fix(server): replaces text logo with proper logo

docs: adds images

README.md

@@ -1,8 +1,10 @@
-# Ngino
+# <img src="docs/images/Ngino_logo_symbol.png" width="25" alt="logo symbol"/> Ngino
 
-Ngino is a small outbound HTTP tunnel for running Ollama (or vLLM, etc.) on GPU workstations while exposing the API from a server that cannot reach those workstations directly.
+Ngino is a small outbound HTTP tunnel for running Ollama, llama.cpp, vLLM, etc. on GPU workstations while exposing the API from a server that cannot reach those workstations directly.
 
-The client opens and maintains a WebSocket connection to the server. The server accepts normal HTTP requests and forwards them through that WebSocket to the client. The client then calls a local upstream such as `http://localhost:11434` and streams the response back.
+<img src="docs/images/Ngino_logo_full.png" alt="logo symbol"/>
+
+The client opens and maintains a WebSocket connection to the server. The server accepts normal HTTP requests from users and forwards them through that WebSocket to the client. The client then calls a local upstream such as `http://localhost:11434` and streams the response back.
 
 <table>
 <tr>
@@ -15,8 +17,8 @@ The server provides
 - An API with
   - Authentication via user keys
   - Authorization (planned)
-- Load balancing (Scale your AI strategy horizontally!)
+- Load balancing (Scale your AI inferencing horizontally by adding more nodes!)
-- (Ollama-only) Model management (install, remove, load, unload models)
+- (Ollama-based) GGUF Model management (install, remove, load, unload models)
 - Client monitoring
   - Who is active
   - What models are running
@@ -129,9 +131,35 @@ Linux:
 ```bash
 sudo bash deploy/install-client.sh --server http://your-server:5050 --token "change-me"
 ```
+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   --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   --llama-cpp-parallel 128
+```
 
 Options: `--server`, `--token` (required); `--client-id`, `--upstream`, `--install-dir`, `--service-name`, `--no-ollama` (optional). Missing required values are prompted interactively.
 
+llama.cpp via Docker options (replaces Ollama for inferencing):
+
+| Option | Description |
+|--------|-------------|
+| `--use-llama-cpp-via-docker` | Use llama.cpp Docker containers instead of Ollama |
+| `--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`.
 
 ## Notes

deploy/install-client.ps1

@@ -14,7 +14,11 @@ param(
     [string]$InstallDir = "$env:ProgramFiles\Ngino Client",
     [string]$ServiceName = "NginoClient",
     [switch]$InsecureSkipTlsVerify,
-    [switch]$NoOllama
+    [switch]$NoOllama,
+    [switch]$UseLlamaCppViaDocker,
+    [string]$UseOllamaModelsPath = "",
+    [string]$LlamaCppDockerImage = "",
+    [int]$LlamaCppBasePort = 0
 )
 
 $ErrorActionPreference = "Stop"
@@ -54,6 +58,14 @@ if ([string]::IsNullOrWhiteSpace($Token)) {
     finally { [Runtime.InteropServices.Marshal]::ZeroFreeBSTR($tokenPointer) }
 }
 if ([string]::IsNullOrWhiteSpace($Token)) { throw "Token is required." }
+
+if ($UseLlamaCppViaDocker -and [string]::IsNullOrWhiteSpace($UseOllamaModelsPath)) {
+    $UseOllamaModelsPath = Read-Host "Ollama models path (e.g. C:\Users\user\.ollama\models)"
+}
+if ($UseLlamaCppViaDocker -and [string]::IsNullOrWhiteSpace($UseOllamaModelsPath)) {
+    throw "Ollama models path is required with -UseLlamaCppViaDocker."
+}
+
 if ($ServiceName -notmatch '^[A-Za-z0-9_.-]+$') { throw "ServiceName contains unsupported characters." }
 
 $scriptDir = $PSScriptRoot
@@ -149,6 +161,18 @@ $serviceEnvironment = @(
     "DOTNET_CLI_TELEMETRY_OPTOUT=1",
     "DOTNET_NOLOGO=1"
 )
+if ($UseLlamaCppViaDocker) {
+    $serviceEnvironment += "NGINO_USE_LLAMA_CPP_VIA_DOCKER=true"
+    if (-not [string]::IsNullOrWhiteSpace($UseOllamaModelsPath)) {
+        $serviceEnvironment += "NGINO_USE_OLLAMA_MODELS_PATH=$UseOllamaModelsPath"
+    }
+    if (-not [string]::IsNullOrWhiteSpace($LlamaCppDockerImage)) {
+        $serviceEnvironment += "NGINO_LLAMA_CPP_DOCKER_IMAGE=$LlamaCppDockerImage"
+    }
+    if ($LlamaCppBasePort -gt 0) {
+        $serviceEnvironment += "NGINO_LLAMA_CPP_BASE_PORT=$LlamaCppBasePort"
+    }
+}
 if ($InsecureSkipTlsVerify) {
     Write-Warn "Server TLS certificate validation is disabled for $ServiceName."
 }
@@ -169,6 +193,9 @@ Write-Host "  Client ID:   $ClientId"
 Write-Host "  Upstream:    $Upstream"
 Write-Host "  Service:     $ServiceName"
 Write-Host "  Install dir: $InstallDir"
+if ($UseLlamaCppViaDocker) {
+    Write-Host "  llama.cpp:   enabled (models: $UseOllamaModelsPath)"
+}
 Write-Host ""
 Write-Host "  Manage:  Get-Service $ServiceName | Start-Service/Stop-Service/Restart-Service"
 Write-Host "  Logs:    Get-WinEvent -LogName Application | Where-Object ProviderName -eq NginoClient"

deploy/install-client.sh

@@ -16,6 +16,13 @@ TOKEN=""
 CLIENT_ID="$(hostname -s 2>/dev/null || echo "linux-client")"
 UPSTREAM="$DEFAULT_UPSTREAM"
 SKIP_OLLAMA=false
+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'
@@ -45,11 +52,27 @@ Optional:
   --install-dir <dir>  Install directory; defaults to $DEFAULT_INSTALL_DIR
   --service-name <n>   systemd service name; defaults to $DEFAULT_SERVICE_NAME
   --no-ollama          Skip Ollama installation and status check
+  --use-llama-cpp-via-docker
+                       Use llama.cpp via Docker for inference instead of Ollama
+  --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 <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 \\
+     --llama-cpp-parallel 128
 EOF
     exit 0
 }
@@ -57,15 +80,22 @@ EOF
 # ── Argument parsing ──────────────────────────────────────────────────────────
 while [[ $# -gt 0 ]]; do
     case "$1" in
-        --server)       SERVER_URL="$2"; shift 2 ;;
+        --server)                   SERVER_URL="$2"; shift 2 ;;
-        --token)        TOKEN="$2"; shift 2 ;;
+        --token)                    TOKEN="$2"; shift 2 ;;
-        --client-id)    CLIENT_ID="$2"; shift 2 ;;
+        --client-id)                CLIENT_ID="$2"; shift 2 ;;
-        --upstream)     UPSTREAM="$2"; shift 2 ;;
+        --upstream)                 UPSTREAM="$2"; shift 2 ;;
-        --install-dir)  INSTALL_DIR="$2"; shift 2 ;;
+        --install-dir)              INSTALL_DIR="$2"; shift 2 ;;
-        --service-name) SERVICE_NAME="$2"; shift 2 ;;
+        --service-name)             SERVICE_NAME="$2"; shift 2 ;;
-        --no-ollama)    SKIP_OLLAMA=true; shift ;;
+        --no-ollama)                SKIP_OLLAMA=true; shift ;;
-        -h|--help)      usage ;;
+        --use-llama-cpp-via-docker) USE_LLAMA_CPP_VIA_DOCKER=true; shift ;;
-        *)              die "Unknown option: $1" ;;
+        --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
 done
 
@@ -85,6 +115,13 @@ if [[ -z "$TOKEN" ]]; then
     die "Token is required."
 fi
 
+if [[ "$USE_LLAMA_CPP_VIA_DOCKER" == "true" && -z "$USE_OLLAMA_MODELS_PATH" ]]; then
+    read -rp "Ollama models path (e.g. /usr/share/ollama/.ollama/models): " USE_OLLAMA_MODELS_PATH
+fi
+if [[ "$USE_LLAMA_CPP_VIA_DOCKER" == "true" && -z "$USE_OLLAMA_MODELS_PATH" ]]; then
+    die "Ollama models path is required with --use-llama-cpp-via-docker."
+fi
+
 # ── Root check ────────────────────────────────────────────────────────────────
 if [[ $EUID -ne 0 ]]; then
     die "This script must be run as root (or with sudo)."
@@ -254,6 +291,14 @@ fi
 
 info "Build successful."
 
+# ── Stop existing service before overwriting binary ─────────────────────────
+SERVICE_FILE="/etc/systemd/system/${SERVICE_NAME}.service"
+
+if systemctl list-unit-files "$SERVICE_NAME.service" &>/dev/null 2>&1 || systemctl is-active --quiet "$SERVICE_NAME" 2>/dev/null; then
+    info "Stopping existing service $SERVICE_NAME..."
+    systemctl stop "$SERVICE_NAME" 2>/dev/null || true
+fi
+
 # ── Install ───────────────────────────────────────────────────────────────────
 info "Installing to $INSTALL_DIR..."
 mkdir -p "$INSTALL_DIR"
@@ -265,23 +310,41 @@ info "Client installed to $INSTALL_DIR."
 # ── Write environment file (avoids shell injection in unit file) ─────────────
 ENV_DIR="/etc/ngino-client"
 mkdir -p "$ENV_DIR"
-printf 'NGINO_TOKEN=%s\n' "$TOKEN" > "$ENV_DIR/env"
+{
+    printf 'NGINO_TOKEN=%s\n' "$TOKEN"
+    if [[ "$USE_LLAMA_CPP_VIA_DOCKER" == "true" ]]; then
+        printf 'NGINO_USE_LLAMA_CPP_VIA_DOCKER=true\n'
+        if [[ -n "$USE_OLLAMA_MODELS_PATH" ]]; then
+            printf 'NGINO_USE_OLLAMA_MODELS_PATH=%s\n' "$USE_OLLAMA_MODELS_PATH"
+        fi
+        if [[ -n "$LLAMA_CPP_DOCKER_IMAGE" ]]; then
+            printf 'NGINO_LLAMA_CPP_DOCKER_IMAGE=%s\n' "$LLAMA_CPP_DOCKER_IMAGE"
+        fi
+        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"
 info "Environment file written to $ENV_DIR/env (mode 0600)."
 
 # ── Create systemd service ───────────────────────────────────────────────────
 SERVICE_FILE="/etc/systemd/system/${SERVICE_NAME}.service"
 
-if systemctl list-unit-files "$SERVICE_NAME.service" &>/dev/null 2>&1; then
-    info "Stopping existing service $SERVICE_NAME..."
-    systemctl stop "$SERVICE_NAME" 2>/dev/null || true
-fi
-
 cat > "$SERVICE_FILE" <<EOF
 [Unit]
 Description=Ngino Tunnel Client
-After=network-online.target
+After=network-online.target docker.service
-Wants=network-online.target
+Wants=network-online.target docker.service
 $([ "$SKIP_OLLAMA" = "false" ] && echo "After=ollama.service")
 $([ "$SKIP_OLLAMA" = "false" ] && echo "Wants=ollama.service")
 
