);
+
+    if (logLevelNumber > addonLoggerLogLevel) {
+        return;
+    }
+
+    if (addonJsLoggerCallbackSet) {
+        std::stringstream* stringStream = new std::stringstream();
+        if (text != nullptr) {
+            *stringStream << text;
+        }
+
+        addon_logger_log* data = new addon_logger_log {
+            logLevelNumber,
+            stringStream,
+        };
+
+        auto status = addonThreadSafeLoggerCallback.NonBlockingCall(data);
+
+        if (status == napi_ok) {
+            return;
+        } else {
+            delete stringStream;
+            delete data;
+        }
+    }
+
+    if (text != nullptr) {
+        if (level == 2) {
+            fputs(text, stderr);
+            fflush(stderr);
+        } else {
+            fputs(text, stdout);
+            fflush(stdout);
+        }
+    }
+}
+
+Napi::Value setLogger(const Napi::CallbackInfo& info) {
+    if (info.Length() < 1 || !info[0].IsFunction()) {
+        if (addonJsLoggerCallbackSet) {
+            addonJsLoggerCallbackSet = false;
+            addonThreadSafeLoggerCallback.Release();
+        }
+
+        return info.Env().Undefined();
+    }
+
+    auto addonLoggerJSCallback = info[0].As<Napi::Function>();
+    AddonThreadSafeLogCallbackFunctionContext* context = new Napi::Reference<Napi::Value>(Napi::Persistent(info.This()));
+    addonThreadSafeLoggerCallback = AddonThreadSafeLogCallbackFunction::New(
+        info.Env(),
+        addonLoggerJSCallback,
+        "loggerCallback",
+        0,
+        1,
+        context,
+        [](Napi::Env, void*, AddonThreadSafeLogCallbackFunctionContext* ctx) {
+            addonJsLoggerCallbackSet = false;
+
+            delete ctx;
+        }
+    );
+    addonJsLoggerCallbackSet = true;
+
+    // prevent blocking the main node process from exiting due to active resources
+    addonThreadSafeLoggerCallback.Unref(info.Env());
+
+    return info.Env().Undefined();
+}
+
+Napi::Value setLoggerLogLevel(const Napi::CallbackInfo& info) {
+    if (info.Length() < 1 || !info[0].IsNumber()) {
+        addonLoggerLogLevel = 5;
+
+        return info.Env().Undefined();
+    }
+
+    addonLoggerLogLevel = info[0].As<Napi::Number>().Int32Value();
+
+    return info.Env().Undefined();
+}

package/llama/addon/globals/addonLog.h

@@ -0,0 +1,21 @@
+#pragma once
+#include "llama.h"
+#include "napi.h"
+
+struct addon_logger_log {
+    public:
+        const int logLevelNumber;
+        const std::stringstream* stringStream;
+};
+
+void addonLlamaCppLogCallback(ggml_log_level level, const char* text, void* user_data);
+
+using AddonThreadSafeLogCallbackFunctionContext = Napi::Reference<Napi::Value>;
+void addonCallJsLogCallback(
+    Napi::Env env, Napi::Function callback, AddonThreadSafeLogCallbackFunctionContext* context, addon_logger_log* data
+);
+using AddonThreadSafeLogCallbackFunction =
+    Napi::TypedThreadSafeFunction<AddonThreadSafeLogCallbackFunctionContext, addon_logger_log, addonCallJsLogCallback>;
+
+Napi::Value setLogger(const Napi::CallbackInfo& info);
+Napi::Value setLoggerLogLevel(const Napi::CallbackInfo& info);

package/llama/addon/globals/addonProgress.cpp

@@ -0,0 +1,15 @@
+#include "addonProgress.h"
+
+void addonCallJsProgressCallback(
+    Napi::Env env, Napi::Function callback, AddonThreadSafeProgressCallbackFunctionContext* context, addon_progress_event* data
+) {
+    if (env != nullptr && callback != nullptr) {
+        try {
+            callback.Call({Napi::Number::New(env, data->progress)});
+        } catch (const Napi::Error& e) {}
+    }
+
+    if (data != nullptr) {
+        delete data;
+    }
+}

package/llama/addon/globals/addonProgress.h

@@ -0,0 +1,15 @@
+#pragma once
+#include "napi.h"
+
+struct addon_progress_event {
+    public:
+        const float progress;
+};
+
+using AddonThreadSafeProgressCallbackFunctionContext = Napi::Reference<Napi::Value>;
+void addonCallJsProgressCallback(
+    Napi::Env env, Napi::Function callback, AddonThreadSafeProgressCallbackFunctionContext* context, addon_progress_event* data
+);
+using AddonThreadSafeProgressEventCallbackFunction =
+    Napi::TypedThreadSafeFunction<AddonThreadSafeProgressCallbackFunctionContext, addon_progress_event, addonCallJsProgressCallback>;
+

