feat(client): file logging + llama.cpp fallback cooldown and parallel options

README.md

@@ -133,11 +133,11 @@ sudo bash deploy/install-client.sh --server http://your-server:5050 --token "cha
 ```
 or using ollama models with llama.cpp backend using ROCm:
 ```bash
-sudo bash deploy/install-client.sh   --server https://ai.domain.tld   --token "change-me"   --use-llama-cpp-via-docker   --use-ollama-models-path /usr/share/ollama/.ollama/models   --llama-cpp-docker-image ghcr.io/ggml-org/llama.cpp:server-rocm   --llama-cpp-base-port 8081
+sudo bash deploy/install-client.sh   --server https://ai.domain.tld   --token "change-me"   --use-llama-cpp-via-docker   --use-ollama-models-path /usr/share/ollama/.ollama/models   --llama-cpp-docker-image ghcr.io/ggml-org/llama.cpp:server-rocm   --llama-cpp-base-port 8081   --llama-cpp-parallel 128
 ```
 or using ollama models with llama.cpp backend using CUDA:
 ```bash
-sudo bash deploy/install-client.sh   --server https://ai.domain.tld   --token "change-me"   --use-llama-cpp-via-docker   --use-ollama-models-path /usr/share/ollama/.ollama/models   --llama-cpp-docker-image ghcr.io/ggml-org/llama.cpp:server-cuda   --llama-cpp-base-port 8081
+sudo bash deploy/install-client.sh   --server https://ai.domain.tld   --token "change-me"   --use-llama-cpp-via-docker   --use-ollama-models-path /usr/share/ollama/.ollama/models   --llama-cpp-docker-image ghcr.io/ggml-org/llama.cpp:server-cuda   --llama-cpp-base-port 8081   --llama-cpp-parallel 128
 ```
 
 Options: `--server`, `--token` (required); `--client-id`, `--upstream`, `--install-dir`, `--service-name`, `--no-ollama` (optional). Missing required values are prompted interactively.
@@ -150,6 +150,15 @@ llama.cpp via Docker options (replaces Ollama for inferencing):
 | `--use-ollama-models-path <dir>` | Path to Ollama models directory (`manifests/blobs`); required with the flag above |
 | `--llama-cpp-docker-image <img>` | Docker image; defaults to auto-detected (rocm/cuda/cpu) |
 | `--llama-cpp-base-port <num>` | Base port for containers; defaults to `8081` |
+| `--llama-cpp-parallel <num>` | llama.cpp parallel slots per container; if unset, llama.cpp's own default is used (which is `1`) |
+| `--llama-cpp-fallback-cooldown <sec>` | Seconds before llama.cpp is retried after a failed container start; defaults to `180` |
+| `--log-dir <dir>` | Directory for log files; defaults to `<app dir>/Logs` |
+
+### llama.cpp fallback to Ollama
+
+Models are served via llama.cpp Docker containers. If a container cannot be started for a model (for example, the model's GGUF blob is incompatible with the llama.cpp build), the client falls back to the Ollama upstream for that model. Transient start failures are remembered for `--llama-cpp-fallback-cooldown` seconds (default 3 minutes) and then retried; a container that starts but exits before becoming ready marks the model as falling back until it is unloaded. `load`/`unload` model commands and on-demand request routing are all covered; a failed container start is detected quickly by watching the container state, and the container log tail is written to the client log to aid debugging.
+
+Note: some hybrid SSM/attention models (e.g. `qwen3.5-coder-next`) are converted by Ollama into a GGUF tensor layout that stock llama.cpp cannot load (`missing tensor 'blk.0.ssm_dt.bias'` and similar). Such models are served via the Ollama fallback above. If you want them to run on llama.cpp instead, use a Hugging Face-converted GGUF (e.g. `unsloth/Qwen3-Coder-Next-GGUF`) rather than the Ollama blob.
 
 The script ensures .NET 10 and Ollama are installed, builds the client self-contained, installs it to `/opt/Ngino-client`, and creates a systemd service (`Ngino-client`). Logs: `journalctl -u Ngino-client -f`.