@@ -318,6 +381,9 @@ echo "  Client ID:   $CLIENT_ID"
 echo "  Upstream:    $UPSTREAM"
 echo "  Service:     $SERVICE_NAME"
 echo "  Install dir: $INSTALL_DIR"
+if [[ "$USE_LLAMA_CPP_VIA_DOCKER" == "true" ]]; then
+    echo "  llama.cpp:   enabled (models: $USE_OLLAMA_MODELS_PATH)"
+fi
 echo
 echo "  Manage:  systemctl {start|stop|restart|status} $SERVICE_NAME"
 echo "  Logs:    journalctl -u $SERVICE_NAME -f"

deploy/uninstall-client.ps1

@@ -0,0 +1,42 @@
+#Requires -Version 5.1
+#Requires -RunAsAdministrator
+
+[CmdletBinding()]
+param(
+    [string]$InstallDir = "$env:ProgramFiles\Ngino Client",
+    [string]$ServiceName = "NginoClient"
+)
+
+$ErrorActionPreference = "Stop"
+
+function Write-Info([string]$Message) { Write-Host "[INFO]  $Message" -ForegroundColor Green }
+
+if ($ServiceName -notmatch '^[A-Za-z0-9_.-]+$') { throw "ServiceName contains unsupported characters." }
+
+# ── Stop and delete the Windows service ───────────────────────────────────────
+$service = Get-Service -Name $ServiceName -ErrorAction SilentlyContinue
+if ($service) {
+    if ($service.Status -ne "Stopped") {
+        Write-Info "Stopping service $ServiceName..."
+        Stop-Service -Name $ServiceName -Force
+        $service.WaitForStatus("Stopped", [TimeSpan]::FromSeconds(30))
+    }
+    Write-Info "Deleting service $ServiceName..."
+    & sc.exe delete $ServiceName
+    if ($LASTEXITCODE -ne 0) { throw "Could not delete Windows service $ServiceName." }
+} else {
+    Write-Info "Service $ServiceName does not exist; skipping."
+}
+
+# ── Remove install directory ──────────────────────────────────────────────────
+if (Test-Path -LiteralPath $InstallDir) {
+    Write-Info "Removing install directory $InstallDir..."
+    Remove-Item -LiteralPath $InstallDir -Recurse -Force
+} else {
+    Write-Info "Install directory $InstallDir does not exist; skipping."
+}
+
+Write-Host ""
+Write-Info "Uninstall complete."
+Write-Host "  Service:     $ServiceName (stopped and deleted)"
+Write-Host "  Install dir: $InstallDir (removed)"

deploy/uninstall-client.sh

@@ -0,0 +1,95 @@
+#!/usr/bin/env bash
+set -euo pipefail
+
+# ── Defaults ──────────────────────────────────────────────────────────────────
+DEFAULT_INSTALL_DIR="/opt/ngino-client"
+DEFAULT_SERVICE_NAME="ngino-client"
+
+SCRIPT_DIR="$(cd "$(dirname "${BASH_SOURCE[0]}")" && pwd)"
+
+INSTALL_DIR="$DEFAULT_INSTALL_DIR"
+SERVICE_NAME="$DEFAULT_SERVICE_NAME"
+
+# ── Colors ────────────────────────────────────────────────────────────────────
+RED='\033[0;31m'
+GREEN='\033[0;32m'
+YELLOW='\033[0;33m'
+NC='\033[0m'
+
+info()    { echo -e "${GREEN}[INFO]${NC}  $*"; }
+warn()    { echo -e "${YELLOW}[WARN]${NC}  $*"; }
+err()     { echo -e "${RED}[ERROR]${NC} $*" >&2; }
+die()     { err "$@"; exit 1; }
+
+# ── Usage ─────────────────────────────────────────────────────────────────────
+usage() {
+    cat <<EOF
+Usage: $0 [OPTIONS]
+
+Uninstalls the Ngino client systemd service and removes its files.
+Does NOT remove .NET SDK, Ollama, or Docker.
+
+Optional:
+  --install-dir <dir>  Install directory; defaults to $DEFAULT_INSTALL_DIR
+  --service-name <n>   systemd service name; defaults to $DEFAULT_SERVICE_NAME
+  -h, --help           Show this help message
+
+Examples:
+  $0
+  $0 --install-dir /opt/ngino-client --service-name ngino-client
+EOF
+    exit 0
+}
+
+# ── Argument parsing ──────────────────────────────────────────────────────────
+while [[ $# -gt 0 ]]; do
+    case "$1" in
+        --install-dir)              INSTALL_DIR="$2"; shift 2 ;;
+        --service-name)             SERVICE_NAME="$2"; shift 2 ;;
+        -h|--help)                  usage ;;
+        *)                          die "Unknown option: $1" ;;
+    esac
+done
+
+# ── Root check ────────────────────────────────────────────────────────────────
+if [[ $EUID -ne 0 ]]; then
+    die "This script must be run as root (or with sudo)."
+fi
+
+SERVICE_FILE="/etc/systemd/system/${SERVICE_NAME}.service"
+ENV_DIR="/etc/ngino-client"
+
+# ── Stop and disable the service ──────────────────────────────────────────────
+if systemctl list-unit-files "$SERVICE_NAME.service" &>/dev/null 2>&1; then
+    info "Stopping service $SERVICE_NAME..."
+    systemctl stop "$SERVICE_NAME" 2>/dev/null || true
+    info "Disabling service $SERVICE_NAME..."
+    systemctl disable "$SERVICE_NAME" 2>/dev/null || true
+fi
+
+# ── Remove service unit file ──────────────────────────────────────────────────
+if [[ -f "$SERVICE_FILE" ]]; then
+    info "Removing service unit file $SERVICE_FILE..."
+    rm -f "$SERVICE_FILE"
+fi
+
+systemctl daemon-reload
+
+# ── Remove install directory ──────────────────────────────────────────────────
+if [[ -d "$INSTALL_DIR" ]]; then
+    info "Removing install directory $INSTALL_DIR..."
+    rm -rf "$INSTALL_DIR"
+fi
+
+# ── Remove environment file ───────────────────────────────────────────────────
+if [[ -d "$ENV_DIR" ]]; then
+    info "Removing environment directory $ENV_DIR..."
+    rm -rf "$ENV_DIR"
+fi
+
+# ── Done ──────────────────────────────────────────────────────────────────────
+echo
+info "Uninstall complete."
+echo "  Service:     $SERVICE_NAME (stopped, disabled, unit removed)"
+echo "  Install dir: $INSTALL_DIR (removed)"
+echo "  Env dir:     $ENV_DIR (removed)"

docs/images/Ngino_logo.svg

@@ -0,0 +1,332 @@
+<?xml version="1.0" encoding="UTF-8" standalone="no"?>
+<svg
+   viewBox="0 0 1280 720"
+   role="img"
+   aria-labelledby="title desc"
+   version="1.1"
+   id="svg10"
+   sodipodi:docname="Ngino_logo.svg"
+   xml:space="preserve"
+   inkscape:version="1.4.3 (0d15f75042, 2025-12-25)"
+   inkscape:export-filename="Ngino_logo_full_dark.png"
+   inkscape:export-xdpi="125.532"
+   inkscape:export-ydpi="125.532"
+   xmlns:inkscape="http://www.inkscape.org/namespaces/inkscape"
+   xmlns:sodipodi="http://sodipodi.sourceforge.net/DTD/sodipodi-0.dtd"
+   xmlns:xlink="http://www.w3.org/1999/xlink"
+   xmlns="http://www.w3.org/2000/svg"
+   xmlns:svg="http://www.w3.org/2000/svg"
+   xmlns:rdf="http://www.w3.org/1999/02/22-rdf-syntax-ns#"
+   xmlns:cc="http://creativecommons.org/ns#"
+   xmlns:dc="http://purl.org/dc/elements/1.1/"><sodipodi:namedview
+     id="namedview10"
+     pagecolor="#000000"
+     bordercolor="#000000"
+     borderopacity="1"
+     inkscape:showpageshadow="2"
+     inkscape:pageopacity="0.0"
+     inkscape:pagecheckerboard="0"
+     inkscape:deskcolor="#000000"
+     inkscape:zoom="0.7671875"
+     inkscape:cx="391.69043"
+     inkscape:cy="406.02851"
+     inkscape:window-width="2560"
+     inkscape:window-height="1412"
+     inkscape:window-x="0"
+     inkscape:window-y="0"
+     inkscape:window-maximized="1"
+     inkscape:current-layer="layer4"
+     inkscape:export-bgcolor="#00000000" /><title
+     id="title">Ngino logo</title><desc
+     id="desc">A fully vector llama outline with one input connection branching to three AI endpoints, beside the Ngino wordmark.</desc><defs
+     id="defs2"><linearGradient
+       id="swatch181"><stop
+         style="stop-color:#000000;stop-opacity:1;"
+         offset="0"
+         id="stop181" /></linearGradient><clipPath
+       clipPathUnits="userSpaceOnUse"
+       id="clipPath15"><rect
+         style="fill:none;stroke:#000000;stroke-linecap:butt;stroke-linejoin:bevel;paint-order:stroke markers fill;stop-color:#000000"
+         id="rect15-7"
+         width="45.034248"
+         height="38.955639"
+         x="262.48285"
+         y="121.00073" /></clipPath><linearGradient
+       id="swatch7"
+       inkscape:swatch="solid"><stop
+         style="stop-color:#000000;stop-opacity:1;"
+         offset="0"
+         id="stop7" /></linearGradient><linearGradient
+       id="linearGradient1"
+       inkscape:swatch="gradient"><stop
+         style="stop-color:#000000;stop-opacity:0;"
+         offset="0"
+         id="stop3" /><stop
+         style="stop-color:#000000;stop-opacity:0;"
+         offset="1"
+         id="stop4" /></linearGradient><linearGradient
+       id="routeGradient"
+       x1="95"
+       y1="0"
+       x2="430"
+       y2="0"
+       gradientUnits="userSpaceOnUse"><stop
+         offset="0"
+         stop-color="#0697b5"
+         id="stop1" /><stop
+         offset="1"
+         stop-color="#45d05f"
+         id="stop2" /></linearGradient><linearGradient
+       inkscape:collect="always"
+       xlink:href="#routeGradient"
+       id="linearGradient10"
+       gradientUnits="userSpaceOnUse"
+       x1="95"
+       y1="0"
+       x2="430"
+       y2="0"
+       gradientTransform="translate(50.835031)" /><linearGradient
+       inkscape:collect="always"
+       xlink:href="#routeGradient"
+       id="linearGradient24"
+       gradientUnits="userSpaceOnUse"
+       x1="95"
+       y1="0"
+       x2="430"
+       y2="0" /><linearGradient
+       inkscape:collect="always"
+       xlink:href="#routeGradient"
+       id="linearGradient25"
+       gradientUnits="userSpaceOnUse"
+       x1="95"
+       y1="0"
+       x2="430"
+       y2="0" /><linearGradient
+       inkscape:collect="always"
+       xlink:href="#routeGradient"
+       id="linearGradient26"
+       gradientUnits="userSpaceOnUse"
+       x1="95"
+       y1="0"
+       x2="430"
+       y2="0" /><linearGradient
+       inkscape:collect="always"
+       xlink:href="#routeGradient"
+       id="linearGradient27"
+       gradientUnits="userSpaceOnUse"
+       x1="95"
+       y1="0"
+       x2="430"
+       y2="0" /><linearGradient
+       inkscape:collect="always"
+       xlink:href="#routeGradient"
+       id="linearGradient28"
+       gradientUnits="userSpaceOnUse"
+       x1="95"
+       y1="0"
+       x2="430"
+       y2="0"
+       gradientTransform="translate(18.248473,-20.855397)" /><linearGradient
+       inkscape:collect="always"
+       xlink:href="#routeGradient"
+       id="linearGradient29"
+       gradientUnits="userSpaceOnUse"
+       x1="95"
+       y1="0"
+       x2="430"
+       y2="0"
+       gradientTransform="translate(18.248473,-20.855397)" /><linearGradient
+       inkscape:collect="always"
+       xlink:href="#routeGradient"
+       id="linearGradient30"
+       gradientUnits="userSpaceOnUse"
+       x1="95"
+       y1="0"
+       x2="430"
+       y2="0"
+       gradientTransform="translate(18.248473,-20.855397)" /><linearGradient
+       inkscape:collect="always"
+       xlink:href="#routeGradient"
+       id="linearGradient31"
+       gradientUnits="userSpaceOnUse"
+       x1="95"
+       y1="0"
+       x2="430"
+       y2="0"
+       gradientTransform="translate(18.248473,-20.855397)" /><filter
+       inkscape:collect="always"
+       style="color-interpolation-filters:sRGB"
+       id="filter215"
+       x="-0.0021624053"
+       y="-0.0047179751"
+       width="1.0043248"
+       height="1.009436"><feGaussianBlur
+         inkscape:collect="always"
+         stdDeviation="1.0812026"
+         id="feGaussianBlur215" /></filter><filter
+       inkscape:collect="always"
+       style="color-interpolation-filters:sRGB"
+       id="filter215-8"
+       x="-0.003844276"
+       y="-0.0074139609"
+       width="1.0076886"
+       height="1.0148279"><feGaussianBlur
+         inkscape:collect="always"
+         stdDeviation="1.0812026"
+         id="feGaussianBlur215-8" /></filter><filter
+       inkscape:collect="always"
+       style="color-interpolation-filters:sRGB"
+       id="filter215-6"
+       x="-0.0051897726"
+       y="-0.0047179751"
+       width="1.0103795"
+       height="1.009436"><feGaussianBlur
+         inkscape:collect="always"
+         stdDeviation="1.0812026"
+         id="feGaussianBlur215-0" /></filter></defs><!-- Transparent background; no embedded raster image. --><rect
+     style="display:inline;opacity:1;fill:#ffffff;fill-opacity:1;stroke:none;stroke-width:50;stroke-linecap:round;stroke-miterlimit:0.8;stroke-dasharray:none;stroke-opacity:1;filter:url(#filter215)"
+     id="rect1"
+     width="1200"
+     height="550"
+     x="26.069246"
+     y="20.855396"
+     inkscape:label="background_full"
+     rx="100"
+     inkscape:highlight-color="#000000"
+     inkscape:export-filename="Ngino_logo_full.png"
+     inkscape:export-xdpi="153.60001"
+     inkscape:export-ydpi="153.60001" /><g
+     inkscape:groupmode="layer"
+     id="layer1"
+     inkscape:label="text"
+     style="display:inline"
+     inkscape:export-filename="Ngino_logo_text.png"
+     inkscape:export-xdpi="125.532"
+     inkscape:export-ydpi="125.532"><rect
+       style="display:inline;fill:#ffffff;fill-opacity:1;stroke:none;stroke-width:50;stroke-linecap:round;stroke-miterlimit:0.8;stroke-dasharray:none;stroke-opacity:1;filter:url(#filter215-8)"
+       id="rect1-1"
+       width="675"
+       height="350"
+       x="458.96759"
+       y="152.75374"
+       inkscape:label="background_text"
+       rx="100"
+       inkscape:highlight-color="#000000"
+       inkscape:export-filename="Ngino_logo_full.png"
+       inkscape:export-xdpi="153.60001"
+       inkscape:export-ydpi="153.60001" /><text
+       xml:space="preserve"
+       style="font-style:normal;font-variant:normal;font-weight:normal;font-stretch:normal;font-size:37.3333px;line-height:1.25;font-family:Sans;-inkscape-font-specification:'Sans, Normal';font-variant-ligatures:normal;font-variant-caps:normal;font-variant-numeric:normal;font-variant-east-asian:normal;display:inline;fill:#000000;fill-opacity:1"
+       x="503.97342"
+       y="430.72964"
+       id="text2"
+       inkscape:label="sub"><tspan
+         sodipodi:role="line"
+         id="tspan2"
+         x="503.97342"
+         y="430.72964">R E V E R S E   P R O X Y   F O R  A I</tspan></text><text
+       xml:space="preserve"
+       style="font-weight:900;font-size:170.667px;line-height:1.25;font-family:Montserrat;-inkscape-font-specification:'Montserrat, Heavy';display:inline;fill:#000000;fill-opacity:1"
+       x="496.78931"
+       y="357.61456"
+       id="text1"
+       inkscape:label="main"
+       inkscape:export-filename="Ngino_logo_text.png"
+       inkscape:export-xdpi="125.532"
+       inkscape:export-ydpi="125.532"><tspan
+         sodipodi:role="line"
+         id="tspan1"
+         x="496.78931"
+         y="357.61456">Ngino</tspan></text></g><!-- Connection nodes --><!-- Wordmark --><rect
+     style="display:inline;fill:#ffffff;fill-opacity:1;stroke:none;stroke-width:50;stroke-linecap:round;stroke-miterlimit:0.8;stroke-dasharray:none;stroke-opacity:1;filter:url(#filter215-6)"
+     id="rect1-8"
+     width="500"
+     height="550"
+     x="28.825022"
+     y="21.104044"
+     inkscape:label="background_logo"
+     rx="100"
+     inkscape:highlight-color="#000000"
+     inkscape:export-filename="Ngino_logo_full.png"
+     inkscape:export-xdpi="153.60001"
+     inkscape:export-ydpi="153.60001" /><g
+     inkscape:groupmode="layer"
+     id="layer2"
+     inkscape:label="Logo"
+     style="display:inline"
+     inkscape:export-filename="Ngino_Logo_Symbol.png"
+     inkscape:export-xdpi="125.532"
+     inkscape:export-ydpi="125.532"><g
+       inkscape:groupmode="layer"
+       id="layer3"
+       inkscape:label="Body"
+       style="display:inline"><path
+         d="m 115.82078,398.64766 c -22.393081,-46.4277 -14.055,-69.1446 2.64154,-97.02036 C 143.97964,277.26884 147,275.30346 190,275.30346 h 56 c 36,0 53.48269,-18.60692 63.48269,-46.60692 6,-16 4,-43 5,-65 0,-20 6,-37.30346 16,-52.30346 l 0.39308,-26.965381 C 333.17923,71.427699 347,53 356,48 l 2,47 c 6,-19 17,-34 31,-40 l -1.69653,38.268839 c 13,3 22.69653,17.517311 28.69653,28.517311 18,3 29,6.17923 31,22.17923 0,14 -3.78615,18.08961 -16.78615,20.08961 l -37,7.21385 c 12,24 28.91039,76.76578 36.91039,108.76578 14,58 -6.24848,146.8554 -29.24848,202.8554 l -13,29 h -47 l -8,-35 c -4,-22 -23,-38 -48,-38 h -53.50305 c -27,0 -45,16 -49,40 l -6,33 h -47.12424 z"
+         id="path2"
+         sodipodi:nodetypes="ccsscccccccccccccccccsscccc"
+         style="display:inline;fill:none;stroke:url(#linearGradient27)"
+         stroke-width="13"
+         inkscape:label="body" /></g><g
+       inkscape:groupmode="layer"
+       id="layer4"
+       inkscape:label="Symbol"
+       style="display:inline"><g
+         fill="url(#routeGradient)"
+         id="g9"
+         style="display:inline;fill:url(#linearGradient27)"
+         transform="translate(18.248473,-20.855397)"
+         inkscape:label="Symbol_circles"><circle
+           cx="354"
+           cy="398"
+           r="15"
+           id="circle9"
+           style="fill:url(#linearGradient26)"
+           inkscape:label="ball_fork_down" /><circle
+           cx="354"
+           cy="356"
+           r="15"
+           id="circle8"
+           style="display:inline;fill:url(#linearGradient25)"
+           inkscape:label="ball_fork_mid" /><circle
+           cx="354"
+           cy="314"
+           r="15"
+           id="circle7"
+           style="display:inline;fill:url(#linearGradient24)"
+           inkscape:label="ball_fork_up" /><circle
+           cx="130.83504"
+           cy="356"
+           r="15"
+           id="circle6"
+           style="display:inline;fill:url(#linearGradient10)"
+           inkscape:label="ball_connector" /></g><g
+         fill="none"
+         stroke="url(#routeGradient)"
+         stroke-width="13"
+         stroke-linecap="round"
+         stroke-linejoin="round"
+         id="g6"
+         style="display:inline"
+         inkscape:label="Symbol_paths"><!-- Llama silhouette --><!-- One input, three routed endpoints --><path
+           d="m 254.24847,335.1446 c 29,0 42,42 78,42 h 40"
+           id="path6"
+           style="display:inline;stroke:url(#linearGradient31)"
+           inkscape:label="fork_down" /><path
+           d="m 254.24847,335.1446 h 118"
+           id="path5"
+           style="display:inline;stroke:url(#linearGradient28)"
+           inkscape:label="fork_mid" /><path
+           d="m 254.24847,335.1446 c 29,0 42,-42 78,-42 h 40"
+           id="path4"
+           style="display:inline;stroke:url(#linearGradient30)"
+           inkscape:label="fork_up" /><path
+           d="M 150.38696,335.1446 H 254.24847"
+           id="path3"
+           sodipodi:nodetypes="cc"
+           style="display:inline;stroke:url(#linearGradient29)"
+           inkscape:label="connector"
+           inkscape:export-filename="Ngino_logo_symbol.png"
+           inkscape:export-xdpi="153.60001"
+           inkscape:export-ydpi="153.60001" /></g></g></g><metadata
+     id="metadata215"><rdf:RDF><cc:Work
+         rdf:about=""><dc:title>Ngino logo</dc:title></cc:Work></rdf:RDF></metadata></svg>

