llm-manager 1.10.0

use regex::Regex;
use std::sync::LazyLock;

fn compile(pattern: &str) -> Regex {
    Regex::new(pattern).expect("invalid regex pattern")
}

// ── Phase detection patterns ──────────────────────────────────────

/// Matches llama.cpp server startup lines: "llama server", "ggml version", etc.
pub static LLAMA_START: LazyLock<Regex> =
    LazyLock::new(|| compile(r"(?i)(llama.*server|ggml\s+version)"));

/// Matches model loader start: "llama_model_loader"
pub static LOADING_MODEL: LazyLock<Regex> = LazyLock::new(|| compile(r"(?i)llama_model_loader"));

/// Matches metadata loaded: "loaded X meta", "meta data"
pub static LOADED_META: LazyLock<Regex> =
    LazyLock::new(|| compile(r"(?i)(loaded\s+\d+\s+meta|meta\s+data)"));

/// Matches tensor loading start: "load tensors:"
pub static LOAD_TENSORS: LazyLock<Regex> = LazyLock::new(|| compile(r"(?i)load tensors:"));

/// Matches "server listening on" or "http server listening"
pub static SERVER_LISTENING: LazyLock<Regex> = LazyLock::new(|| {
    compile(r"(?i)(server listening|http server listening|load_model:\s*initializing\s+slots)")
});

// ── Detail parsing patterns ───────────────────────────────────────

/// Matches "loading tensor 12 of 345" or "loading tensor 12 out of 345"
pub static LOADING_TENSOR: LazyLock<Regex> =
    LazyLock::new(|| compile(r"(?i)loading tensor\s+(\d+)\s+(?:of|out of)\s+(\d+)"));

/// Matches "offloading 32 repeating layers to GPU"
pub static OFFLOADING_LAYERS: LazyLock<Regex> =
    LazyLock::new(|| compile(r"(?i)offloading\s+(\d+)\s+repeating layers"));

/// Matches "offloaded X/Y layers" or "offloaded X out of Y layers"
pub static OFFLOADED_LAYERS: LazyLock<Regex> =
    LazyLock::new(|| compile(r"(?i)offloaded\s+(\d+)\s*(?:out\s+of|/)\s*(\d+)\s*layers"));

/// Matches "Vulkan0 model buffer size = 12345.67 MiB" or "CPU model buffer size = 12345.67 MiB"
pub static MODEL_BUFFER_SIZE: LazyLock<Regex> = LazyLock::new(|| {
    compile(
        r"(?i)((?:vulkan\d|roc\d|cuda|cpu|metal)?(?:\s*))model buffer size\s*=\s*([\d.]+)\s*MiB",
    )
});

/// Matches "kv buffer size = 1234.56 MiB"
pub static KV_BUFFER_SIZE: LazyLock<Regex> =
    LazyLock::new(|| compile(r"(?i)kv buffer size\s*=\s*([\d.]+)\s*MiB"));

/// Matches "loading model: 12 of 345" or "loading model: 12/345"
pub static LOADING_MODEL_PROGRESS: LazyLock<Regex> =
    LazyLock::new(|| compile(r"(?i)loading model:\s*(\d+)\s*(?:of|out of|/)\s*(\d+)"));

// ── Error detection ───────────────────────────────────────────────

/// Detects whether a log line indicates a loading error.
pub fn is_loading_error(msg: &str) -> bool {
    let lower = msg.to_lowercase();
    // OOM patterns — check first (most common)
    lower.contains("out of memory")
        || lower.contains("out_of_memory")
        || lower.contains("out_of_device_memory")
        || lower.contains("outofmemory")
        || lower.contains("outofdevicememory")
        // Actual loading error patterns — must be specific to model loading
        || lower.contains("failed to load")
        || lower.contains("failed loading")
        || lower.contains("failed to initialize")
        || lower.contains("load failed")
        || lower.contains("load_model: error")
        || lower.contains("error loading")
        || lower.contains("exception")
        || lower.contains("vk::systemerror")
}

/// Detects whether the error is specifically an OOM.
pub fn is_oom_error(msg: &str) -> bool {
    let lower = msg.to_lowercase();
    lower.contains("out of memory")
        || lower.contains("out_of_memory")
        || lower.contains("out_of_device_memory")
        || lower.contains("outofmemory")
        || lower.contains("outofdevicememory")
}