 

deploy/install-client.sh

@@ -20,6 +20,9 @@ USE_LLAMA_CPP_VIA_DOCKER=false
 USE_OLLAMA_MODELS_PATH=""
 LLAMA_CPP_DOCKER_IMAGE=""
 LLAMA_CPP_BASE_PORT=""
+LLAMA_CPP_PARALLEL=""
+LLAMA_CPP_FALLBACK_COOLDOWN=""
+LOG_DIR=""
 
 # ── Colors ────────────────────────────────────────────────────────────────────
 RED='\033[0;31m'
@@ -57,13 +60,19 @@ Optional:
                        llama.cpp Docker image; defaults to auto-detected (rocm/cuda/cpu)
   --llama-cpp-base-port <num>
                        Base port for llama.cpp containers; defaults to 8081
+  --llama-cpp-parallel <num>
+                       llama.cpp parallel slots per container; if unset, llama.cpp's own default is used
+  --llama-cpp-fallback-cooldown <sec>
+                       Seconds before llama.cpp is retried after a failed container start; defaults to 180
+  --log-dir <dir>      Directory for log files; defaults to <install-dir>/Logs
   -h, --help           Show this help message
 
 Examples:
   $0 --server http://gpu-server:5050 --token "my-secret"
   $0 --server http://gpu-server:5050 --token "my-secret" --no-ollama
   $0 --server http://gpu-server:5050 --token "my-secret" \\
-     --use-llama-cpp-via-docker --use-ollama-models-path /usr/share/ollama/.ollama/models
+     --use-llama-cpp-via-docker --use-ollama-models-path /usr/share/ollama/.ollama/models \\
+     --llama-cpp-parallel 128
 EOF
     exit 0
 }
@@ -82,6 +91,9 @@ while [[ $# -gt 0 ]]; do
         --use-ollama-models-path)   USE_OLLAMA_MODELS_PATH="$2"; shift 2 ;;
         --llama-cpp-docker-image)   LLAMA_CPP_DOCKER_IMAGE="$2"; shift 2 ;;
         --llama-cpp-base-port)      LLAMA_CPP_BASE_PORT="$2"; shift 2 ;;
+        --llama-cpp-parallel)       LLAMA_CPP_PARALLEL="$2"; shift 2 ;;
+        --llama-cpp-fallback-cooldown) LLAMA_CPP_FALLBACK_COOLDOWN="$2"; shift 2 ;;
+        --log-dir)                  LOG_DIR="$2"; shift 2 ;;
         -h|--help)                  usage ;;
         *)                          die "Unknown option: $1" ;;
     esac
@@ -311,6 +323,15 @@ mkdir -p "$ENV_DIR"
         if [[ -n "$LLAMA_CPP_BASE_PORT" ]]; then
             printf 'NGINO_LLAMA_CPP_BASE_PORT=%s\n' "$LLAMA_CPP_BASE_PORT"
         fi
+        if [[ -n "$LLAMA_CPP_PARALLEL" ]]; then
+            printf 'NGINO_LLAMA_CPP_PARALLEL=%s\n' "$LLAMA_CPP_PARALLEL"
+        fi
+        if [[ -n "$LLAMA_CPP_FALLBACK_COOLDOWN" ]]; then
+            printf 'NGINO_LLAMA_CPP_FALLBACK_COOLDOWN_SECONDS=%s\n' "$LLAMA_CPP_FALLBACK_COOLDOWN"
+        fi
+    fi
+    if [[ -n "$LOG_DIR" ]]; then
+        printf 'NGINO_LOG_DIR=%s\n' "$LOG_DIR"
     fi
 } > "$ENV_DIR/env"
 chmod 600 "$ENV_DIR/env"

src/Ngino.Client/ClientOptions.cs

@@ -28,6 +28,12 @@ internal sealed class ClientOptions
 
     public int LlamaCppBasePort { get; init; } = 8081;
 
+    public int? LlamaCppParallel { get; init; }
+
+    public TimeSpan LlamaCppFallbackCooldown { get; init; } = TimeSpan.FromMinutes(3);
+
+    public string? LogDirectory { get; init; }
+
     public Uri TunnelUri
     {
         get
@@ -68,7 +74,10 @@ internal sealed class ClientOptions
             UseLlamaCppViaDocker = ReadBool(values, false, "use-llama-cpp-via-docker", "NGINO_USE_LLAMA_CPP_VIA_DOCKER"),
             UseOllamaModelsPath = NormalizeDirectoryPath(Read(values, "use-ollama-models-path", "NGINO_USE_OLLAMA_MODELS_PATH")),
             LlamaCppDockerImage = Read(values, "llama-cpp-docker-image", "NGINO_LLAMA_CPP_DOCKER_IMAGE"),
-            LlamaCppBasePort = ReadInt(values, 8081, "llama-cpp-base-port", "NGINO_LLAMA_CPP_BASE_PORT")
+            LlamaCppBasePort = ReadInt(values, 8081, "llama-cpp-base-port", "NGINO_LLAMA_CPP_BASE_PORT"),
+            LlamaCppParallel = ReadOptionalInt(values, "llama-cpp-parallel", "NGINO_LLAMA_CPP_PARALLEL"),
+            LlamaCppFallbackCooldown = TimeSpan.FromSeconds(ReadInt(values, 180, "llama-cpp-fallback-cooldown", "NGINO_LLAMA_CPP_FALLBACK_COOLDOWN_SECONDS")),
+            LogDirectory = NormalizeDirectoryPath(Read(values, "log-dir", "NGINO_LOG_DIR"))
         };
     }
 
@@ -87,6 +96,10 @@ internal sealed class ClientOptions
           --use-ollama-models-path <dir> Path to Ollama models directory (manifests/blobs), required with --use-llama-cpp-via-docker
           --llama-cpp-docker-image <img> llama.cpp Docker image; defaults to auto-detected (rocm/cuda/cpu)
           --llama-cpp-base-port <num>    Base port for llama.cpp containers; defaults to 8081
+          --llama-cpp-parallel <num>     llama.cpp parallel slots per container; if unset, llama.cpp's own default is used
+          --llama-cpp-fallback-cooldown <sec>
+                                         Seconds before llama.cpp is retried after a failed container start; defaults to 180
