mudler/LocalAI · warning

LoRA directory does not exist or is not a directory: %s\n

Error message

LoRA directory does not exist or is not a directory: %s\n

What it means

Warning from discover_lora_files() in the stablediffusion-ggml shim: the configured LoRA directory does not exist or is not a directory, so the name->path LoRA lookup map is empty. Generation still runs; any <lora:...> tags in prompts will later fail to resolve (see the 'LoRA file not found' warning).

Source

Thrown at backend/go/stablediffusion-ggml/cpp/gosd.cpp:119

        embedding_vec.push_back(item);
        fprintf(stderr, "Found embedding: %s -> %s\n", item.name, item.path);
    }

    fprintf(stderr, "Loaded %zu embeddings from %s\n", embedding_vec.size(), embedding_dir);
}

// Discover LoRA files in directory and build a map of name -> path
static std::map<std::string, std::string> discover_lora_files(const char* lora_dir) {
    std::map<std::string, std::string> lora_map;

    if (!lora_dir || strlen(lora_dir) == 0) {
        fprintf(stderr, "LoRA directory not specified\n");
        return lora_map;
    }

    if (!std::filesystem::exists(lora_dir) || !std::filesystem::is_directory(lora_dir)) {
        fprintf(stderr, "LoRA directory does not exist or is not a directory: %s\n", lora_dir);
        return lora_map;
    }

    static const std::vector<std::string> valid_ext = {".safetensors", ".ckpt", ".pt", ".gguf"};

    fprintf(stderr, "Discovering LoRA files in: %s\n", lora_dir);

    for (const auto& entry : std::filesystem::directory_iterator(lora_dir)) {
        if (!entry.is_regular_file()) {
            continue;
        }

        auto path = entry.path();
        std::string ext = path.extension().string();

        bool valid = false;
        for (const auto& e : valid_ext) {
            if (ext == e) {

View on GitHub (pinned to 44413a9d06)

Solutions

1. Create the LoRA directory or correct the lora_dir option value
2. Confirm it is a directory, not a file (e.g. pointing lora_dir directly at a .safetensors file triggers this)
3. If LoRAs are not used, leave lora_dir unset to avoid the warning (the load-time warning 212 will note it)

Example fix

# before
options: "...,lora_dir=/models/loras"  # missing

# after
mkdir -p /models/loras && cp ~/loras/*.safetensors /models/loras/

// Go caller, before load
if opts.LoraDir != "" {
    info, err := os.Stat(opts.LoraDir)
    if err != nil || !info.IsDir() {
        return fmt.Errorf("lora dir %q missing or not a directory", opts.LoraDir)
    }
}

When it happens

Trigger: Setting lora_dir=<path> in the model load options where the path is missing, mistyped, a plain file, or relative to the wrong working directory.

Common situations: LoRA weights stored elsewhere than the configured default; docker volume not mounted at the expected path; per-user model directories where only some users have the lora folder.

Related errors

• Embedding directory does not exist or is not a directory: %s
• LoRA directory not set, cannot parse LoRAs from prompt\n
• WARNING: LoRA file not found: %s\n
• Invalid lora_apply_mode: %s, using default\n
• WARNING: LoRA model directory not set. LoRAs in prompts will

AI-assisted analysis of mudler/LocalAI@44413a9d06 (2026-08-15). Data as JSON: /api/errors/62ecdc0249b6426d. Report an issue: GitHub.