fix non-AVX CPU detection (#2141)

* chat: fix non-AVX CPU detection on Windows
* bindings: throw exception instead of logging to console

Signed-off-by: Jared Van Bortel <jared@nomic.ai>

gpt4all-backend/llmodel.cpp

@@ -19,33 +19,27 @@
 
 std::string s_implementations_search_path = ".";
 
-static bool has_at_least_minimal_hardware() {
-#if defined(__x86_64__) || defined(_M_X64)
-    #ifndef _MSC_VER
-        return __builtin_cpu_supports("avx");
-    #else
-        int cpuInfo[4];
-        __cpuid(cpuInfo, 1);
-        return cpuInfo[2] & (1 << 28);
-    #endif
-#else
-    return true; // Don't know how to handle non-x86_64
-#endif
-}
+#if !(defined(__x86_64__) || defined(_M_X64))
+    // irrelevant on non-x86_64
+    #define cpu_supports_avx()  -1
+    #define cpu_supports_avx2() -1
+#elif defined(_MSC_VER)
+    // MSVC
+    static int get_cpu_info(int func_id, int reg_id) {
+        int info[4];
+        __cpuid(info, func_id);
+        return info[reg_id];
+    }
 
-static bool requires_avxonly() {
-#if defined(__x86_64__) || defined(_M_X64)
-    #ifndef _MSC_VER
-        return !__builtin_cpu_supports("avx2");
-    #else
-        int cpuInfo[4];
-        __cpuidex(cpuInfo, 7, 0);
-        return !(cpuInfo[1] & (1 << 5));
-    #endif
+    // AVX via EAX=1: Processor Info and Feature Bits, bit 28 of ECX
+    #define cpu_supports_avx()  (get_cpu_info(1, 2) & (1 << 28))
+    // AVX2 via EAX=7, ECX=0: Extended Features, bit 5 of EBX
+    #define cpu_supports_avx2() (get_cpu_info(7, 1) & (1 <<  5))
 #else
-    return false; // Don't know how to handle non-x86_64
+    // gcc/clang
+    #define cpu_supports_avx()  __builtin_cpu_supports("avx")
+    #define cpu_supports_avx2() __builtin_cpu_supports("avx2")
 #endif
-}
 
 LLModel::Implementation::Implementation(Dlhandle &&dlhandle_)
     : m_dlhandle(new Dlhandle(std::move(dlhandle_))) {
@@ -71,21 +65,25 @@ LLModel::Implementation::Implementation(Implementation &&o)
 }
 
 LLModel::Implementation::~Implementation() {
-    if (m_dlhandle) delete m_dlhandle;
+    delete m_dlhandle;
 }
 
-bool LLModel::Implementation::isImplementation(const Dlhandle &dl) {
+static bool isImplementation(const Dlhandle &dl) {
     return dl.get<bool(uint32_t)>("is_g4a_backend_model_implementation");
 }
 
 const std::vector<LLModel::Implementation> &LLModel::Implementation::implementationList() {
+    if (cpu_supports_avx() == 0) {
+        throw std::runtime_error("CPU does not support AVX");
+    }
+
     // NOTE: allocated on heap so we leak intentionally on exit so we have a chance to clean up the
     // individual models without the cleanup of the static list interfering
     static auto* libs = new std::vector<Implementation>([] () {
         std::vector<Implementation> fres;
 
-        std::string impl_name_re = "(bert|gptj|llamamodel-mainline)";
-        if (requires_avxonly()) {
+        std::string impl_name_re = "(gptj|llamamodel-mainline)";
+        if (cpu_supports_avx2() == 0) {
             impl_name_re += "-avxonly";
         } else {
             impl_name_re += "-(default|metal)";
@@ -107,9 +105,8 @@ const std::vector<LLModel::Implementation> &LLModel::Implementation::implementat
                     // Add to list if model implementation
                     try {
                         Dlhandle dl(p.string());
-                        if (!Implementation::isImplementation(dl)) {
+                        if (!isImplementation(dl))
                             continue;
-                        }
                         fres.emplace_back(Implementation(std::move(dl)));
                     } catch (...) {}
                 }
@@ -134,18 +131,13 @@ const LLModel::Implementation* LLModel::Implementation::implementation(const cha
         return &i;
     }
 