src/Ngino.Client/ClientOptions.cs

@@ -20,6 +20,20 @@ internal sealed class ClientOptions
 
     public bool InsecureSkipTlsVerify { get; init; }
 
+    public bool UseLlamaCppViaDocker { get; init; }
+
+    public string? UseOllamaModelsPath { get; init; }
+
+    public string? LlamaCppDockerImage { get; init; }
+
+    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
@@ -56,21 +70,36 @@ internal sealed class ClientOptions
             ClientId = Read(values, "client-id", "NGINO_CLIENT_ID") ?? Environment.MachineName.ToLowerInvariant(),
             ReconnectDelay = TimeSpan.FromSeconds(ReadInt(values, 5, "reconnect-delay", "NGINO_RECONNECT_DELAY_SECONDS")),
             ChunkSize = ReadInt(values, 64 * 1024, "chunk-size", "NGINO_CHUNK_SIZE"),
-            InsecureSkipTlsVerify = ReadBool(values, false, "insecure-skip-tls-verify", "NGINO_INSECURE_SKIP_TLS_VERIFY")
+            InsecureSkipTlsVerify = ReadBool(values, false, "insecure-skip-tls-verify", "NGINO_INSECURE_SKIP_TLS_VERIFY"),
+            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"),
+            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"))
         };
     }
 
     public static string Usage =>
         """
         Ngino.Client options:
-          --server <url>             Server base URL, e.g. http://my-server:5050
+          --server <url>                 Server base URL, e.g. http://my-server:5050
-          --upstream <url>           Local upstream URL, e.g. http://localhost:11434
+          --upstream <url>               Local upstream URL, e.g. http://localhost:11434
-          --token <value>            Optional token matching the server
+          --token <value>                Optional token matching the server
-          --client-id <name>         Identifies this machine on the server; defaults to the machine name
+          --client-id <name>             Identifies this machine on the server; defaults to the machine name
-          --tunnel-path <path>       Defaults to /_ngino/tunnel
+          --tunnel-path <path>           Defaults to /_ngino/tunnel
-          --reconnect-delay <sec>    Defaults to 5
+          --reconnect-delay <sec>        Defaults to 5
-          --chunk-size <bytes>       Defaults to 65536
+          --chunk-size <bytes>           Defaults to 65536
-          --insecure-skip-tls-verify Disable server TLS certificate validation (unsafe)
+          --insecure-skip-tls-verify     Disable server TLS certificate validation (unsafe)
+          --use-llama-cpp-via-docker     Use llama.cpp via Docker for inference instead of Ollama
+          --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)
@@ -92,7 +121,7 @@ internal sealed class ClientOptions
                 continue;
             }
 
-            if (keyValue[0].Equals("insecure-skip-tls-verify", StringComparison.OrdinalIgnoreCase))
+            if (IsBoolFlag(keyValue[0]))
             {
                 values[keyValue[0]] = "true";
                 continue;
@@ -109,6 +138,16 @@ internal sealed class ClientOptions
         return values;
     }
 
+    private static bool IsBoolFlag(string key)
+    {
+        return key switch
+        {
+            "insecure-skip-tls-verify" => true,
+            "use-llama-cpp-via-docker" => true,
+            _ => false
+        };
+    }
+
     private static string? Read(Dictionary<string, string> values, params string[] keys)
     {
         foreach (var key in keys)
@@ -134,6 +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);
@@ -154,4 +199,14 @@ internal sealed class ClientOptions
 
     private static string NormalizePath(string path) =>
         path.StartsWith('/') ? path : $"/{path}";
+
+    private static string? NormalizeDirectoryPath(string? path)
+    {
+        if (string.IsNullOrWhiteSpace(path))
+        {
+            return null;
+        }
+
+        return Path.GetFullPath(path);
+    }
 }

