improve mixpanel usage statistics (#2238)

Other changes:
- Always display first start dialog if privacy options are unset (e.g. if the user closed GPT4All without selecting them)
- LocalDocs scanQueue is now always deferred
- Fix a potential crash in magic_match
- LocalDocs indexing is now started after the first start dialog is dismissed so usage stats are included

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

gpt4all-backend/gptj.cpp

@@ -785,7 +785,7 @@ const std::vector<LLModel::Token> &GPTJ::endTokens() const
     return fres;
 }
 
-std::string get_arch_name(gguf_context *ctx_gguf) {
+const char *get_arch_name(gguf_context *ctx_gguf) {
     std::string arch_name;
     const int kid = gguf_find_key(ctx_gguf, "general.architecture");
     enum gguf_type ktype = gguf_get_kv_type(ctx_gguf, kid);
@@ -814,21 +814,25 @@ DLL_EXPORT const char *get_build_variant() {
     return GGML_BUILD_VARIANT;
 }
 
-DLL_EXPORT bool magic_match(const char * fname) {
+DLL_EXPORT char *get_file_arch(const char *fname) {
     struct ggml_context * ctx_meta = NULL;
     struct gguf_init_params params = {
         /*.no_alloc = */ true,
         /*.ctx      = */ &ctx_meta,
     };
     gguf_context *ctx_gguf = gguf_init_from_file(fname, params);
-    if (!ctx_gguf)
-        return false;
 
-    bool isValid = gguf_get_version(ctx_gguf) <= 3;
-    isValid = isValid && get_arch_name(ctx_gguf) == "gptj";
+    char *arch = nullptr;
+    if (ctx_gguf && gguf_get_version(ctx_gguf) <= 3) {
+        arch = strdup(get_arch_name(ctx_gguf));
+    }
 
     gguf_free(ctx_gguf);
-    return isValid;
+    return arch;
+}
+
+DLL_EXPORT bool is_arch_supported(const char *arch) {
+    return !strcmp(arch, "gptj");
 }
 
 DLL_EXPORT LLModel *construct() {

gpt4all-backend/llamamodel.cpp

@@ -104,7 +104,7 @@ static int llama_sample_top_p_top_k(
     return llama_sample_token(ctx, &candidates_p);
 }
 
-std::string get_arch_name(gguf_context *ctx_gguf) {
+const char *get_arch_name(gguf_context *ctx_gguf) {
     std::string arch_name;
     const int kid = gguf_find_key(ctx_gguf, "general.architecture");
     enum gguf_type ktype = gguf_get_kv_type(ctx_gguf, kid);
@@ -961,25 +961,23 @@ DLL_EXPORT const char *get_build_variant() {
     return GGML_BUILD_VARIANT;
 }
 
-DLL_EXPORT bool magic_match(const char *fname) {
-    auto * ctx = load_gguf(fname);
-    std::string arch = get_arch_name(ctx);
-
-    bool valid = true;
-
-    if (std::find(KNOWN_ARCHES.begin(), KNOWN_ARCHES.end(), arch) == KNOWN_ARCHES.end()) {
-        // not supported by this version of llama.cpp
-        if (arch != "gptj") { // we support this via another module
-            std::cerr << __func__ << ": unsupported model architecture: " << arch << "\n";
+DLL_EXPORT char *get_file_arch(const char *fname) {
+    auto *ctx = load_gguf(fname);
+    char *arch = nullptr;
+    if (ctx) {
+        std::string archStr = get_arch_name(ctx);
+        if (is_embedding_arch(archStr) && gguf_find_key(ctx, (archStr + ".pooling_type").c_str()) < 0) {
+            // old bert.cpp embedding model
+        } else {
+            arch = strdup(archStr.c_str());
         }
-        valid = false;
     }
-
-    if (valid && is_embedding_arch(arch) && gguf_find_key(ctx, (arch + ".pooling_type").c_str()) < 0)
-        valid = false; // old pre-llama.cpp embedding model, e.g. all-MiniLM-L6-v2-f16.gguf
-
     gguf_free(ctx);
-    return valid;
+    return arch;
+}
+
+DLL_EXPORT bool is_arch_supported(const char *arch) {
+    return std::find(KNOWN_ARCHES.begin(), KNOWN_ARCHES.end(), std::string(arch)) < KNOWN_ARCHES.end();
 }
 