+          --log-dir <dir>                Directory for log files; defaults to <app dir>/Logs
         """;
 
     private static Dictionary<string, string> ParseArgs(string[] args)
@@ -160,6 +173,12 @@ internal sealed class ClientOptions
         return int.TryParse(value, out var parsed) && parsed > 0 ? parsed : fallback;
     }
 
+    private static int? ReadOptionalInt(Dictionary<string, string> values, params string[] keys)
+    {
+        var value = Read(values, keys);
+        return int.TryParse(value, out var parsed) && parsed > 0 ? parsed : null;
+    }
+
     private static bool ReadBool(Dictionary<string, string> values, bool fallback, params string[] keys)
     {
         var value = Read(values, keys);

src/Ngino.Client/FileLogger.cs

@@ -0,0 +1,113 @@
+using System.Text;
+using Microsoft.Extensions.Logging;
+
+namespace Ngino.Client;
+
+internal sealed class FileLoggerProvider : ILoggerProvider
+{
+    private const long DefaultMaxFileSizeBytes = 5L * 1024 * 1024;
+    private const string LogFileName = "ngino-client.log";
+    private const string RotatedLogFileName = "ngino-client.log.1";
+
+    private readonly string _directory;
+    private readonly long _maxFileSizeBytes;
+    private readonly object _lock = new();
+    private StreamWriter _writer = null!;
+    private string _currentFile = null!;
+
+    public FileLoggerProvider(string directory, long maxFileSizeBytes = DefaultMaxFileSizeBytes)
+    {
+        _directory = directory;
+        _maxFileSizeBytes = maxFileSizeBytes;
+        Directory.CreateDirectory(directory);
+        OpenFile();
+    }
+
+    public string LogDirectory => _directory;
+
+    public ILogger CreateLogger(string categoryName) => new FileLogger(this, categoryName);
+
+    public void WriteLog(DateTime timestamp, LogLevel level, string category, string message)
+    {
+        lock (_lock)
+        {
+            var line = $"{timestamp:yyyy-MM-dd HH:mm:ss.fff} [{level}] {category}: {message}";
+            if (_writer.BaseStream.Length + line.Length + 2 > _maxFileSizeBytes)
+            {
+                RotateFile();
+            }
+
+            _writer.WriteLine(line);
+        }
+    }
+
+    public void Dispose()
+    {
+        lock (_lock)
+        {
+            _writer.Dispose();
+        }
+    }
+
+    private void OpenFile()
+    {
+        _currentFile = Path.Combine(_directory, LogFileName);
+        _writer = new StreamWriter(
+            new FileStream(_currentFile, FileMode.Append, FileAccess.Write, FileShare.ReadWrite),
+            new UTF8Encoding(encoderShouldEmitUTF8Identifier: false))
+        {
+            AutoFlush = true
+        };
+    }
+
+    private void RotateFile()
+    {
+        _writer.Dispose();
+
+        var rotatedFile = Path.Combine(_directory, RotatedLogFileName);
+        try
+        {
+            File.Delete(rotatedFile);
+            if (File.Exists(_currentFile))
+            {
+                File.Move(_currentFile, rotatedFile);
+            }
+        }
+        catch (IOException)
+        {
+        }
+        catch (UnauthorizedAccessException)
+        {
+        }
+
+        OpenFile();
+    }
+}
+
+internal sealed class FileLogger(FileLoggerProvider provider, string category) : ILogger
+{
+    public IDisposable? BeginScope<TState>(TState state) where TState : notnull => null;
+
+    public bool IsEnabled(LogLevel logLevel) => logLevel >= LogLevel.Trace;
+
+    public void Log<TState>(
+        LogLevel logLevel,
+        EventId eventId,
+        TState state,
+        Exception? exception,
+        Func<TState, Exception?, string> formatter)
+    {
+        if (!IsEnabled(logLevel))
+        {
+            return;
+        }
+
+        var message = formatter(state, exception);
+        if (exception is not null)
+        {
+            message += Environment.NewLine + exception;
+        }
+
+        provider.WriteLog(DateTime.Now, logLevel, category, message);
+    }
+}

src/Ngino.Client/LlamaCppManager.cs

@@ -16,24 +16,35 @@ internal sealed partial class LlamaCppManager : IAsyncDisposable
     private const string NginoContainerLabel = "ngino-llamacpp";
     private static readonly TimeSpan DockerTimeout = TimeSpan.FromSeconds(60);
     private static readonly TimeSpan ContainerStartTimeout = TimeSpan.FromMinutes(5);
+    private static readonly TimeSpan DefaultFallbackCooldown = TimeSpan.FromMinutes(3);
 
     private readonly string _blobsPath;
     private readonly string _manifestsPath;
     private readonly string _dockerImage;
     private readonly int _basePort;
+    private readonly int? _parallel;