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

@@ -0,0 +1,856 @@
+using System.Collections.Concurrent;
+using System.Diagnostics;
+using System.Net.Http;
+using System.Net.Sockets;
+using System.Text.Json;
+using System.Text.RegularExpressions;
+using Microsoft.Extensions.Logging;
+using Microsoft.Extensions.Logging.Abstractions;
+
+namespace Ngino.Client;
+
+internal sealed partial class LlamaCppManager : IAsyncDisposable
+{
+    private static readonly JsonSerializerOptions JsonOptions = new(JsonSerializerDefaults.Web);
+    private const int DefaultBasePort = 8081;
+    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,
+        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;
+    }
+
+    public string DockerImage => _dockerImage;
+
+    public List<LlamaCppModel> DiscoverModels()
+    {
+        var models = new List<LlamaCppModel>();
+
+        if (!Directory.Exists(_manifestsPath))
+        {
+            _logger.LogWarning("Ollama manifests path not found: {ManifestsPath}", _manifestsPath);
+            return models;
+        }
+
+        foreach (var manifestPath in Directory.EnumerateFiles(_manifestsPath, "*", SearchOption.AllDirectories))
+        {
+            try
+            {
+                var model = ParseManifest(manifestPath);
+                if (model is not null)
+                {
+                    models.Add(model);
+                }
+            }
+            catch (Exception ex)
+            {
+                _logger.LogDebug(ex, "Failed to parse manifest: {ManifestPath}", manifestPath);
+            }
+        }
+
+        return models;
+    }
+
+    public List<LlamaCppModel> DiscoverModelsWithBlob()
+    {
+        return DiscoverModels().Where(m => File.Exists(m.BlobPath)).ToList();
+    }
+
+    public bool IsModelActive(string ollamaModelName)
+    {
+        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))
+        {
+            return false;
+        }
+
+        var containerName = SanitizeContainerName($"ngino-llamacpp-{ollamaModelName}");
+        var (exitCode, output) = await RunDockerWithOutputAsync(
+            ["ps", "--filter", $"name=^{containerName}$", "--format", "{{.ID}}"],
+            CancellationToken.None);
+
+        var running = exitCode == 0 && !string.IsNullOrWhiteSpace(output);
+        if (!running)
+        {
+            _logger.LogWarning(
+                "llama.cpp container {ContainerName} is no longer running. Invalidating cached port for {Model}.",
+                containerName, ollamaModelName);
+            RemoveModelPort(ollamaModelName);
+        }
+
+        return running;
+    }
+
+    public bool RemoveModelMapping(string ollamaModelName)
+    {
+        if (RemoveModelPort(ollamaModelName))
+        {
+            _logger.LogWarning(
+                "Removed stale llama.cpp port mapping for {Model}.",
+                ollamaModelName);
+            return true;
+        }
+
+        return false;
+    }
+
+    public Uri? GetUpstream(string ollamaModelName)
+    {
+        if (_modelPorts.TryGetValue(ollamaModelName, out var port))
+        {
+            return new Uri($"http://localhost:{port}");
+        }
+
+        return null;
+    }
+
+    public async Task<bool> StartModelContainerAsync(LlamaCppModel model, CancellationToken cancellationToken)
+    {
+        var ollamaName = model.OllamaName;
+        if (string.IsNullOrWhiteSpace(ollamaName))
+        {
+            _logger.LogWarning("Cannot start container: model has no Ollama name");
+            return false;
+        }
+
+        var startLock = _modelStartLocks.GetOrAdd(ollamaName, static _ => new SemaphoreSlim(1, 1));
+        await startLock.WaitAsync(cancellationToken);
+        try
+        {
+            return await StartModelContainerCoreAsync(model, cancellationToken);
+        }
+        finally
+        {
+            startLock.Release();
+        }
+    }
+
+    private async Task<bool> StartModelContainerCoreAsync(LlamaCppModel model, CancellationToken cancellationToken)
+    {
+        var ollamaName = model.OllamaName;
+        if (string.IsNullOrWhiteSpace(ollamaName))
+        {
+            return false;
+        }
+
+        if (_modelPorts.ContainsKey(ollamaName))
+        {
+            _logger.LogInformation("Model {Model} already has a running container", ollamaName);
+            return true;
+        }
+
+        if (IsModelOnFallback(ollamaName))
+        {
+            _logger.LogInformation(
+                "Model {Model} previously failed to load via llama.cpp. Skipping container start.",
+                ollamaName);
+            return false;
+        }
+
+        if (!File.Exists(model.BlobPath))
+        {
+            _logger.LogError("Model blob not found: {BlobPath}", model.BlobPath);
+            MarkModelAsFallback(ollamaName);
+            return false;
+        }
+
+        var port = FindAvailablePort();
+        var containerName = SanitizeContainerName($"ngino-llamacpp-{ollamaName}");
+
+        try
+        {
+            var existingPort = await FindExistingContainerPortAsync(containerName);
+            if (existingPort.HasValue)
+            {
+                _logger.LogInformation(
+                    "Reusing existing container for {Model} on port {Port}", ollamaName, existingPort.Value);
+                _modelPorts[ollamaName] = existingPort.Value;
+                return true;
+            }
+
+            await RunDockerAsync(["rm", "-f", containerName], CancellationToken.None);
+
+            var args = BuildDockerRunArgs(containerName, model, port);
+            _logger.LogInformation(
+                "Starting llama.cpp container for {Model} on port {Port}: docker {Args}",
+                ollamaName, port, string.Join(" ", args));
+
+            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;
+            }
+
+            _logger.LogInformation(
+                "llama.cpp container for {Model} started on port {Port}. Waiting for it to become ready...",
+                ollamaName, port);
+
+            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} ({Result}). Stopping it.",
+                    ollamaName, port, ContainerStartTimeout, result);
+
+                try
+                {
+                    await RunDockerAsync(["stop", "--time", "10", containerName], CancellationToken.None);
+                    await RunDockerAsync(["rm", "-f", containerName], CancellationToken.None);
+                }
+                catch (Exception cleanupException)
+                {
+                    _logger.LogWarning(cleanupException, "Failed to clean up container {ContainerName}", containerName);
+                }
+
+                return false;
+            }
+
+            _modelPorts[ollamaName] = port;
+            _logger.LogInformation("llama.cpp container for {Model} is ready on port {Port}.", ollamaName, port);
+            return true;
+        }
+        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 (!RemoveModelPort(ollamaModelName))
+        {
+            _logger.LogWarning("No running container found for model {Model}", ollamaModelName);
+            return false;
+        }
+
+        var containerName = SanitizeContainerName($"ngino-llamacpp-{ollamaModelName}");
+
+        _logger.LogInformation("Stopping llama.cpp container {ContainerName}", containerName);
+
+        try
+        {
+            await RunDockerAsync(["stop", "--time", "10", containerName], cancellationToken);
+            await RunDockerAsync(["rm", "-f", containerName], cancellationToken);
+            ClearModelFallback(ollamaModelName);
+            return true;
+        }
+        catch (Exception ex)
+        {
+            _logger.LogError(ex, "Failed to stop container {ContainerName}", containerName);
+            return false;
+        }
+    }
+
+    public async Task StopAllContainersAsync()
+    {
+        _logger.LogInformation("Stopping all llama.cpp containers...");
+
+        try
+        {
+            var (exitCode, output) = await RunDockerWithOutputAsync(
+                ["ps", "-q", "--filter", $"label={NginoContainerLabel}"],
+                CancellationToken.None);
+
+            if (exitCode == 0 && !string.IsNullOrWhiteSpace(output))
+            {
+                var containerIds = output.Split('\n', StringSplitOptions.RemoveEmptyEntries | StringSplitOptions.TrimEntries);
+                foreach (var id in containerIds)
+                {
+                    await RunDockerAsync(["stop", "--time", "10", id], CancellationToken.None);
+                    await RunDockerAsync(["rm", "-f", id], CancellationToken.None);
+                }
+            }
+        }
+        catch (Exception ex)
+        {
+            _logger.LogWarning(ex, "Failed to stop all llama.cpp containers");
+        }
+
+        _modelPorts.Clear();
+        _reservedPorts.Clear();
+        _fallbackModels.Clear();
+    }
+
+    public async Task<bool> TestDockerAsync()
+    {
+        try
+        {
+            var (exitCode, _) = await RunDockerWithOutputAsync(["info", "--format", "{{.ServerVersion}}"], CancellationToken.None);
+            return exitCode == 0;
+        }
+        catch
+        {
+            return false;
+        }
+    }
+
+    public async ValueTask DisposeAsync()
+    {
+        await StopAllContainersAsync();
+    }
+
+    private LlamaCppModel? ParseManifest(string manifestPath)
+    {
+        var json = File.ReadAllText(manifestPath);
+        using var document = JsonDocument.Parse(json);
+        var root = document.RootElement;
+
+        if (!root.TryGetProperty("layers", out var layers) || layers.ValueKind != JsonValueKind.Array)
+        {
+            return null;
+        }
+
+        string? modelDigest = null;
+        foreach (var layer in layers.EnumerateArray())
+        {
+            if (layer.TryGetProperty("mediaType", out var mediaType)
+                && mediaType.GetString() == "application/vnd.ollama.image.model"
+                && layer.TryGetProperty("digest", out var digest))
+            {
+                modelDigest = digest.GetString();
+                break;
+            }
+        }
+
+        if (string.IsNullOrWhiteSpace(modelDigest))
+        {
+            return null;
+        }
+
+        var modelName = ResolveModelName(manifestPath);
+        if (modelName is null)
+        {
+            return null;
+        }
+
+        var blobName = modelDigest.Replace(":", "-", StringComparison.Ordinal);
+        var blobPath = Path.GetFullPath(Path.Combine(_blobsPath, blobName));
+
+        return new LlamaCppModel
+        {
+            OllamaName = modelName,
+            BlobDigest = blobName,
+            BlobPath = blobPath,
+            ManifestPath = manifestPath
+        };
+    }
+
+    private static string? ResolveModelName(string manifestPath)
+    {
+        var normalizedPath = manifestPath.Replace('\\', '/');
+        var parts = normalizedPath.Split('/', StringSplitOptions.RemoveEmptyEntries);
+
+        var manifestIndex = Array.FindLastIndex(parts, p =>
+            p.Equals("manifests", StringComparison.OrdinalIgnoreCase));
+
+        if (manifestIndex < 0 || manifestIndex >= parts.Length - 1)
+        {
+            return null;
+        }
+
+        var relativeParts = parts[(manifestIndex + 1)..];
+
+        if (relativeParts.Length < 2)
+        {
+            return null;
+        }
+
+        var registry = relativeParts[0];
+
+        if (registry.Equals("registry.ollama.ai", StringComparison.OrdinalIgnoreCase))
+        {
+            if (relativeParts.Length < 3)
+            {
+                return null;
+            }
+
+            if (relativeParts[1].Equals("library", StringComparison.OrdinalIgnoreCase) && relativeParts.Length >= 4)
+            {
+                return $"{relativeParts[2]}:{relativeParts[3]}";
+            }
+
+            if (relativeParts.Length == 3)
+            {
+                return $"{relativeParts[1]}:{relativeParts[2]}";
+            }
+
+            return null;
+        }
+
+        if (relativeParts.Length >= 3)
+        {
+            var tag = relativeParts[^1];
+            var modelPath = string.Join("/", relativeParts.Take(relativeParts.Length - 1));
+            return $"{modelPath}:{tag}";
+        }
+
+        return null;
+    }
+
+    private string[] BuildDockerRunArgs(string containerName, LlamaCppModel model, int port)
+    {
+        var blobsDir = Path.GetDirectoryName(Path.GetFullPath(model.BlobPath))!;
+        var blobFile = Path.GetFileName(model.BlobPath);
+
+        var args = new List<string>
+        {
+            "run",
+            "-d",
+            "--label", $"{NginoContainerLabel}=true",
+            "--name", containerName,
+            "-p", $"{port}:{port}",
+            "-v", $"{blobsDir}:/models/blobs:ro",
+        };
+
+        if (HasRocmDevices())
+        {
+            args.Add("--device=/dev/kfd");
+            args.Add("--device=/dev/dri");
+            args.Add("--group-add=video");
+        }
+
+        if (HasNvidiaGpu() && !HasRocmDevices())
+        {
+            args.Add("--gpus=all");
+        }
+
+        args.Add(_dockerImage);
+        args.Add("--embeddings");
+        args.Add("-m");
+        args.Add($"/models/blobs/{blobFile}");
+        args.Add("-ngl");
+        args.Add("auto");
+        if (_parallel.HasValue)
+        {
+            args.Add("--parallel");
+            args.Add(_parallel.Value.ToString());
+        }
+        args.Add("--host");
+        args.Add("0.0.0.0");
+        args.Add("--port");
+        args.Add(port.ToString());
+
+        return [.. args];
+    }
+
+    private async Task<int?> FindExistingContainerPortAsync(string containerName)
+    {
+        var (exitCode, output) = await RunDockerWithOutputAsync(
+            ["ps", "--filter", $"name=^{containerName}$", "--format", "{{.Ports}}"],
+            CancellationToken.None);
+
+        if (exitCode != 0 || string.IsNullOrWhiteSpace(output))
+        {
+            return null;
+        }
+
+        foreach (var line in output.Split('\n', StringSplitOptions.RemoveEmptyEntries | StringSplitOptions.TrimEntries))
+        {
+            var port = ParseHostPort(line);
+            if (port.HasValue)
+            {
+                return port;
+            }
+        }
+
+        return null;
+    }
+
+    private static int? ParseHostPort(string ports)
+    {
+        foreach (var mapping in ports.Split(','))
+        {
+            var trimmed = mapping.Trim();
+            var arrowIndex = trimmed.IndexOf("->", StringComparison.Ordinal);
+            if (arrowIndex < 0)
+            {
+                continue;
+            }
+
+            var hostPart = trimmed[..arrowIndex].Trim();
+            var colonIndex = hostPart.LastIndexOf(':');
+            if (colonIndex < 0)
+            {
+                continue;
+            }
+
+            if (int.TryParse(hostPart[(colonIndex + 1)..], out var port))
+            {
+                return port;
+            }
+        }
+
+        return null;
+    }
+
+    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))
+            {
+                port++;
+            }
+
+            _reservedPorts[port] = 0;
+            return port;
+        }
+    }
+
+    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;
+
+        var tcpResult = await WaitForTcpPortAsync(host, port, containerName, deadline, cancellationToken);
+        if (tcpResult != ContainerStartResult.Ready)
+        {
+            return tcpResult;
+        }
+
+        using var handler = new SocketsHttpHandler
+        {
+            ConnectTimeout = TimeSpan.FromSeconds(3),
+            UseProxy = false
+        };
+        using var httpClient = new HttpClient(handler) { Timeout = TimeSpan.FromSeconds(10) };
+
+        while (DateTime.UtcNow < deadline)
+        {
+            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 ContainerStartResult.Ready;
+                }
+            }
+            catch (OperationCanceledException) when (!cancellationToken.IsCancellationRequested)
+            {
+            }
+            catch (HttpRequestException ex)
+            {
+                _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 ContainerStartResult.TimedOut;
+    }
+
+    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 ContainerStartResult.Ready;
+            }
+            catch (OperationCanceledException)
+            {
+                throw;
+            }
+            catch (SocketException)
+            {
+            }
+
+            await Task.Delay(TimeSpan.FromSeconds(2), cancellationToken);
+        }
+
+        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)
+    {
+        var sanitized = InvalidContainerNameChars().Replace(name, "_");
+        return sanitized.Trim('_').ToLowerInvariant();
+    }
+
+    private async Task<int> RunDockerAsync(string[] args, CancellationToken cancellationToken)
+    {
+        var (exitCode, _) = await RunDockerWithOutputAsync(args, cancellationToken);
+        return exitCode;
+    }
+
+    private async Task<(int ExitCode, string Output)> RunDockerWithOutputAsync(
+        string[] args, CancellationToken cancellationToken)
+    {
+        var process = new Process
+        {
+            StartInfo = new ProcessStartInfo
+            {
+                FileName = "docker",
+                Arguments = string.Join(" ", args.Select(a => a.Contains(' ') ? $"\"{a}\"" : a)),
+                RedirectStandardOutput = true,
+                RedirectStandardError = true,
+                UseShellExecute = false,
+                CreateNoWindow = true
+            }
+        };
+
+        process.Start();
+
+        var readOutput = process.StandardOutput.ReadToEndAsync(cancellationToken);
+        var readError = process.StandardError.ReadToEndAsync(cancellationToken);
+        var waitTask = process.WaitForExitAsync(cancellationToken);
+
+        var completed = await Task.WhenAny(waitTask, Task.Delay(DockerTimeout, cancellationToken));
+
+        string output;
+        string error;
+
+        try
+        {
+            output = await readOutput;
+            error = await readError;
+        }
+        catch
+        {
+            output = "";
+            error = "timed out";
+        }
+
+        if (completed != waitTask)
+        {
+            _logger.LogWarning("Docker command timed out: docker {Args}", string.Join(" ", args));
+            try { process.Kill(entireProcessTree: true); } catch { }
+            return (-1, error);
+        }
+
+        if (!string.IsNullOrWhiteSpace(error))
+        {
+            output = $"{output}\n{error}".Trim();
+        }
+
+        return (process.ExitCode, output);
+    }
+
+    private static string GetDefaultDockerImage()
+    {
+        if (HasRocmDevices())
+        {
+            return "ghcr.io/ggml-org/llama.cpp:server-rocm";
+        }
+
+        if (HasNvidiaGpu())
+        {
+            return "ghcr.io/ggml-org/llama.cpp:server-cuda";
+        }
+
+        return "ghcr.io/ggml-org/llama.cpp:server";
+    }
+
+    private static bool HasRocmDevices() => File.Exists("/dev/kfd") && Directory.Exists("/dev/dri");
+
+    private static bool HasNvidiaGpu()
+    {
+        try
+        {
+            return File.Exists("/proc/driver/nvidia/version")
+                || Directory.Exists("/proc/driver/nvidia/gpus");
+        }
+        catch
+        {
+            return false;
+        }
+    }
+
+    [GeneratedRegex(@"[^a-zA-Z0-9_.-]")]
+    private static partial Regex InvalidContainerNameChars();
+
+    private enum ContainerStartResult
+    {
+        Ready,
+        ContainerExited,
+        TimedOut
+    }
+}
+
+internal sealed record LlamaCppModel
+{
+    public required string OllamaName { get; init; }
+    public required string BlobDigest { get; init; }
+    public required string BlobPath { get; init; }
+    public required string ManifestPath { get; init; }
+}