 DLL_EXPORT LLModel *construct() {

gpt4all-backend/llmodel.cpp

@@ -8,6 +8,7 @@
 #include <fstream>
 #include <iostream>
 #include <memory>
+#include <optional>
 #include <regex>
 #include <sstream>
 #include <string>
@@ -49,14 +50,17 @@ LLModel::Implementation::Implementation(Dlhandle &&dlhandle_)
     auto get_build_variant = m_dlhandle->get<const char *()>("get_build_variant");
     assert(get_build_variant);
     m_buildVariant = get_build_variant();
-    m_magicMatch = m_dlhandle->get<bool(const char*)>("magic_match");
-    assert(m_magicMatch);
+    m_getFileArch = m_dlhandle->get<char *(const char *)>("get_file_arch");
+    assert(m_getFileArch);
+    m_isArchSupported = m_dlhandle->get<bool(const char *)>("is_arch_supported");
+    assert(m_isArchSupported);
     m_construct = m_dlhandle->get<LLModel *()>("construct");
     assert(m_construct);
 }
 
 LLModel::Implementation::Implementation(Implementation &&o)
-    : m_magicMatch(o.m_magicMatch)
+    : m_getFileArch(o.m_getFileArch)
+    , m_isArchSupported(o.m_isArchSupported)
     , m_construct(o.m_construct)
     , m_modelType(o.m_modelType)
     , m_buildVariant(o.m_buildVariant)
@@ -123,18 +127,26 @@ const std::vector<LLModel::Implementation> &LLModel::Implementation::implementat
 
 const LLModel::Implementation* LLModel::Implementation::implementation(const char *fname, const std::string& buildVariant) {
     bool buildVariantMatched = false;
+    std::optional<std::string> archName;
     for (const auto& i : implementationList()) {
         if (buildVariant != i.m_buildVariant) continue;
         buildVariantMatched = true;
 
-        if (!i.m_magicMatch(fname)) continue;
-        return &i;
+        char *arch = i.m_getFileArch(fname);
+        if (!arch) continue;
+        archName = arch;
+
+        bool archSupported = i.m_isArchSupported(arch);
+        free(arch);
+        if (archSupported) return &i;
     }
 
     if (!buildVariantMatched)
-        throw std::runtime_error("Could not find any implementations for build variant: " + buildVariant);
+        throw MissingImplementationError("Could not find any implementations for build variant: " + buildVariant);
+    if (!archName)
+        throw UnsupportedModelError("Unsupported file format");
 
-    return nullptr; // unsupported model format
+    throw BadArchError(std::move(*archName));
 }
 
 LLModel *LLModel::Implementation::construct(const std::string &modelPath, std::string buildVariant, int n_ctx) {
@@ -144,7 +156,11 @@ LLModel *LLModel::Implementation::construct(const std::string &modelPath, std::s
     #if defined(__APPLE__) && defined(__arm64__) // FIXME: See if metal works for intel macs
         if (buildVariant == "auto") {
             size_t total_mem = getSystemTotalRAMInBytes();
-            impl = implementation(modelPath.c_str(), "metal");
+            try {
+                impl = implementation(modelPath.c_str(), "metal");
+            } catch (const std::exception &e) {
+                // fall back to CPU
+            }
             if(impl) {
                 LLModel* metalimpl = impl->m_construct();
                 metalimpl->m_implementation = impl;
@@ -177,7 +193,6 @@ LLModel *LLModel::Implementation::construct(const std::string &modelPath, std::s
             }
         }
         impl = implementation(modelPath.c_str(), buildVariant);
-        if (!impl) return nullptr;
     }
 
     // Construct and return llmodel implementation

gpt4all-backend/llmodel.h

@@ -17,6 +17,29 @@ class LLModel {
 public:
     using Token = int32_t;
 
+    class BadArchError: public std::runtime_error {
+    public:
+        BadArchError(std::string arch)
+            : runtime_error("Unsupported model architecture: " + arch)
+            , m_arch(std::move(arch))
+            {}
+
+        const std::string &arch() const noexcept { return m_arch; }
+
+    private:
+        std::string m_arch;
+    };
+
+    class MissingImplementationError: public std::runtime_error {
+    public:
+        using std::runtime_error::runtime_error;
+    };
+
+    class UnsupportedModelError: public std::runtime_error {
+    public:
+        using std::runtime_error::runtime_error;
+    };
+
     struct GPUDevice {
         int index;
         int type;
@@ -53,7 +76,8 @@ public:
         static const Implementation *implementation(const char *fname, const std::string &buildVariant);
         static LLModel *constructDefaultLlama();
 
-        bool (*m_magicMatch)(const char *fname);
+        char *(*m_getFileArch)(const char *fname);
+        bool (*m_isArchSupported)(const char *arch);
         LLModel *(*m_construct)();
 
         std::string_view m_modelType;

gpt4all-backend/llmodel_c.cpp

@@ -40,11 +40,6 @@ llmodel_model llmodel_model_create2(const char *model_path, const char *build_va
         return nullptr;
     }
 
-    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;