+    private readonly TimeSpan _fallbackCooldown;
     private readonly ILogger _logger;
     private readonly ConcurrentDictionary<string, int> _modelPorts = new(StringComparer.OrdinalIgnoreCase);
+    private readonly ConcurrentDictionary<int, byte> _reservedPorts = new();
+    private readonly ConcurrentDictionary<string, DateTime> _fallbackModels = new(StringComparer.OrdinalIgnoreCase);
+    private readonly ConcurrentDictionary<string, SemaphoreSlim> _modelStartLocks = new(StringComparer.OrdinalIgnoreCase);
+    private readonly object _portAllocationLock = new();
 
     public LlamaCppManager(
         string ollamaModelsPath,
         string? dockerImage,
         int? basePort,
-        ILogger? logger = null)
+        ILogger? logger = null,
+        TimeSpan fallbackCooldown = default,
+        int? parallel = null)
     {
         _manifestsPath = Path.Combine(ollamaModelsPath, "manifests");
         _blobsPath = Path.Combine(ollamaModelsPath, "blobs");
         _dockerImage = dockerImage ?? GetDefaultDockerImage();
         _basePort = basePort ?? DefaultBasePort;
+        _fallbackCooldown = fallbackCooldown > TimeSpan.Zero ? fallbackCooldown : DefaultFallbackCooldown;
+        _parallel = parallel is > 0 ? parallel : null;
         _logger = logger ?? NullLogger<LlamaCppManager>.Instance;
     }
 
@@ -78,6 +89,45 @@ internal sealed partial class LlamaCppManager : IAsyncDisposable
         return _modelPorts.ContainsKey(ollamaModelName);
     }
 
+    public void MarkModelAsFallback(string ollamaModelName)
+    {
+        _fallbackModels[ollamaModelName] = DateTime.UtcNow.Add(_fallbackCooldown);
+        _logger.LogWarning(
+            "Model {Model} will fall back to the Ollama upstream for {Cooldown} before llama.cpp is retried.",
+            ollamaModelName, _fallbackCooldown);
+    }
+
+    public void MarkModelAsPermanentFallback(string ollamaModelName)
+    {
+        _fallbackModels[ollamaModelName] = DateTime.MaxValue;
+        _logger.LogError(
+            "Model {Model} exited its llama.cpp container before becoming ready. It will fall back to the Ollama upstream until it is unloaded.",
+            ollamaModelName);
+    }
+
+    public bool IsModelOnFallback(string ollamaModelName)
+    {
+        if (_fallbackModels.TryGetValue(ollamaModelName, out var expiresAt))
+        {
+            if (expiresAt > DateTime.UtcNow)
+            {
+                return true;
+            }
+
+            _fallbackModels.TryRemove(ollamaModelName, out _);
+        }
+
+        return false;
+    }
+
+    public void ClearModelFallback(string ollamaModelName)
+    {
+        if (_fallbackModels.TryRemove(ollamaModelName, out _))
+        {
+            _logger.LogInformation("Cleared llama.cpp fallback marker for model {Model}.", ollamaModelName);
+        }
+    }
+
     public async Task<bool> IsContainerRunningAsync(string ollamaModelName)
     {
         if (!_modelPorts.ContainsKey(ollamaModelName))
@@ -96,7 +146,7 @@ internal sealed partial class LlamaCppManager : IAsyncDisposable
             _logger.LogWarning(
                 "llama.cpp container {ContainerName} is no longer running. Invalidating cached port for {Model}.",
                 containerName, ollamaModelName);
-            _modelPorts.TryRemove(ollamaModelName, out _);
+            RemoveModelPort(ollamaModelName);
         }
 
         return running;
@@ -104,11 +154,11 @@ internal sealed partial class LlamaCppManager : IAsyncDisposable
 
     public bool RemoveModelMapping(string ollamaModelName)
     {
-        if (_modelPorts.TryRemove(ollamaModelName, out var port))
+        if (RemoveModelPort(ollamaModelName))
         {
             _logger.LogWarning(
-                "Removed stale llama.cpp port mapping for {Model} (port {Port}).",
-                ollamaModelName, port);
+                "Removed stale llama.cpp port mapping for {Model}.",
+                ollamaModelName);
             return true;
         }
 
@@ -134,21 +184,52 @@ internal sealed partial class LlamaCppManager : IAsyncDisposable
             return false;
         }
 
+        var startLock = _modelStartLocks.GetOrAdd(ollamaName, static _ => new SemaphoreSlim(1, 1));
+        await startLock.WaitAsync(cancellationToken);
+        try
+        {
+            return await StartModelContainerCoreAsync(model, cancellationToken);
+        }
+        finally
+        {
+            startLock.Release();
+        }
+    }
+
+    private async Task<bool> StartModelContainerCoreAsync(LlamaCppModel model, CancellationToken cancellationToken)
+    {
+        var ollamaName = model.OllamaName;