src/Ngino.Client/OllamaToLlamaCppTranslator.cs

@@ -0,0 +1,605 @@
+using System.Text;
+using System.Text.Json;
+using Microsoft.Extensions.Logging;
+using Ngino.Protocol;
+
+namespace Ngino.Client;
+
+internal sealed class OllamaToLlamaCppTranslator
+{
+    private static readonly JsonSerializerOptions JsonOptions = new(JsonSerializerDefaults.Web);
+
+    private readonly string _modelName;
+    private readonly ILogger _logger;
+
+    public OllamaToLlamaCppTranslator(string modelName, ILogger logger)
+    {
+        _modelName = modelName;
+        _logger = logger;
+    }
+
+    public bool TryTranslatePath(string method, string pathAndQuery, out string newPath)
+    {
+        var path = pathAndQuery.Split('?')[0];
+        newPath = path switch
+        {
+            "/api/generate" => "/completion",
+            "/api/chat" => "/v1/chat/completions",
+            "/api/embed" or "/api/embeddings" => "/v1/embeddings",
+            _ => null!
+        };
+        return newPath is not null;
+    }
+
+    public byte[] TranslateBody(string pathAndQuery, byte[] body)
+    {
+        if (body is null || body.Length == 0)
+            return body;
+
+        var path = pathAndQuery.Split('?')[0];
+
+        try
+        {
+            return path switch
+            {
+                "/api/generate" => TranslateGenerateBody(body),
+                "/api/chat" => TranslateChatBody(body),
+                "/api/embed" or "/api/embeddings" => TranslateEmbedBody(body),
+                _ => body
+            };
+        }
+        catch (Exception ex)
+        {
+            _logger.LogWarning(ex, "Failed to translate request body for {Path}", path);
+            return body;
+        }
+    }
+
+    private byte[] TranslateGenerateBody(byte[] body)
+    {
+        using var doc = JsonDocument.Parse(body);
+        var root = doc.RootElement;
+
+        var result = new Dictionary<string, object?>();
+
+        if (root.TryGetProperty("prompt", out var prompt))
+            result["prompt"] = prompt.GetString() ?? "";
+
+        result["stream"] = true;
+
+        CopyOptions(root, result);
+
+        if (!result.ContainsKey("n_predict"))
+            result["n_predict"] = 2048;
+
+        return JsonSerializer.SerializeToUtf8Bytes(result, JsonOptions);
+    }
+
+    private byte[] TranslateChatBody(byte[] body)
+    {
+        using var doc = JsonDocument.Parse(body);
+        var root = doc.RootElement;
+
+        var result = new Dictionary<string, object?>();
+
+        if (root.TryGetProperty("messages", out var messages))
+            result["messages"] = messages.Deserialize<object>(JsonOptions);
+
+        result["stream"] = true;
+
+        if (root.TryGetProperty("model", out var model))
+            result["model"] = model.GetString();
+
+        CopyOptions(root, result, chat: true);
+
+        if (!result.ContainsKey("max_tokens"))
+            result["max_tokens"] = 2048;
+
+        return JsonSerializer.SerializeToUtf8Bytes(result, JsonOptions);
+    }
+
+    public static bool ExtractOriginalStream(byte[] originalBody, string pathAndQuery)
+    {
+        var path = pathAndQuery.Split('?')[0];
+        if (path is not "/api/generate" and not "/api/chat")
+            return true;
+
+        try
+        {
+            using var doc = JsonDocument.Parse(originalBody);
+            var root = doc.RootElement;
+            if (root.TryGetProperty("stream", out var stream))
+                return stream.ValueKind != JsonValueKind.False;
+        }
+        catch
+        {
+        }
+
+        return true;
+    }
+
+    private byte[] TranslateEmbedBody(byte[] body)
+    {
+        using var doc = JsonDocument.Parse(body);
+        var root = doc.RootElement;
+
+        var result = new Dictionary<string, object?>
+        {
+            ["model"] = _modelName
+        };
+
+        if (root.TryGetProperty("input", out var input))
+            result["input"] = input.Deserialize<object>(JsonOptions);
+
+        return JsonSerializer.SerializeToUtf8Bytes(result, JsonOptions);
+    }
+
+    public Func<HttpResponseMessage, CancellationToken, Task> CreateResponseHandler(
+        Func<TunnelMessage, CancellationToken, Task> sendAsync,
+        string requestId,
+        string originalPath,
+        Func<bool> originalRequestedStream) =>
+        async (response, ct) =>
+        {
+            await TranslateAndSendResponse(response, sendAsync, requestId, originalPath, originalRequestedStream(), ct);
+        };
+
+    private async Task TranslateAndSendResponse(
+        HttpResponseMessage httpResponse,
+        Func<TunnelMessage, CancellationToken, Task> sendAsync,
+        string requestId,
+        string originalPath,
+        bool originalRequestedStream,
+        CancellationToken cancellationToken)
+    {
+        var path = originalPath.Split('?')[0];
+
+        switch (path)
+        {
+            case "/api/tags":
+                await SynthesizeTagsResponse(sendAsync, requestId, cancellationToken);
+                return;
+
+            case "/api/ps":
+                await SynthesizePsResponse(sendAsync, requestId, cancellationToken);
+                return;
+        }
+
+        if (!httpResponse.IsSuccessStatusCode)
+        {
+            await ForwardRawResponse(httpResponse, sendAsync, requestId, cancellationToken);
+            return;
+        }
+
+        if (path is "/api/generate" or "/api/chat")
+        {
+            if (originalRequestedStream)
+            {
+                await TranslateStreaming(path, httpResponse, sendAsync, requestId, cancellationToken);
+            }
+            else
+            {
+                await TranslateNonStreaming(path, httpResponse, sendAsync, requestId, cancellationToken);
+            }
+        }
+        else
+        {
+            await ForwardRawResponse(httpResponse, sendAsync, requestId, cancellationToken);
+        }
+    }
+
+    private async Task ForwardRawResponse(
+        HttpResponseMessage httpResponse,
+        Func<TunnelMessage, CancellationToken, Task> sendAsync,
+        string requestId,
+        CancellationToken cancellationToken)
+    {
+        await sendAsync(new TunnelMessage
+        {
+            Type = TunnelMessageTypes.HttpResponseHeaders,
+            RequestId = requestId,
+            StatusCode = (int)httpResponse.StatusCode,
+            ReasonPhrase = httpResponse.ReasonPhrase
+        }, cancellationToken);
+
+        await using var stream = await httpResponse.Content.ReadAsStreamAsync(cancellationToken);
+        var buffer = new byte[64 * 1024];
+        while (true)
+        {
+            var bytesRead = await stream.ReadAsync(buffer, cancellationToken);
+            if (bytesRead == 0)
+                break;
+            await sendAsync(new TunnelMessage
+            {
+                Type = TunnelMessageTypes.HttpResponseBody,
+                RequestId = requestId,
+                Body = buffer.AsSpan(0, bytesRead).ToArray()
+            }, cancellationToken);
+        }
+
+        await sendAsync(new TunnelMessage
+        {
+            Type = TunnelMessageTypes.HttpResponseComplete,
+            RequestId = requestId
+        }, cancellationToken);
+    }
+
+    private async Task SynthesizeTagsResponse(
+        Func<TunnelMessage, CancellationToken, Task> sendAsync,
+        string requestId,
+        CancellationToken cancellationToken)
+    {
+        var modelList = new
+        {
+            models = new[]
+            {
+                new
+                {
+                    name = _modelName,
+                    model = _modelName,
+                    modified_at = DateTime.UtcNow.ToString("o"),
+                    size = 0L,
+                    digest = "sha256:" + _modelName,
+                    details = new
+                    {
+                        format = "gguf",
+                        family = "llama",
+                        parameter_size = "",
+                        quantization_level = ""
+                    }
+                }
+            }
+        };
+
+        var body = JsonSerializer.SerializeToUtf8Bytes(modelList, JsonOptions);
+
+        await sendAsync(new TunnelMessage
+        {
+            Type = TunnelMessageTypes.HttpResponseHeaders,
+            RequestId = requestId,
+            StatusCode = 200,
+            ReasonPhrase = "OK"
+        }, cancellationToken);
+
+        await sendAsync(new TunnelMessage
+        {
+            Type = TunnelMessageTypes.HttpResponseBody,
+            RequestId = requestId,
+            Body = body
+        }, cancellationToken);
+
+        await sendAsync(new TunnelMessage
+        {
+            Type = TunnelMessageTypes.HttpResponseComplete,
+            RequestId = requestId
+        }, cancellationToken);
+    }
+
+    private async Task SynthesizePsResponse(
+        Func<TunnelMessage, CancellationToken, Task> sendAsync,
+        string requestId,
+        CancellationToken cancellationToken)
+    {
+        var body = JsonSerializer.SerializeToUtf8Bytes(new { models = Array.Empty<object>() }, JsonOptions);
+
+        await sendAsync(new TunnelMessage
+        {
+            Type = TunnelMessageTypes.HttpResponseHeaders,
+            RequestId = requestId,
+            StatusCode = 200,
+            ReasonPhrase = "OK"
+        }, cancellationToken);
+
+        await sendAsync(new TunnelMessage
+        {
+            Type = TunnelMessageTypes.HttpResponseBody,
+            RequestId = requestId,
+            Body = body
+        }, cancellationToken);
+
+        await sendAsync(new TunnelMessage
+        {
+            Type = TunnelMessageTypes.HttpResponseComplete,
+            RequestId = requestId
+        }, cancellationToken);
+    }
+
+    private async Task TranslateNonStreaming(
+        string path,
+        HttpResponseMessage httpResponse,
+        Func<TunnelMessage, CancellationToken, Task> sendAsync,
+        string requestId,
+        CancellationToken cancellationToken)
+    {
+        await sendAsync(new TunnelMessage
+        {
+            Type = TunnelMessageTypes.HttpResponseHeaders,
+            RequestId = requestId,
+            StatusCode = (int)httpResponse.StatusCode,
+            ReasonPhrase = httpResponse.ReasonPhrase
+        }, cancellationToken);
+
+        var body = await httpResponse.Content.ReadAsByteArrayAsync(cancellationToken);
+        byte[] translatedBody;
+
+        try
+        {
+            translatedBody = path switch
+            {
+                "/api/generate" => TranslateNonStreamingGenerate(body),
+                "/api/chat" => TranslateNonStreamingChat(body),
+                _ => body
+            };
+        }
+        catch (Exception ex)
+        {
+            _logger.LogWarning(ex, "Failed to translate non-streaming response");
+            translatedBody = body;
+        }
+
+        await sendAsync(new TunnelMessage
+        {
+            Type = TunnelMessageTypes.HttpResponseBody,
+            RequestId = requestId,
+            Body = translatedBody
+        }, cancellationToken);
+
+        await sendAsync(new TunnelMessage
+        {
+            Type = TunnelMessageTypes.HttpResponseComplete,
+            RequestId = requestId
+        }, cancellationToken);
+    }
+
+    private async Task TranslateStreaming(
+        string path,
+        HttpResponseMessage httpResponse,
+        Func<TunnelMessage, CancellationToken, Task> sendAsync,
+        string requestId,
+        CancellationToken cancellationToken)
+    {
+        await sendAsync(new TunnelMessage
+        {
+            Type = TunnelMessageTypes.HttpResponseHeaders,
+            RequestId = requestId,
+            StatusCode = 200,
+            ReasonPhrase = "OK"
+        }, cancellationToken);
+
+        await using var stream = await httpResponse.Content.ReadAsStreamAsync(cancellationToken);
+        using var reader = new StreamReader(stream, Encoding.UTF8);
+
+        while (!cancellationToken.IsCancellationRequested)
+        {
+            var line = await reader.ReadLineAsync(cancellationToken);
+            if (line is null)
+                break;
+
+            if (!line.StartsWith("data: ", StringComparison.Ordinal))
+                continue;
+
+            var jsonStr = line[6..];
+            if (jsonStr == "[DONE]")
+                continue;
+
+            try
+            {
+                using var doc = JsonDocument.Parse(jsonStr);
+                var root = doc.RootElement;
+
+                byte[]? chunk = path switch
+                {
+                    "/api/generate" => TranslateGenerateStreamChunk(root),
+                    "/api/chat" => TranslateChatStreamChunk(root),
+                    _ => null
+                };
+
+                if (chunk is not null)
+                {
+                    await sendAsync(new TunnelMessage
+                    {
+                        Type = TunnelMessageTypes.HttpResponseBody,
+                        RequestId = requestId,
+                        Body = chunk
+                    }, cancellationToken);
+                }
+            }
+            catch (JsonException)
+            {
+            }
+        }
+
+        await sendAsync(new TunnelMessage
+        {
+            Type = TunnelMessageTypes.HttpResponseBody,
+            RequestId = requestId,
+            Body = Encoding.UTF8.GetBytes(
+                $"{{\"model\":\"{EscapeJson(_modelName)}\",\"response\":\"\",\"done\":true}}\n")
+        }, cancellationToken);
+
+        await sendAsync(new TunnelMessage
+        {
+            Type = TunnelMessageTypes.HttpResponseComplete,
+            RequestId = requestId
+        }, cancellationToken);
+    }
+
+    private byte[] TranslateNonStreamingGenerate(byte[] body)
+    {
+        using var doc = JsonDocument.Parse(body);
+        var root = doc.RootElement;
+
+        var result = new Dictionary<string, object?>
+        {
+            ["model"] = _modelName,
+            ["created_at"] = DateTime.UtcNow.ToString("o"),
+            ["response"] = root.TryGetProperty("content", out var content) ? content.GetString() : "",
+            ["done"] = root.TryGetProperty("stop", out var stop) && stop.GetBoolean()
+        };
+
+        CopyTimings(root, result);
+        return JsonSerializer.SerializeToUtf8Bytes(result, JsonOptions);
+    }
+
+    private byte[] TranslateNonStreamingChat(byte[] body)
+    {
+        using var doc = JsonDocument.Parse(body);
+        var root = doc.RootElement;
+
+        var message = root.GetProperty("choices")[0].GetProperty("message");
+        var content = message.GetProperty("content").GetString() ?? "";
+
+        var result = new Dictionary<string, object?>
+        {
+            ["model"] = _modelName,
+            ["created_at"] = DateTime.UtcNow.ToString("o"),
+            ["message"] = new Dictionary<string, object?>
+            {
+                ["role"] = "assistant",
+                ["content"] = content
+            },
+            ["done"] = true
+        };
+
+        if (root.TryGetProperty("usage", out var usage))
+        {
+            if (usage.TryGetProperty("completion_tokens", out var comp))
+                result["eval_count"] = comp.GetInt32();
+            if (usage.TryGetProperty("prompt_tokens", out var prompt))
+                result["prompt_eval_count"] = prompt.GetInt32();
+        }
+
+        CopyTimings(root, result);
+        return JsonSerializer.SerializeToUtf8Bytes(result, JsonOptions);
+    }
+
+    private byte[]? TranslateGenerateStreamChunk(JsonElement root)
+    {
+        var done = root.TryGetProperty("stop", out var stop) && stop.GetBoolean();
+        var text = root.TryGetProperty("content", out var content) ? content.GetString() ?? "" : "";
+
+        var result = new Dictionary<string, object?>
+        {
+            ["model"] = _modelName,
+            ["created_at"] = DateTime.UtcNow.ToString("o"),
+            ["response"] = text,
+            ["done"] = done
+        };
+
+        if (done)
+            CopyTimings(root, result);
+
+        var bytes = JsonSerializer.SerializeToUtf8Bytes(result, JsonOptions);
+        var withNewline = new byte[bytes.Length + 1];
+        bytes.CopyTo(withNewline, 0);
+        withNewline[^1] = (byte)'\n';
+        return withNewline;
+    }
+
+    private byte[]? TranslateChatStreamChunk(JsonElement root)
+    {
+        if (!root.TryGetProperty("choices", out var choices) || choices.GetArrayLength() == 0)
+            return null;
+
+        var choice = choices[0];
+        var delta = choice.GetProperty("delta");
+        var finishReason = choice.TryGetProperty("finish_reason", out var fr) ? fr.GetString() : null;
+        var done = finishReason is not null && finishReason != "null" && finishReason != "";
+
+        var content = delta.TryGetProperty("content", out var c) ? c.GetString() ?? "" : "";
+        var role = delta.TryGetProperty("role", out var r) ? r.GetString() : null;
+
+        var result = new Dictionary<string, object?>
+        {
+            ["model"] = _modelName,
+            ["created_at"] = DateTime.UtcNow.ToString("o"),
+            ["message"] = new Dictionary<string, object?>
+            {
+                ["role"] = role ?? "assistant",
+                ["content"] = content
+            },
+            ["done"] = done
+        };
+
+        var bytes = JsonSerializer.SerializeToUtf8Bytes(result, JsonOptions);
+        var withNewline = new byte[bytes.Length + 1];
+        bytes.CopyTo(withNewline, 0);
+        withNewline[^1] = (byte)'\n';
+        return withNewline;
+    }
+
+    private static void CopyOptions(JsonElement root, Dictionary<string, object?> target, bool chat = false)
+    {
+        if (!root.TryGetProperty("options", out var options) || options.ValueKind != JsonValueKind.Object)
+            return;
+
+        foreach (var opt in options.EnumerateObject())
+        {
+            var key = (chat ? MapChatOptionName(opt.Name) : MapGenerateOptionName(opt.Name)) ?? opt.Name;
+            target[key] = ValueToObject(opt.Value);
+        }
+    }
+
+    private static string? MapGenerateOptionName(string name) => name switch
+    {
+        "num_predict" => "n_predict",
+        "temperature" => "temperature",
+        "top_p" => "top_p",
+        "top_k" => "top_k",
+        "seed" => "seed",
+        "stop" => "stop",
+        "repeat_penalty" => "repeat_penalty",
+        "repeat_last_n" => "repeat_last_n",
+        "frequency_penalty" => "frequency_penalty",
+        "presence_penalty" => "presence_penalty",
+        "mirostat" => "mirostat",
+        "mirostat_tau" => "mirostat_tau",
+        "mirostat_eta" => "mirostat_eta",
+        "num_ctx" => "n_ctx",
+        "num_batch" => "n_batch",
+        _ => null
+    };
+
+    private static string? MapChatOptionName(string name) => name switch
+    {
+        "num_predict" => "max_tokens",
+        "temperature" => "temperature",
+        "top_p" => "top_p",
+        "seed" => "seed",
+        "stop" => "stop",
+        "frequency_penalty" => "frequency_penalty",
+        "presence_penalty" => "presence_penalty",
+        _ => null
+    };
+
+    private static void CopyTimings(JsonElement root, Dictionary<string, object?> target)
+    {
+        if (!root.TryGetProperty("timings", out var timings))
+            return;
+
+        if (timings.TryGetProperty("predicted_n", out var predN))
+            target["eval_count"] = predN.GetInt32();
+        if (timings.TryGetProperty("predicted_ms", out var predMs))
+            target["eval_duration"] = (long)(predMs.GetDouble() * 1_000_000);
+        if (timings.TryGetProperty("prompt_n", out var promptN))
+            target["prompt_eval_count"] = promptN.GetInt32();
+        if (timings.TryGetProperty("prompt_ms", out var promptMs))
+            target["prompt_eval_duration"] = (long)(promptMs.GetDouble() * 1_000_000);
+    }
+
+    private static object? ValueToObject(JsonElement element)
+    {
+        return element.ValueKind switch
+        {
+            JsonValueKind.String => element.GetString(),
+            JsonValueKind.Number => element.TryGetInt64(out var l) ? (object)l : element.GetDouble(),
+            JsonValueKind.True => true,
+            JsonValueKind.False => false,
+            JsonValueKind.Null => null,
+            _ => element.Deserialize<object>(JsonOptions)
+        };
+    }
+
+    private static string EscapeJson(string value) =>
+        value.Replace("\\", "\\\\").Replace("\"", "\\\"");
+}