package/llama/addon/globals/getGpuInfo.cpp

@@ -0,0 +1,108 @@
+#include "getGpuInfo.h"
+#include "addonLog.h"
+
+#ifdef GPU_INFO_USE_CUDA
+#  include "../../gpuInfo/cuda-gpu-info.h"
+#endif
+#ifdef GPU_INFO_USE_VULKAN
+#  include "../../gpuInfo/vulkan-gpu-info.h"
+#endif
+#ifdef GPU_INFO_USE_METAL
+#  include "../../gpuInfo/metal-gpu-info.h"
+#endif
+
+
+#ifdef GPU_INFO_USE_CUDA
+void logCudaError(const char* message) {
+    addonLlamaCppLogCallback(GGML_LOG_LEVEL_ERROR, (std::string("CUDA error: ") + std::string(message)).c_str(), nullptr);
+}
+#endif
+#ifdef GPU_INFO_USE_VULKAN
+void logVulkanWarning(const char* message) {
+    addonLlamaCppLogCallback(GGML_LOG_LEVEL_WARN, (std::string("Vulkan warning: ") + std::string(message)).c_str(), nullptr);
+}
+#endif
+
+Napi::Value getGpuVramInfo(const Napi::CallbackInfo& info) {
+    uint64_t total = 0;
+    uint64_t used = 0;
+
+#ifdef GPU_INFO_USE_CUDA
+    size_t cudaDeviceTotal = 0;
+    size_t cudaDeviceUsed = 0;
+    bool cudeGetInfoSuccess = gpuInfoGetTotalCudaDevicesInfo(&cudaDeviceTotal, &cudaDeviceUsed, logCudaError);
+
+    if (cudeGetInfoSuccess) {
+        total += cudaDeviceTotal;
+        used += cudaDeviceUsed;
+    }
+#endif
+
+#ifdef GPU_INFO_USE_VULKAN
+    uint64_t vulkanDeviceTotal = 0;
+    uint64_t vulkanDeviceUsed = 0;
+    const bool vulkanDeviceSupportsMemoryBudgetExtension = gpuInfoGetTotalVulkanDevicesInfo(&vulkanDeviceTotal, &vulkanDeviceUsed, logVulkanWarning);
+
+    if (vulkanDeviceSupportsMemoryBudgetExtension) {
+        total += vulkanDeviceTotal;
+        used += vulkanDeviceUsed;
+    }
+#endif
+
+#ifdef GPU_INFO_USE_METAL
+    uint64_t metalDeviceTotal = 0;
+    uint64_t metalDeviceUsed = 0;
+    getMetalGpuInfo(&metalDeviceTotal, &metalDeviceUsed);
+
+    total += metalDeviceTotal;
+    used += metalDeviceUsed;
+#endif
+
+    Napi::Object result = Napi::Object::New(info.Env());
+    result.Set("total", Napi::Number::From(info.Env(), total));
+    result.Set("used", Napi::Number::From(info.Env(), used));
+
+    return result;
+}
+
+Napi::Value getGpuDeviceInfo(const Napi::CallbackInfo& info) {
+    std::vector<std::string> deviceNames;
+
+#ifdef GPU_INFO_USE_CUDA
+    gpuInfoGetCudaDeviceNames(&deviceNames, logCudaError);
+#endif
+
+#ifdef GPU_INFO_USE_VULKAN
+    gpuInfoGetVulkanDeviceNames(&deviceNames, logVulkanWarning);
+#endif
+
+#ifdef GPU_INFO_USE_METAL
+    getMetalGpuDeviceNames(&deviceNames);
+#endif
+
+    Napi::Object result = Napi::Object::New(info.Env());
+
+    Napi::Array deviceNamesNapiArray = Napi::Array::New(info.Env(), deviceNames.size());
+    for (size_t i = 0; i < deviceNames.size(); ++i) {
+        deviceNamesNapiArray[i] = Napi::String::New(info.Env(), deviceNames[i]);
+    }
+    result.Set("deviceNames", deviceNamesNapiArray);
+
+    return result;
+}
+
+Napi::Value getGpuType(const Napi::CallbackInfo& info) {
+#ifdef GPU_INFO_USE_CUDA
+    return Napi::String::New(info.Env(), "cuda");
+#endif
+
+#ifdef GPU_INFO_USE_VULKAN
+    return Napi::String::New(info.Env(), "vulkan");
+#endif
+
+#ifdef GPU_INFO_USE_METAL
+    return Napi::String::New(info.Env(), "metal");
+#endif
+
+    return info.Env().Undefined();
+}