-    if (!buildVariantMatched) {
-        std::cerr << "LLModel ERROR: Could not find any implementations for build variant: " << buildVariant << "\n";
-    }
-    return nullptr;
+    if (!buildVariantMatched)
+        throw std::runtime_error("Could not find any implementations for build variant: " + buildVariant);
+
+    return nullptr; // unsupported model format
 }
 
 LLModel *LLModel::Implementation::construct(const std::string &modelPath, std::string buildVariant, int n_ctx) {
-    if (!has_at_least_minimal_hardware()) {
-        std::cerr << "LLModel ERROR: CPU does not support AVX\n";
-        return nullptr;
-    }
-
     // Get correct implementation
     const Implementation* impl = nullptr;
 
@@ -178,7 +170,7 @@ LLModel *LLModel::Implementation::construct(const std::string &modelPath, std::s
     if (!impl) {
         //TODO: Auto-detect CUDA/OpenCL
         if (buildVariant == "auto") {
-            if (requires_avxonly()) {
+            if (cpu_supports_avx2() == 0) {
                 buildVariant = "avxonly";
             } else {
                 buildVariant = "default";
@@ -196,15 +188,24 @@ LLModel *LLModel::Implementation::construct(const std::string &modelPath, std::s
 
 LLModel *LLModel::Implementation::constructDefaultLlama() {
     static std::unique_ptr<LLModel> llama([]() -> LLModel * {
+        const std::vector<LLModel::Implementation> *impls;
+        try {
+            impls = &implementationList();
+        } catch (const std::runtime_error &e) {
+            std::cerr << __func__ << ": implementationList failed: " << e.what() << "\n";
+            return nullptr;
+        }
+
         const LLModel::Implementation *impl = nullptr;
-        for (const auto &i : implementationList()) {
+        for (const auto &i: *impls) {
             if (i.m_buildVariant == "metal" || i.m_modelType != "LLaMA") continue;
             impl = &i;
         }
         if (!impl) {
-            std::cerr << "LLModel ERROR: Could not find CPU LLaMA implementation\n";
+            std::cerr << __func__ << ": could not find llama.cpp implementation\n";
             return nullptr;
         }
+
         auto fres = impl->m_construct();
         fres->m_implementation = impl;
         return fres;
@@ -240,3 +241,7 @@ void LLModel::Implementation::setImplementationsSearchPath(const std::string& pa
 const std::string& LLModel::Implementation::implementationsSearchPath() {
     return s_implementations_search_path;
 }
+
+bool LLModel::Implementation::hasSupportedCPU() {
+    return cpu_supports_avx() != 0;
+}

gpt4all-backend/llmodel.h

@@ -30,7 +30,6 @@ public:
 
     class Implementation {
     public:
-        Implementation(Dlhandle &&);
         Implementation(const Implementation &) = delete;
         Implementation(Implementation &&);
         ~Implementation();
@@ -38,9 +37,6 @@ public:
         std::string_view modelType() const { return m_modelType; }
         std::string_view buildVariant() const { return m_buildVariant; }
 
-        static bool isImplementation(const Dlhandle &dl);
-        static const std::vector<Implementation> &implementationList();
-        static const Implementation *implementation(const char *fname, const std::string &buildVariant);
         static LLModel *construct(const std::string &modelPath, std::string buildVariant = "auto", int n_ctx = 2048);
         static std::vector<GPUDevice> availableGPUDevices();
         static int32_t maxContextLength(const std::string &modelPath);
@@ -48,8 +44,13 @@ public:
         static bool isEmbeddingModel(const std::string &modelPath);
         static void setImplementationsSearchPath(const std::string &path);
         static const std::string &implementationsSearchPath();
+        static bool hasSupportedCPU();
 
     private:
+        Implementation(Dlhandle &&);
+
+        static const std::vector<Implementation> &implementationList();
+        static const Implementation *implementation(const char *fname, const std::string &buildVariant);
         static LLModel *constructDefaultLlama();
 
         bool (*m_magicMatch)(const char *fname);

gpt4all-backend/llmodel_c.cpp

@@ -13,8 +13,6 @@ struct LLModelWrapper {
     ~LLModelWrapper() { delete llModel; }
 };
 
-thread_local static std::string last_error_message;
-
 llmodel_model llmodel_model_create(const char *model_path) {
     const char *error;
     auto fres = llmodel_model_create2(model_path, "auto", &error);
@@ -24,24 +22,30 @@ llmodel_model llmodel_model_create(const char *model_path) {
     return fres;
 }
 