+        if (string.IsNullOrWhiteSpace(ollamaName))
+        {
+            return false;
+        }
+
         if (_modelPorts.ContainsKey(ollamaName))
         {
             _logger.LogInformation("Model {Model} already has a running container", ollamaName);
             return true;
         }
 
+        if (IsModelOnFallback(ollamaName))
+        {
+            _logger.LogInformation(
+                "Model {Model} previously failed to load via llama.cpp. Skipping container start.",
+                ollamaName);
+            return false;
+        }
+
         if (!File.Exists(model.BlobPath))
         {
             _logger.LogError("Model blob not found: {BlobPath}", model.BlobPath);
+            MarkModelAsFallback(ollamaName);
             return false;
         }
 
         var port = FindAvailablePort();
         var containerName = SanitizeContainerName($"ngino-llamacpp-{ollamaName}");
 
+        try
+        {
             var existingPort = await FindExistingContainerPortAsync(containerName);
             if (existingPort.HasValue)
             {
@@ -165,14 +246,13 @@ internal sealed partial class LlamaCppManager : IAsyncDisposable
                 "Starting llama.cpp container for {Model} on port {Port}: docker {Args}",
                 ollamaName, port, string.Join(" ", args));
 
-        try
-        {
             var (exitCode, output) = await RunDockerWithOutputAsync(args, cancellationToken);
             if (exitCode != 0)
             {
                 _logger.LogError(
                     "Failed to start llama.cpp container for {Model}, exit code: {ExitCode}, output: {Output}",
                     ollamaName, exitCode, output);
+                MarkModelAsFallback(ollamaName);
                 return false;
             }
 
@@ -180,12 +260,21 @@ internal sealed partial class LlamaCppManager : IAsyncDisposable
                 "llama.cpp container for {Model} started on port {Port}. Waiting for it to become ready...",
                 ollamaName, port);
 
-            var ready = await WaitForServerReadyAsync("localhost", port, cancellationToken);
-            if (!ready)
+            var result = await WaitForServerReadyAsync("localhost", port, containerName, cancellationToken);
+            if (result != ContainerStartResult.Ready)
             {
+                if (result == ContainerStartResult.ContainerExited)
+                {
+                    MarkModelAsPermanentFallback(ollamaName);
+                }
+                else
+                {
+                    MarkModelAsFallback(ollamaName);
+                }
+
                 _logger.LogError(
-                    "llama.cpp container for {Model} did not become ready on port {Port} within {Timeout}. Stopping it.",
-                    ollamaName, port, ContainerStartTimeout);
+                    "llama.cpp container for {Model} did not become ready on port {Port} within {Timeout} ({Result}). Stopping it.",
+                    ollamaName, port, ContainerStartTimeout, result);
 
                 try
                 {
@@ -207,13 +296,18 @@ internal sealed partial class LlamaCppManager : IAsyncDisposable
         catch (Exception ex)
         {
             _logger.LogError(ex, "Failed to start llama.cpp container for {Model}", ollamaName);
+            MarkModelAsFallback(ollamaName);
             return false;
         }
+        finally
+        {
+            ReleaseReservedPort(port);
+        }
     }
 
     public async Task<bool> StopModelContainerAsync(string ollamaModelName, CancellationToken cancellationToken)
     {
-        if (!_modelPorts.TryRemove(ollamaModelName, out _))
+        if (!RemoveModelPort(ollamaModelName))
         {
             _logger.LogWarning("No running container found for model {Model}", ollamaModelName);
             return false;
@@ -227,6 +321,7 @@ internal sealed partial class LlamaCppManager : IAsyncDisposable
         {
             await RunDockerAsync(["stop", "--time", "10", containerName], cancellationToken);
             await RunDockerAsync(["rm", "-f", containerName], cancellationToken);
+            ClearModelFallback(ollamaModelName);
             return true;
         }
         catch (Exception ex)
@@ -262,6 +357,8 @@ internal sealed partial class LlamaCppManager : IAsyncDisposable
         }
 
         _modelPorts.Clear();
+        _reservedPorts.Clear();
+        _fallbackModels.Clear();
     }
 
     public async Task<bool> TestDockerAsync()
@@ -389,7 +486,6 @@ internal sealed partial class LlamaCppManager : IAsyncDisposable
         {
             "run",
             "-d",
-            "--rm",
             "--label", $"{NginoContainerLabel}=true",