src/Ngino.Client/Program.cs

@@ -16,6 +16,21 @@ try
         Console.WriteLine("  WARNING: server TLS certificate validation is disabled");
     }
 
+    if (options.UseLlamaCppViaDocker)
+    {
+        Console.WriteLine($"  llama.cpp via Docker: enabled");
+        Console.WriteLine($"  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);
@@ -24,6 +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

@@ -16,11 +16,13 @@ internal sealed class TunnelClient
     private const string EmbeddingWarmupInput = "Ngino warmup";
 
     private readonly ConcurrentDictionary<string, UpstreamRequest> _activeRequests = new();
+    private readonly ConcurrentDictionary<string, PendingRequestBody> _pendingRequestBodies = new();
     private readonly HttpClient _httpClient;
     private readonly ClientOptions _options;
     private readonly ILogger<TunnelClient> _logger;
     private readonly object _modelSnapshotLock = new();
     private readonly SemaphoreSlim _sendLock = new(1, 1);
+    private readonly LlamaCppManager? _llamaCppManager;
     private List<string> _lastActiveModels = [];
     private List<string> _lastModels = [];
 
@@ -32,10 +34,41 @@ internal sealed class TunnelClient
         {
             Timeout = Timeout.InfiniteTimeSpan
         };
+
+        if (_options.UseLlamaCppViaDocker)
+        {
+            if (string.IsNullOrWhiteSpace(_options.UseOllamaModelsPath))
+            {
+                throw new InvalidOperationException(
+                    "--use-ollama-models-path is required when --use-llama-cpp-via-docker is set.");
+            }
+
+            _llamaCppManager = new LlamaCppManager(
+                _options.UseOllamaModelsPath,
+                _options.LlamaCppDockerImage,
+                _options.LlamaCppBasePort,
+                _logger,
+                _options.LlamaCppFallbackCooldown,
+                _options.LlamaCppParallel);
+        }
     }
 
     public async Task RunAsync(CancellationToken cancellationToken)
     {
+        if (_llamaCppManager is not null)
+        {
+            _logger.LogInformation("Testing Docker availability...");
+            var dockerAvailable = await _llamaCppManager.TestDockerAsync();
+            if (!dockerAvailable)
+            {
+                _logger.LogWarning("Docker is not available. llama.cpp via Docker will not work.");
+            }
+            else
+            {
+                _logger.LogInformation("Docker is available. Using llama.cpp image: {Image}", _llamaCppManager.DockerImage);
+            }
+        }
+
         while (!cancellationToken.IsCancellationRequested)
         {
             using var socket = new ClientWebSocket();
@@ -154,6 +187,16 @@ internal sealed class TunnelClient
 
     private async Task<List<string>> GetUpstreamModelsAsync(CancellationToken cancellationToken)
     {
+        if (_llamaCppManager is not null)
+        {
+            var models = _llamaCppManager.DiscoverModelsWithBlob();
+            return models
+                .Select(m => m.OllamaName)
+                .Where(name => !string.IsNullOrWhiteSpace(name))
+                .OrderBy(name => name, StringComparer.OrdinalIgnoreCase)
+                .ToList();
+        }
+
         using var request = new HttpRequestMessage(HttpMethod.Get, new Uri(_options.Upstream, "/api/tags"));
         using var response = await _httpClient.SendAsync(request, HttpCompletionOption.ResponseHeadersRead, cancellationToken);
         response.EnsureSuccessStatusCode();
@@ -166,6 +209,17 @@ internal sealed class TunnelClient
 
     private async Task<List<string>> GetActiveUpstreamModelsAsync(CancellationToken cancellationToken)
     {
+        if (_llamaCppManager is not null)
+        {
+            var models = _llamaCppManager.DiscoverModelsWithBlob();
+            return models
+                .Where(m => _llamaCppManager.IsModelActive(m.OllamaName))
+                .Select(m => m.OllamaName)
+                .Where(name => !string.IsNullOrWhiteSpace(name))
+                .OrderBy(name => name, StringComparer.OrdinalIgnoreCase)
+                .ToList();
+        }
+
         using var request = new HttpRequestMessage(HttpMethod.Get, new Uri(_options.Upstream, "/api/ps"));
         using var response = await _httpClient.SendAsync(request, HttpCompletionOption.ResponseHeadersRead, cancellationToken);
 