+static void llmodel_set_error(const char **errptr, const char *message) {
+    thread_local static std::string last_error_message;
+    if (errptr) {
+        last_error_message = message;
+        *errptr = last_error_message.c_str();
+    }
+}
+
 llmodel_model llmodel_model_create2(const char *model_path, const char *build_variant, const char **error) {
-    auto wrapper = new LLModelWrapper;
-
+    LLModel *llModel;
     try {
-        wrapper->llModel = LLModel::Implementation::construct(model_path, build_variant);
-        if (!wrapper->llModel) {
-            last_error_message = "Model format not supported (no matching implementation found)";
-        }
+        llModel = LLModel::Implementation::construct(model_path, build_variant);
     } catch (const std::exception& e) {
-        last_error_message = e.what();
+        llmodel_set_error(error, e.what());
+        return nullptr;
     }
 
-    if (!wrapper->llModel) {
-        delete std::exchange(wrapper, nullptr);
-        if (error) {
-            *error = last_error_message.c_str();
-        }
+    if (!llModel) {
+        llmodel_set_error(error, "Model format not supported (no matching implementation found)");
+        return nullptr;
     }
+
+    auto wrapper = new LLModelWrapper;
+    wrapper->llModel = llModel;
     return wrapper;
 }
 
@@ -159,8 +163,7 @@ float *llmodel_embed(
     auto *wrapper = static_cast<LLModelWrapper *>(model);
 
     if (!texts || !*texts) {
-        if (error)
-            *error = strdup("'texts' is NULL or empty");
+        llmodel_set_error(error, "'texts' is NULL or empty");
         return nullptr;
     }
 
@@ -183,8 +186,7 @@ float *llmodel_embed(
         embedding = new float[embd_size];
         wrapper->llModel->embed(textsVec, embedding, prefixStr, dimensionality, do_mean, atlas);
     } catch (std::exception const &e) {
-        if (error)
-            *error = strdup(e.what());
+        llmodel_set_error(error, e.what());
         return nullptr;
     }
 

gpt4all-bindings/python/gpt4all/_pyllmodel.py

@@ -185,7 +185,7 @@ class LLModel:
         model = llmodel.llmodel_model_create2(self.model_path, b"auto", ctypes.byref(err))
         if model is None:
             s = err.value
-            raise ValueError(f"Unable to instantiate model: {'null' if s is None else s.decode()}")
+            raise RuntimeError(f"Unable to instantiate model: {'null' if s is None else s.decode()}")
         self.model = model
 
     def __del__(self):

gpt4all-chat/llm.cpp

@@ -1,4 +1,5 @@
 #include "llm.h"
+#include "../gpt4all-backend/llmodel.h"
 #include "../gpt4all-backend/sysinfo.h"
 
 #include <QCoreApplication>
@@ -25,22 +26,8 @@ LLM *LLM::globalInstance()
 
 LLM::LLM()
     : QObject{nullptr}
-    , m_compatHardware(true)
+    , m_compatHardware(LLModel::Implementation::hasSupportedCPU())
 {
-#if defined(__x86_64__)
-    #ifndef _MSC_VER
-        const bool minimal(__builtin_cpu_supports("avx"));
-    #else
-        int cpuInfo[4];
-        __cpuid(cpuInfo, 1);
-        const bool minimal(cpuInfo[2] & (1 << 28));
-    #endif
-#else
-    const bool minimal = true; // Don't know how to handle non-x86_64
-#endif
-
-    m_compatHardware = minimal;
-
     QNetworkInformation::loadDefaultBackend();
     auto * netinfo = QNetworkInformation::instance();
     if (netinfo) {

gpt4all-chat/modellist.cpp

@@ -228,11 +228,11 @@ int ModelInfo::maxContextLength() const
     if (!installed || isOnline) return -1;
     if (m_maxContextLength != -1) return m_maxContextLength;
     auto path = (dirpath + filename()).toStdString();
-    int layers = LLModel::Implementation::maxContextLength(path);
-    if (layers < 0) {
-        layers = 4096; // fallback value
+    int n_ctx = LLModel::Implementation::maxContextLength(path);
+    if (n_ctx < 0) {
+        n_ctx = 4096; // fallback value
     }
-    m_maxContextLength = layers;
+    m_maxContextLength = n_ctx;
     return m_maxContextLength;
 }