             "--name", containerName,
             "-p", $"{port}:{port}",
@@ -413,9 +509,12 @@ internal sealed partial class LlamaCppManager : IAsyncDisposable
         args.Add("-m");
         args.Add($"/models/blobs/{blobFile}");
         args.Add("-ngl");
-        args.Add("999");
+        args.Add("auto");
+        if (_parallel.HasValue)
+        {
             args.Add("--parallel");
-        args.Add("4");
+            args.Add(_parallel.Value.ToString());
+        }
         args.Add("--host");
         args.Add("0.0.0.0");
         args.Add("--port");
@@ -475,8 +574,15 @@ internal sealed partial class LlamaCppManager : IAsyncDisposable
     }
 
     private int FindAvailablePort()
+    {
+        lock (_portAllocationLock)
         {
             var usedPorts = new HashSet<int>(_modelPorts.Values);
+            foreach (var reservedPort in _reservedPorts.Keys)
+            {
+                usedPorts.Add(reservedPort);
+            }
+
             var port = _basePort;
 
             while (usedPorts.Contains(port))
@@ -484,17 +590,36 @@ internal sealed partial class LlamaCppManager : IAsyncDisposable
                 port++;
             }
 
+            _reservedPorts[port] = 0;
             return port;
         }
+    }
 
-    private static async Task<bool> WaitForServerReadyAsync(
-        string host, int port, CancellationToken cancellationToken)
+    private void ReleaseReservedPort(int port)
+    {
+        _reservedPorts.TryRemove(port, out _);
+    }
+
+    private bool RemoveModelPort(string ollamaModelName)
+    {
+        if (_modelPorts.TryRemove(ollamaModelName, out var port))
+        {
+            ReleaseReservedPort(port);
+            return true;
+        }
+
+        return false;
+    }
+
+    private async Task<ContainerStartResult> WaitForServerReadyAsync(
+        string host, int port, string containerName, CancellationToken cancellationToken)
     {
         var deadline = DateTime.UtcNow + ContainerStartTimeout;
 
-        if (!await WaitForTcpPortAsync(host, port, deadline, cancellationToken))
+        var tcpResult = await WaitForTcpPortAsync(host, port, containerName, deadline, cancellationToken);
+        if (tcpResult != ContainerStartResult.Ready)
         {
-            return false;
+            return tcpResult;
         }
 
         using var handler = new SocketsHttpHandler
@@ -508,41 +633,57 @@ internal sealed partial class LlamaCppManager : IAsyncDisposable
         {
             cancellationToken.ThrowIfCancellationRequested();
 
+            if (!await IsDockerContainerRunningAsync(containerName))
+            {
+                await LogContainerOutputAsync(containerName);
+                return ContainerStartResult.ContainerExited;
+            }
+
             try
             {
                 using var response = await httpClient.GetAsync(
                     $"http://{host}:{port}/health", cancellationToken);
                 if (response.IsSuccessStatusCode)
                 {
-                    return true;
+                    return ContainerStartResult.Ready;
                 }
             }
             catch (OperationCanceledException) when (!cancellationToken.IsCancellationRequested)
             {
             }
-            catch (Exception)
+            catch (HttpRequestException ex)
             {
-                return false;
+                _logger.LogDebug(ex, "Health probe of llama.cpp container {ContainerName} failed; retrying.", containerName);
+            }
+            catch (IOException ex)
+            {
+                _logger.LogDebug(ex, "Health probe of llama.cpp container {ContainerName} failed; retrying.", containerName);
             }
 
             await Task.Delay(TimeSpan.FromSeconds(2), cancellationToken);
         }
 
-        return false;
+        return ContainerStartResult.TimedOut;
     }
 
-    private static async Task<bool> WaitForTcpPortAsync(
-        string host, int port, DateTime deadline, CancellationToken cancellationToken)
+    private async Task<ContainerStartResult> WaitForTcpPortAsync(
+        string host, int port, string containerName, DateTime deadline, CancellationToken cancellationToken)
     {
         while (DateTime.UtcNow < deadline)
         {
             cancellationToken.ThrowIfCancellationRequested();
 
+            if (!await IsDockerContainerRunningAsync(containerName))
+            {
+                await LogContainerOutputAsync(containerName);
+                return ContainerStartResult.ContainerExited;
+            }
+
             try
             {
                 using var client = new TcpClient();