@@ -290,7 +344,8 @@ internal sealed class TunnelClient
         switch (message.Type)
         {
             case TunnelMessageTypes.HttpRequest:
-                StartRequest(socket, message, cancellationToken);
+                _pendingRequestBodies[message.RequestId] = new PendingRequestBody();
+                _ = Task.Run(() => StartRequest(socket, message, cancellationToken), cancellationToken);
                 break;
 
             case TunnelMessageTypes.HttpRequestBody:
@@ -298,6 +353,10 @@ internal sealed class TunnelClient
                 {
                     requestWithBody.AddBody(message.Body ?? []);
                 }
+                else if (_pendingRequestBodies.TryGetValue(message.RequestId, out var pendingBody))
+                {
+                    pendingBody.AddBody(message.Body ?? []);
+                }
                 break;
 
             case TunnelMessageTypes.HttpRequestComplete:
@@ -305,6 +364,10 @@ internal sealed class TunnelClient
                 {
                     completedRequest.CompleteBody();
                 }
+                else if (_pendingRequestBodies.TryGetValue(message.RequestId, out var pendingBody))
+                {
+                    pendingBody.Complete();
+                }
                 break;
 
             case TunnelMessageTypes.Cancel:
@@ -312,6 +375,10 @@ internal sealed class TunnelClient
                 {
                     cancelledRequest.Cancel();
                 }
+                else
+                {
+                    _pendingRequestBodies.TryRemove(message.RequestId, out _);
+                }
                 break;
 
             case TunnelMessageTypes.ModelCommand:
@@ -363,6 +430,11 @@ internal sealed class TunnelClient
             throw new InvalidOperationException("Model command is missing a model name.");
         }
 
+        if (_llamaCppManager is not null)
+        {
+            return await ExecuteModelCommandWithLlamaCppAsync(message, cancellationToken);
+        }
+
         using var request = BuildModelCommandRequest(message);
         using var response = await _httpClient.SendAsync(request, HttpCompletionOption.ResponseHeadersRead, cancellationToken);
         var body = await response.Content.ReadAsByteArrayAsync(cancellationToken);
@@ -379,6 +451,126 @@ internal sealed class TunnelClient
         return BuildModelCommandResult(message.RequestId, response, body);
     }
 
+    private async Task<TunnelMessage> ExecuteModelCommandWithLlamaCppAsync(
+        TunnelMessage message, CancellationToken cancellationToken)
+    {
+        var modelName = message.Model?.Trim();
+        var normalizedCommand = NormalizeModelCommand(message.Command);
+
+        switch (normalizedCommand)
+        {
+            case "load":
+            {
+                var models = _llamaCppManager!.DiscoverModelsWithBlob();
+                var model = models.FirstOrDefault(m =>
+                    string.Equals(m.OllamaName, modelName, StringComparison.OrdinalIgnoreCase));
+
+                if (model is null)
+                {
+                    return new TunnelMessage
+                    {
+                        Type = TunnelMessageTypes.ModelCommandResult,
+                        RequestId = message.RequestId,
+                        StatusCode = 404,
+                        Error = $"Model '{modelName}' not found in Ollama models path."
+                    };
+                }
+
+                var started = await _llamaCppManager.StartModelContainerAsync(model, cancellationToken);
+                if (!started)
+                {
+                    _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", []);
+            }
+
+            case "unload":
+            {
+                var stopped = await _llamaCppManager!.StopModelContainerAsync(modelName!, cancellationToken);
+                if (!stopped)
+                {
+                    _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", []);
+            }
+
+            case "pull":
+            case "delete":
+            {
+                using var request = BuildModelCommandRequest(_options.Upstream, message.Command, message.Model);
+                using var response = await _httpClient.SendAsync(request, HttpCompletionOption.ResponseHeadersRead, cancellationToken);
+                var body = await response.Content.ReadAsByteArrayAsync(cancellationToken);
+                return BuildModelCommandResult(message.RequestId, response, body);
+            }
+
+            case "show":
+            {
+                var models = _llamaCppManager!.DiscoverModelsWithBlob();
+                var model = models.FirstOrDefault(m =>
+                    string.Equals(m.OllamaName, modelName, StringComparison.OrdinalIgnoreCase));
+
+                if (model is null)
+                {
+                    return new TunnelMessage
+                    {
+                        Type = TunnelMessageTypes.ModelCommandResult,
+                        RequestId = message.RequestId,
+                        StatusCode = 404,
+                        Error = $"Model '{modelName}' not found in Ollama models path."
+                    };
+                }
+
+                var showResponse = new
+                {
+                    modelfile = $"# llama.cpp via Docker\nFROM {model.BlobDigest}\n",
+                    details = new
+                    {
+                        format = "gguf",
+                        family = "llama",
+                        parameter_size = "",
+                        quantization_level = ""
+                    },
+                    model_info = new { }
+                };
+
+                var body = JsonSerializer.SerializeToUtf8Bytes(showResponse, JsonOptions);
+                return new TunnelMessage
+                {
+                    Type = TunnelMessageTypes.ModelCommandResult,
+                    RequestId = message.RequestId,
+                    StatusCode = 200,
+                    ReasonPhrase = "OK",
+                    Body = body
+                };
+            }
+
+            default:
+                throw new InvalidOperationException($"Unsupported model command '{message.Command}' with llama.cpp.");
+        }
+    }
+
     private static TunnelMessage BuildModelCommandResult(
         string requestId,
         HttpResponseMessage response,
@@ -394,6 +586,19 @@ internal sealed class TunnelClient
         };
     }
 
+    private static TunnelMessage BuildModelCommandResult(
+        string requestId, int statusCode, string reasonPhrase, byte[] body)
+    {
+        return new TunnelMessage
+        {
+            Type = TunnelMessageTypes.ModelCommandResult,
+            RequestId = requestId,
+            StatusCode = statusCode,
+            ReasonPhrase = reasonPhrase,
+            Body = body
+        };
+    }
+
     private static bool ShouldRetryModelCommandWithEmbedding(
         string? command,
         HttpResponseMessage response,
@@ -489,15 +694,122 @@ internal sealed class TunnelClient
     private static StringContent JsonContent<T>(T value) =>
         new(JsonSerializer.Serialize(value, JsonOptions), Encoding.UTF8, "application/json");
 
-    private void StartRequest(ClientWebSocket socket, TunnelMessage message, CancellationToken cancellationToken)
+    private async Task StartRequest(ClientWebSocket socket, TunnelMessage message, CancellationToken cancellationToken)
     {
+        Uri? effectiveUpstream = null;
+        var modelName = UpstreamRequest.ExtractModelName(message);
+
+        if (_llamaCppManager is not null)
+        {
+            if (modelName is not null)
+            {
+                effectiveUpstream = _llamaCppManager.GetUpstream(modelName);
+                if (effectiveUpstream is not null)
+                {
+                    var running = await _llamaCppManager.IsContainerRunningAsync(modelName);
+                    if (!running)
+                    {
+                        _logger.LogWarning(
+                            "Cached llama.cpp container for model '{Model}' is not running anymore. Starting a fresh one on demand...",
+                            modelName);
+                        effectiveUpstream = null;
+                    }
+                }
+
+                if (effectiveUpstream is null)
+                {
+                    var model = _llamaCppManager.DiscoverModelsWithBlob()
+                        .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...",
+                                modelName);
+
+                            var started = await _llamaCppManager.StartModelContainerAsync(model, cancellationToken);
+                            if (started)
+                            {
+                                effectiveUpstream = _llamaCppManager.GetUpstream(modelName);
+                            }
+                            else
+                            {
+                                _logger.LogWarning(
+                                    "Unable to load model '{Model}' via llama.cpp. Falling back to the Ollama upstream.",
+                                    modelName);
+                            }
+                        }
+                    }
+                    else
+                    {
+                        _logger.LogWarning(
+                            "Model '{Model}' was not found in the Ollama models path. Falling back to default upstream.",
+                            modelName);
+                    }
+                }
+            }
+        }
+
+        // When llama.cpp backend is active, set up request/response translation
+        Func<string, string?>? pathTransform = null;
+        Func<byte[], byte[]>? bodyTransform = null;
+        Func<HttpResponseMessage, CancellationToken, Task>? responseHandler = null;
+        string? translatorModelName = null;
+
+        if (_llamaCppManager is not null && effectiveUpstream is not null && modelName is not null)
+        {
+            var translator = new OllamaToLlamaCppTranslator(modelName, _logger);
+            translatorModelName = modelName;
+            bool originalRequestedStream = true;
+
+            pathTransform = path =>
+            {
+                if (translator.TryTranslatePath(message.Method ?? "GET", path, out var newPath))
+                {
+                    return newPath;
+                }
+                return null;
+            };
+
+            bodyTransform = body =>
+            {
+                originalRequestedStream = OllamaToLlamaCppTranslator.ExtractOriginalStream(body, message.PathAndQuery ?? "/");
+                return translator.TranslateBody(message.PathAndQuery ?? "/", body);
+            };
+
+            responseHandler = translator.CreateResponseHandler(
+                (response, token) => SendAsync(socket, response, token),
+                message.RequestId,
+                message.PathAndQuery ?? "/",
+                () => originalRequestedStream);
+        }
+
         var request = new UpstreamRequest(
             _options,
             _httpClient,
             message,
             (response, token) => SendAsync(socket, response, token),
             requestId => _activeRequests.TryRemove(requestId, out _),
-            cancellationToken);
+            cancellationToken,
+            effectiveUpstream: effectiveUpstream,
+            responseHandler: responseHandler,
+            pathTransform: pathTransform,
+            bodyTransform: bodyTransform,
+            onConnectionRefused: () =>
+            {
+                if (_llamaCppManager is not null && modelName is not null)
+                {
+                    _llamaCppManager.RemoveModelMapping(modelName);
+                }
+            });
 
         if (!_activeRequests.TryAdd(message.RequestId, request))
         {
@@ -510,10 +822,16 @@ internal sealed class TunnelClient
                     Error = "Duplicate request id."
                 },
                 cancellationToken);
+            _pendingRequestBodies.TryRemove(message.RequestId, out _);
             return;
         }
 
-        _ = Task.Run(request.RunAsync, cancellationToken);
+        if (_pendingRequestBodies.TryRemove(message.RequestId, out var pendingBody))
+        {
+            pendingBody.TransferTo(request);
+        }
+
+        await request.RunAsync();
     }
 
     private async Task SendAsync(ClientWebSocket socket, TunnelMessage message, CancellationToken cancellationToken)
@@ -541,5 +859,51 @@ internal sealed class TunnelClient
                 request.Cancel();
             }
         }
+
+        _pendingRequestBodies.Clear();
+    }
+}
+
+internal sealed class PendingRequestBody
+{
+    private readonly List<byte[]> _chunks = [];
+    private bool _completed;
+    private readonly object _lock = new();
+
+    public void AddBody(byte[] chunk)
+    {
+        if (chunk.Length == 0)
+            return;
+
+        lock (_lock)
+        {
+            _chunks.Add(chunk);
+        }
+    }
+
+    public void Complete()
+    {
+        lock (_lock)
+        {
+            _completed = true;
+        }
+    }
+
+    public void TransferTo(UpstreamRequest request)
+    {
+        lock (_lock)
+        {
+            foreach (var chunk in _chunks)
+            {
+                request.AddBody(chunk);
+            }
+
+            _chunks.Clear();
+
+            if (_completed)
+            {
+                request.CompleteBody();
+            }
+        }
     }
 }

src/Ngino.Client/UpstreamRequest.cs

@@ -1,4 +1,5 @@
 using System.Net.Http.Headers;
+using System.Net.Sockets;
 using System.Threading.Channels;
 using Ngino.Protocol;
 
@@ -19,7 +20,8 @@ internal sealed class UpstreamRequest
         "Trailer",
         "Transfer-Encoding",
         "Upgrade",
-        ProtocolConstants.TokenHeader
+        ProtocolConstants.TokenHeader,
+        ProtocolConstants.ModelHeader
     };
 
     private readonly CancellationTokenSource _cancellationTokenSource;
@@ -33,8 +35,13 @@ internal sealed class UpstreamRequest
     private readonly HttpClient _httpClient;
     private readonly TunnelMessage _initialMessage;
     private readonly Action<string> _onComplete;
-    private readonly ClientOptions _options;
+    private readonly Uri _upstream;
     private readonly Func<TunnelMessage, CancellationToken, Task> _sendAsync;
+    private readonly Func<HttpResponseMessage, CancellationToken, Task>? _responseHandler;
+    private readonly Func<string, string?>? _pathTransform;
+    private readonly Func<byte[], byte[]>? _bodyTransform;
+    private readonly Action? _onConnectionRefused;
+    private readonly List<byte[]> _bufferedBody = [];
 
     public UpstreamRequest(
         ClientOptions options,
@@ -42,13 +49,22 @@ internal sealed class UpstreamRequest
         TunnelMessage initialMessage,
         Func<TunnelMessage, CancellationToken, Task> sendAsync,
         Action<string> onComplete,
-        CancellationToken cancellationToken)
+        CancellationToken cancellationToken,
+        Uri? effectiveUpstream = null,
+        Func<HttpResponseMessage, CancellationToken, Task>? responseHandler = null,
+        Func<string, string?>? pathTransform = null,
+        Func<byte[], byte[]>? bodyTransform = null,
+        Action? onConnectionRefused = null)
     {
-        _options = options;
         _httpClient = httpClient;
         _initialMessage = initialMessage;
         _sendAsync = sendAsync;
         _onComplete = onComplete;
+        _upstream = effectiveUpstream ?? options.Upstream;
+        _responseHandler = responseHandler;
+        _pathTransform = pathTransform;
+        _bodyTransform = bodyTransform;
+        _onConnectionRefused = onConnectionRefused;
         _cancellationTokenSource = CancellationTokenSource.CreateLinkedTokenSource(cancellationToken);
 
         if (!initialMessage.HasBody)
@@ -61,12 +77,36 @@ internal sealed class UpstreamRequest
     {
         if (body.Length > 0)
         {
-            _requestBody.Writer.TryWrite(body);
+            if (_bodyTransform is not null)
+            {
+                _bufferedBody.Add(body);
+            }
+            else
+            {
+                _requestBody.Writer.TryWrite(body);
+            }
         }
     }
 