package/llama/addon/globals/getGpuInfo.h

@@ -0,0 +1,6 @@
+#pragma once
+#include "napi.h"
+
+Napi::Value getGpuVramInfo(const Napi::CallbackInfo& info);
+Napi::Value getGpuDeviceInfo(const Napi::CallbackInfo& info);
+Napi::Value getGpuType(const Napi::CallbackInfo& info);

package/llama/binariesGithubRelease.json

@@ -1,3 +1,3 @@
 {
-    "release": "b2074"
+    "release": "b3808"
 }

package/llama/gpuInfo/cuda-gpu-info.cu

@@ -0,0 +1,120 @@
+#include <stddef.h>
+#include <vector>
+#include <string>
+
+#if defined(GPU_INFO_USE_HIPBLAS)
+#include <hip/hip_runtime.h>
+#include <hipblas/hipblas.h>
+#define cudaGetDevice hipGetDevice
+#define cudaGetDeviceCount hipGetDeviceCount
+#define cudaGetErrorString hipGetErrorString
+#define cudaMemGetInfo hipMemGetInfo
+#define cudaSetDevice hipSetDevice
+#define cudaSuccess hipSuccess
+#else
+#include <cuda_runtime.h>
+#include <cuda.h>
+#endif
+
+
+typedef void (*gpuInfoCudaErrorLogCallback_t)(const char* message);
+
+bool gpuInfoSetCudaDevice(const int device, gpuInfoCudaErrorLogCallback_t errorLogCallback) {
+    int current_device;
+    auto getDeviceResult = cudaGetDevice(&current_device);
+
+    if (getDeviceResult != cudaSuccess) {
+        errorLogCallback(cudaGetErrorString(getDeviceResult));
+        return false;
+    }
+
+    if (device == current_device) {
+        return true;
+    }
+
+    const auto setDeviceResult = cudaSetDevice(device);
+
+    if (setDeviceResult != cudaSuccess) {
+        errorLogCallback(cudaGetErrorString(setDeviceResult));
+        return false;
+    }
+
+    return true;
+}
+
+bool gpuInfoGetCudaDeviceInfo(int device, size_t * total, size_t * used, gpuInfoCudaErrorLogCallback_t errorLogCallback) {
+    gpuInfoSetCudaDevice(device, errorLogCallback);
+
+    size_t freeMem;
+    size_t totalMem;
+    auto getMemInfoResult = cudaMemGetInfo(&freeMem, &totalMem);
+
+    if (getMemInfoResult != cudaSuccess) {
+        errorLogCallback(cudaGetErrorString(getMemInfoResult));
+        return false;
+    }
+
+    *total = totalMem;
+    *used = totalMem - freeMem;
+
+    return true;
+}
+
+int gpuInfoGetCudaDeviceCount(gpuInfoCudaErrorLogCallback_t errorLogCallback) {
+    int deviceCount;
+    auto getDeviceCountResult = cudaGetDeviceCount(&deviceCount);
+
+    if (getDeviceCountResult != cudaSuccess) {
+        errorLogCallback(cudaGetErrorString(getDeviceCountResult));
+        return -1;
+    }
+
+    return deviceCount;
+}
+
+bool gpuInfoGetTotalCudaDevicesInfo(size_t * total, size_t * used, gpuInfoCudaErrorLogCallback_t errorLogCallback) {
+    int deviceCount = gpuInfoGetCudaDeviceCount(errorLogCallback);
+
+    if (deviceCount < 0) {
+        return false;
+    }
+
+    size_t usedMem = 0;
+    size_t totalMem = 0;
+
+    for (int i = 0; i < deviceCount; i++) {
+        size_t deviceUsedMem;
+        size_t deviceTotalMem;
+
+        if (!gpuInfoGetCudaDeviceInfo(i, &deviceTotalMem, &deviceUsedMem, errorLogCallback)) {
+            return false;
+        }
+
+        usedMem += deviceUsedMem;
+        totalMem += deviceTotalMem;
+    }
+
+    *total = totalMem;
+    *used = usedMem;
+
+    return true;
+}
+
+void gpuInfoGetCudaDeviceNames(std::vector<std::string> * deviceNames, gpuInfoCudaErrorLogCallback_t errorLogCallback) {
+    int deviceCount = gpuInfoGetCudaDeviceCount(errorLogCallback);
+
+    if (deviceCount < 0) {
+        return;
+    }
+
+    for (int i = 0; i < deviceCount; i++) {
+        cudaDeviceProp prop;
+        auto getDevicePropertiesResult = cudaGetDeviceProperties(&prop, i);
+
+        if (getDevicePropertiesResult != cudaSuccess) {
+            errorLogCallback(cudaGetErrorString(getDevicePropertiesResult));
+        } else {
+            (*deviceNames).push_back(std::string(prop.name));
+        }
+    }
+}