                 await client.ConnectAsync(host, port, cancellationToken);
-                return true;
+                return ContainerStartResult.Ready;
             }
             catch (OperationCanceledException)
             {
@@ -555,7 +696,49 @@ internal sealed partial class LlamaCppManager : IAsyncDisposable
             await Task.Delay(TimeSpan.FromSeconds(2), cancellationToken);
         }
 
-        return false;
+        return ContainerStartResult.TimedOut;
+    }
+
+    private async Task<bool> IsDockerContainerRunningAsync(string containerName)
+    {
+        try
+        {
+            var (exitCode, output) = await RunDockerWithOutputAsync(
+                ["inspect", "-f", "{{.State.Running}}", containerName],
+                CancellationToken.None);
+            return exitCode == 0 && string.Equals(output.Trim(), "true", StringComparison.OrdinalIgnoreCase);
+        }
+        catch
+        {
+            return true;
+        }
+    }
+
+    private async Task LogContainerOutputAsync(string containerName)
+    {
+        try
+        {
+            var (_, output) = await RunDockerWithOutputAsync(
+                ["logs", "--tail", "100", containerName],
+                CancellationToken.None);
+
+            if (!string.IsNullOrWhiteSpace(output))
+            {
+                _logger.LogError(
+                    "llama.cpp container {ContainerName} exited before becoming ready. Last output:\n{Output}",
+                    containerName, output);
+            }
+            else
+            {
+                _logger.LogError(
+                    "llama.cpp container {ContainerName} exited before becoming ready, but produced no output.",
+                    containerName);
+            }
+        }
+        catch (Exception ex)
+        {
+            _logger.LogWarning(ex, "Failed to read logs of container {ContainerName}", containerName);
+        }
     }
 
     private static string SanitizeContainerName(string name)
@@ -655,6 +838,13 @@ internal sealed partial class LlamaCppManager : IAsyncDisposable
 
     [GeneratedRegex(@"[^a-zA-Z0-9_.-]")]
     private static partial Regex InvalidContainerNameChars();
+
+    private enum ContainerStartResult
+    {
+        Ready,
+        ContainerExited,
+        TimedOut
+    }
 }
 
 internal sealed record LlamaCppModel

src/Ngino.Client/Program.cs

@@ -22,8 +22,15 @@ try
         Console.WriteLine($"  ollama models path: {options.UseOllamaModelsPath ?? "(not set)"}");
         Console.WriteLine($"  llama.cpp docker image: {options.LlamaCppDockerImage ?? "(auto)"}");
         Console.WriteLine($"  llama.cpp base port: {options.LlamaCppBasePort}");
+        Console.WriteLine(
+            options.LlamaCppParallel.HasValue
+                ? $"  llama.cpp parallel slots: {options.LlamaCppParallel.Value}"
+                : "  llama.cpp parallel slots: (llama.cpp default)");
     }
 
+    var logDirectory = options.LogDirectory ?? Path.Combine(AppContext.BaseDirectory, "Logs");
+    Console.WriteLine($"  log directory: {logDirectory}");
+
     // Args are parsed by ClientOptions; keep them away from the host configuration.
     var builder = Host.CreateApplicationBuilder(new HostApplicationBuilderSettings { Args = [] });
     builder.Services.AddSingleton(options);
@@ -32,6 +39,7 @@ try
     builder.Services.AddWindowsService(service => service.ServiceName = "NginoClient");
     // The EventLog provider defaults to Warning; connection state is worth seeing there.
     builder.Logging.AddFilter<Microsoft.Extensions.Logging.EventLog.EventLogLoggerProvider>("Ngino.Client", LogLevel.Information);
+    builder.Logging.AddProvider(new FileLoggerProvider(logDirectory));
 
     await builder.Build().RunAsync();
     return 0;

src/Ngino.Client/TunnelClient.cs

@@ -47,7 +47,9 @@ internal sealed class TunnelClient
                 _options.UseOllamaModelsPath,
                 _options.LlamaCppDockerImage,
                 _options.LlamaCppBasePort,
-                _logger);
+                _logger,
+                _options.LlamaCppFallbackCooldown,
+                _options.LlamaCppParallel);
         }
     }
 
@@ -477,13 +479,16 @@ internal sealed class TunnelClient
                 var started = await _llamaCppManager.StartModelContainerAsync(model, cancellationToken);
                 if (!started)
                 {
-                    return new TunnelMessage
-                    {
-                        Type = TunnelMessageTypes.ModelCommandResult,
-                        RequestId = message.RequestId,
-                        StatusCode = 500,
-                        Error = $"Unable to load model '{modelName}'."