-    public void CompleteBody() =>
+    public void CompleteBody()
+    {
+        if (_bodyTransform is not null && _bufferedBody.Count > 0)
+        {
+            var totalLength = _bufferedBody.Sum(b => b.Length);
+            var concatenated = new byte[totalLength];
+            var offset = 0;
+            foreach (var chunk in _bufferedBody)
+            {
+                chunk.CopyTo(concatenated, offset);
+                offset += chunk.Length;
+            }
+
+            var transformed = _bodyTransform(concatenated);
+            _requestBody.Writer.TryWrite(transformed);
+        }
+
         _requestBody.Writer.TryComplete();
+    }
 
     public void Cancel()
     {
@@ -84,22 +124,40 @@ internal sealed class UpstreamRequest
                 HttpCompletionOption.ResponseHeadersRead,
                 _cancellationTokenSource.Token);
 
-            await SendResponseHeadersAsync(response);
+            if (_responseHandler is not null)
-            await SendResponseBodyAsync(response);
+            {
+                await _responseHandler(response, _cancellationTokenSource.Token);
+            }
+            else
+            {
+                await SendResponseHeadersAsync(response);
+                await SendResponseBodyAsync(response);
 
-            await _sendAsync(
+                await _sendAsync(
-                new TunnelMessage
+                    new TunnelMessage
-                {
+                    {
-                    Type = TunnelMessageTypes.HttpResponseComplete,
+                        Type = TunnelMessageTypes.HttpResponseComplete,
-                    RequestId = _initialMessage.RequestId
+                        RequestId = _initialMessage.RequestId
-                },
+                    },
-                _cancellationTokenSource.Token);
+                    _cancellationTokenSource.Token);
+            }
         }
         catch (OperationCanceledException) when (_cancellationTokenSource.IsCancellationRequested)
         {
         }
         catch (Exception exception)
         {
+            if (IsConnectionRefused(exception))
+            {
+                try
+                {
+                    _onConnectionRefused?.Invoke();
+                }
+                catch
+                {
+                }
+            }
+
             await SendErrorAsync(exception);
         }
         finally
@@ -110,10 +168,39 @@ internal sealed class UpstreamRequest
         }
     }
 
+    public static string? ExtractModelName(TunnelMessage message)
+    {
+        if (message.Headers is null)
+        {
+            return null;
+        }
+
+        foreach (var header in message.Headers)
+        {
+            if (string.Equals(header.Name, ProtocolConstants.ModelHeader, StringComparison.OrdinalIgnoreCase)
+                && !string.IsNullOrWhiteSpace(header.Value))
+            {
+                return header.Value;
+            }
+        }
+
+        return null;
+    }
+
     private HttpRequestMessage BuildHttpRequest()
     {
+        var path = _initialMessage.PathAndQuery ?? "/";
+        if (_pathTransform is not null)
+        {
+            var transformed = _pathTransform(path);
+            if (transformed is not null)
+            {
+                path = transformed;
+            }
+        }
+
         var method = new HttpMethod(_initialMessage.Method ?? HttpMethod.Get.Method);
-        var request = new HttpRequestMessage(method, BuildUpstreamUri(_options.Upstream, _initialMessage.PathAndQuery));
+        var request = new HttpRequestMessage(method, BuildUpstreamUri(_upstream, path));
 
         if (_initialMessage.HasBody)
         {
@@ -185,7 +272,7 @@ internal sealed class UpstreamRequest
     private async Task SendResponseBodyAsync(HttpResponseMessage response)
     {
         await using var stream = await response.Content.ReadAsStreamAsync(_cancellationTokenSource.Token);
-        var buffer = new byte[_options.ChunkSize];
+        var buffer = new byte[_upstream switch { _ => 64 * 1024 }];
 
         while (true)
         {
@@ -224,6 +311,25 @@ internal sealed class UpstreamRequest
         }
     }
 
+    internal static bool IsConnectionRefused(Exception exception)
+    {
+        for (var current = exception; current is not null; current = current.InnerException)
+        {
+            if (current is SocketException socketException
+                && socketException.SocketErrorCode == SocketError.ConnectionRefused)
+            {
+                return true;
+            }
+
+            if (current.Message.Contains("Connection refused", StringComparison.OrdinalIgnoreCase))
+            {
+                return true;
+            }
+        }
+
+        return false;
+    }
+
     private static List<HeaderPair> CollectResponseHeaders(HttpResponseMessage response)
     {
         var headers = new List<HeaderPair>();

src/Ngino.Protocol/ProtocolConstants.cs

@@ -7,4 +7,5 @@ public static class ProtocolConstants
     public const string TokenHeader = "X-Ngino-Token";
     public const string ClientIdHeader = "X-Ngino-Client-Id";
     public const string ReplacedCloseDescription = "ngino-replaced";
+public const string ModelHeader = "X-Ngino-Model";
 }

src/Ngino.Server/ManagementStore.cs

@@ -1215,7 +1215,7 @@ internal sealed class ManagementStore
             using var connection = OpenConnection();
             using var command = connection.CreateCommand();
             command.CommandText = """
-                SELECT g.name, gm.client_id, gm.client_pattern
+                SELECT g.id, gm.client_id, gm.client_pattern
                 FROM group_members gm
                 INNER JOIN groups g ON gm.group_id = g.id
                 """;
@@ -1223,14 +1223,14 @@ internal sealed class ManagementStore
             using var reader = command.ExecuteReader();
             while (reader.Read())
             {
-                var groupName = reader.GetString(0);
+                var groupId = reader.GetString(0);
                 var explicitClientId = reader.IsDBNull(1) ? null : reader.GetString(1);
                 var pattern = reader.IsDBNull(2) ? null : reader.GetString(2);
 
                 if (!string.IsNullOrWhiteSpace(explicitClientId)
                     && result.TryGetValue(explicitClientId, out var explicitGroups))
                 {
-                    explicitGroups.Add(groupName);
+                    explicitGroups.Add(groupId);
                 }
                 else if (!string.IsNullOrWhiteSpace(pattern))
                 {
@@ -1248,7 +1248,7 @@ internal sealed class ManagementStore
                     {
                         if (regex.IsMatch(clientId) && result.TryGetValue(clientId, out var groups))
                         {
-                            groups.Add(groupName);
+                            groups.Add(groupId);
                         }
                     }
                 }

src/Ngino.Server/Program.cs

@@ -269,6 +269,11 @@ app.UseStaticFiles();
 
 app.MapAdminEndpoints(settings);
 
+app.MapGet("/favicon.ico", () =>
+    Results.File(
+        Path.Combine(app.Environment.WebRootPath, "favicon.ico"),
+        "image/x-icon"));
+
 app.MapGet("/", () => Results.Redirect("/admin"));
 
 app.MapGet(settings.StatusPath, (HttpContext context, TunnelHub hub, ServerSettings serverSettings, EmbeddingCache embeddingCache, ManagementStore managementStore) =>

src/Ngino.Server/ReverseProxyEndpoint.cs

@@ -10,6 +10,8 @@ internal static class ReverseProxyEndpoint
 {
     private const string UnauthorizedMessage = "Missing or invalid Ngino token.";
 
+    private const string OllamaVersion = "0.32.5";
+
     private static readonly HashSet<string> HopByHopHeaders = new(StringComparer.OrdinalIgnoreCase)
     {
         "Connection",
@@ -72,6 +74,12 @@ internal static class ReverseProxyEndpoint
             return;
         }
 
+        if (IsVersionRequest(context.Request, proxyPath))
+        {
+            await WriteVersionResponseAsync(context);
+            return;
+        }
+
         if (TryGetClientAddress(proxyPath, out var pathClientId, out var clientPath))
         {
             if (!groupAccess.IsClientAllowed(pathClientId))
@@ -317,6 +325,15 @@ internal static class ReverseProxyEndpoint
         HttpMethods.IsGet(request.Method)
         && string.Equals(proxyPath.Value, "/api/tags", StringComparison.OrdinalIgnoreCase);
 
+    private static bool IsVersionRequest(HttpRequest request, PathString proxyPath) =>
+        HttpMethods.IsGet(request.Method)
+        && string.Equals(proxyPath.Value, "/api/version", StringComparison.OrdinalIgnoreCase);
+
+    private static Task WriteVersionResponseAsync(HttpContext context) =>
+        context.Response.WriteAsJsonAsync(
+            new { version = OllamaVersion },
+            context.RequestAborted);
+
     private static async Task HandleTagsAsync(
         HttpContext context,
         TunnelHub hub,
@@ -535,6 +552,12 @@ internal static class ReverseProxyEndpoint
                 Headers = CollectRequestHeaders(context.Request, settings, managementStore)
             };
 
+            if (!string.IsNullOrWhiteSpace(requestedModel))
+            {
+                requestMessage.Headers ??= [];
+                requestMessage.Headers.Add(new HeaderPair(ProtocolConstants.ModelHeader, requestedModel));
+            }
+
             await connection.SendAsync(requestMessage, context.RequestAborted);
             requestBodyTask = ForwardRequestBodyAsync(context.Request, connection, requestId, hasBody, settings, logger);
             _ = requestBodyTask.ContinueWith(

src/Ngino.Server/wwwroot/admin/app.css

@@ -58,16 +58,12 @@ textarea {
 }
 
 .brand-mark {
-  display: grid;
-  place-items: center;
   width: 34px;
   height: 34px;
   border: 1px solid #4c5d74;
   border-radius: 6px;
   background: #223044;
-  color: #c8f2df;
+  object-fit: cover;
-  font-weight: 800;
-  letter-spacing: 0;
 }
 
 .brand strong,

src/Ngino.Server/wwwroot/admin/app.js

@@ -491,6 +491,7 @@ function renderClients() {
 
 function clientsTable(clients) {
   const clientGroups = state.summary?.clientGroups || {};
+  const groupNames = Object.fromEntries((state.summary?.groups || []).map((g) => [g.id, g.name]));
   const rows = clients.map((client) => {
     const groups = clientGroups[client.id] || [];
     return `
@@ -516,7 +517,7 @@ function clientsTable(clients) {
       <td>${modelBadges(client.activeModels)}</td>
       <td>
         <div class="badge-row">
-          ${groups.length ? groups.map((g) => `<a class="badge" href="#groups/${encodeURIComponent(g)}">${escapeHtml(g)}</a>`).join("") : `<span class="cell-sub">None</span>`}
+          ${groups.length ? groups.map((g) => `<a class="badge" href="#groups/${encodeURIComponent(g)}">${escapeHtml(groupNames[g] || g)}</a>`).join("") : `<span class="cell-sub">None</span>`}
         </div>
       </td>
       <td>

src/Ngino.Server/wwwroot/admin/index.html

@@ -10,7 +10,7 @@
   <div class="app-shell">
     <aside class="sidebar">
       <div class="brand">
-        <span class="brand-mark">RL</span>
+        <img class="brand-mark" src="/admin/Ngino_logo_symbol.png" alt="Ngino">
         <span>
           <strong>Ngino</strong>
           <small>Admin</small>

tests/Ngino.Client.Tests/UpstreamRequestTests.cs

@@ -1,3 +1,4 @@
+using System.Net.Sockets;
 using Ngino.Client;
 using Xunit;
 
@@ -32,4 +33,29 @@ public sealed class UpstreamRequestTests
 
         Assert.Contains("origin-form path", exception.Message);
     }
+
+    [Fact]
+    public void IsConnectionRefused_DetectsSocketConnectionRefused()
+    {
+        var socketException = new SocketException((int)SocketError.ConnectionRefused);
+        var exception = new HttpRequestException("Connection refused", socketException);
+
+        Assert.True(UpstreamRequest.IsConnectionRefused(exception));
+    }
+
+    [Fact]
+    public void IsConnectionRefused_DetectsConnectionRefusedByMessage()
+    {
+        var exception = new HttpRequestException("Connection refused (localhost:8081)");
+
+        Assert.True(UpstreamRequest.IsConnectionRefused(exception));
+    }
+
+    [Fact]
+    public void IsConnectionRefused_IgnoresUnrelatedFailures()
+    {
+        var exception = new HttpRequestException("Connection reset by peer");
+
+        Assert.False(UpstreamRequest.IsConnectionRefused(exception));
+    }
 }