-                    };
+                    _logger.LogWarning(
+                        "Unable to load model '{Model}' via llama.cpp. Falling back to Ollama upstream.",
+                        modelName);
+
+                    using var fallbackRequest = BuildModelCommandRequest(_options.Upstream, "load", modelName);
+                    using var fallbackResponse = await _httpClient.SendAsync(
+                        fallbackRequest, HttpCompletionOption.ResponseHeadersRead, cancellationToken);
+                    var fallbackBody = await fallbackResponse.Content.ReadAsByteArrayAsync(cancellationToken);
+
+                    return BuildModelCommandResult(message.RequestId, fallbackResponse, fallbackBody);
                 }
 
                 return BuildModelCommandResult(message.RequestId, 200, "OK", []);
@@ -494,13 +499,18 @@ internal sealed class TunnelClient
                 var stopped = await _llamaCppManager!.StopModelContainerAsync(modelName!, cancellationToken);
                 if (!stopped)
                 {
-                    return new TunnelMessage
-                    {
-                        Type = TunnelMessageTypes.ModelCommandResult,
-                        RequestId = message.RequestId,
-                        StatusCode = 404,
-                        Error = $"No running llama.cpp container for model '{modelName}'."
-                    };
+                    _logger.LogWarning(
+                        "No running llama.cpp container for model '{Model}'. Falling back to Ollama upstream to unload it.",
+                        modelName);
+
+                    _llamaCppManager.ClearModelFallback(modelName!);
+
+                    using var fallbackRequest = BuildModelCommandRequest(_options.Upstream, "unload", modelName);
+                    using var fallbackResponse = await _httpClient.SendAsync(
+                        fallbackRequest, HttpCompletionOption.ResponseHeadersRead, cancellationToken);
+                    var fallbackBody = await fallbackResponse.Content.ReadAsByteArrayAsync(cancellationToken);
+
+                    return BuildModelCommandResult(message.RequestId, fallbackResponse, fallbackBody);
                 }
 
                 return BuildModelCommandResult(message.RequestId, 200, "OK", []);
@@ -712,6 +722,14 @@ internal sealed class TunnelClient
                         .FirstOrDefault(m => string.Equals(m.OllamaName, modelName, StringComparison.OrdinalIgnoreCase));
 
                     if (model is not null)
+                    {
+                        if (_llamaCppManager.IsModelOnFallback(modelName))
+                        {
+                            _logger.LogInformation(
+                                "Model '{Model}' previously failed to load via llama.cpp. Routing directly to the Ollama upstream.",
+                                modelName);
+                        }
+                        else
                         {
                             _logger.LogInformation(
                                 "Request for model '{Model}' but no llama.cpp container is running. Starting one on demand...",
@@ -724,8 +742,10 @@ internal sealed class TunnelClient
                             }
                             else
                             {
-                            await SendModelLoadErrorAsync(socket, message, modelName, cancellationToken);
-                            return;
+                                _logger.LogWarning(
+                                    "Unable to load model '{Model}' via llama.cpp. Falling back to the Ollama upstream.",
+                                    modelName);
+                            }
                         }
                     }
                     else
@@ -830,50 +850,6 @@ internal sealed class TunnelClient
         }
     }
 
-    private async Task SendModelLoadErrorAsync(
-        ClientWebSocket socket, TunnelMessage message, string modelName, CancellationToken cancellationToken)
-    {
-        _logger.LogWarning(
-            "Unable to load model '{Model}' via llama.cpp. Notifying caller.", modelName);
-
-        var body = JsonSerializer.SerializeToUtf8Bytes(
-            new { error = $"Unable to load model '{modelName}'" },
-            JsonOptions);
-
-        _pendingRequestBodies.TryRemove(message.RequestId, out _);
-
-        await SendAsync(
-            socket,
-            new TunnelMessage
-            {
-                Type = TunnelMessageTypes.HttpResponseHeaders,
-                RequestId = message.RequestId,
-                StatusCode = 500,
-                ReasonPhrase = "Internal Server Error",
-                Headers = [new HeaderPair("Content-Type", "application/json")]
-            },
-            CancellationToken.None);
-
-        await SendAsync(
-            socket,
-            new TunnelMessage
-            {
-                Type = TunnelMessageTypes.HttpResponseBody,
-                RequestId = message.RequestId,
-                Body = body
-            },
-            CancellationToken.None);
-
-        await SendAsync(
-            socket,
-            new TunnelMessage
-            {
-                Type = TunnelMessageTypes.HttpResponseComplete,
-                RequestId = message.RequestId
-            },
-            CancellationToken.None);
-    }
-
     private void CancelAllActiveRequests()
     {
         foreach (var pair in _activeRequests.ToArray())