Implemented new Japanese only text mode with WebUI toggle, utilizing a llama3.1 swallow dataset model. Next up is Japanese TTS.

bot/api.py

@@ -226,6 +226,54 @@ def calm_miku_endpoint():
     
     return {"status": "ok", "message": "Miku has been calmed down"}
 
+# ========== Language Mode Management ==========
+@app.get("/language")
+def get_language_mode():
+    """Get current language mode (english or japanese)"""
+    return {
+        "language_mode": globals.LANGUAGE_MODE,
+        "available_languages": ["english", "japanese"],
+        "current_model": globals.JAPANESE_TEXT_MODEL if globals.LANGUAGE_MODE == "japanese" else globals.TEXT_MODEL
+    }
+
+@app.post("/language/toggle")
+def toggle_language_mode():
+    """Toggle between English and Japanese modes"""
+    if globals.LANGUAGE_MODE == "english":
+        globals.LANGUAGE_MODE = "japanese"
+        new_mode = "japanese"
+        model_used = globals.JAPANESE_TEXT_MODEL
+        logger.info("Switched to Japanese mode (using Llama 3.1 Swallow)")
+    else:
+        globals.LANGUAGE_MODE = "english"
+        new_mode = "english"
+        model_used = globals.TEXT_MODEL
+        logger.info("Switched to English mode (using default model)")
+    
+    return {
+        "status": "ok",
+        "language_mode": new_mode,
+        "model_now_using": model_used,
+        "message": f"Miku is now speaking in {new_mode.upper()}!"
+    }
+
+@app.post("/language/set")
+def set_language_mode(language: str = "english"):
+    """Set language mode to either 'english' or 'japanese'"""
+    if language.lower() not in ["english", "japanese"]:
+        return {"error": f"Invalid language mode '{language}'. Use 'english' or 'japanese'."}, 400
+    
+    globals.LANGUAGE_MODE = language.lower()
+    model_used = globals.JAPANESE_TEXT_MODEL if language.lower() == "japanese" else globals.TEXT_MODEL
+    logger.info(f"Language mode set to {language.lower()} (using {model_used})")
+    
+    return {
+        "status": "ok",
+        "language_mode": language.lower(),
+        "model_now_using": model_used,
+        "message": f"Miku is now speaking in {language.upper()}!"
+    }
+
 # ========== Evil Mode Management ==========
 @app.get("/evil-mode")
 def get_evil_mode_status():

bot/globals.py

@@ -26,8 +26,12 @@ LLAMA_AMD_URL = os.getenv("LLAMA_AMD_URL", "http://llama-swap-amd:8080")  # Seco
 TEXT_MODEL = os.getenv("TEXT_MODEL", "llama3.1")
 VISION_MODEL = os.getenv("VISION_MODEL", "vision")
 EVIL_TEXT_MODEL = os.getenv("EVIL_TEXT_MODEL", "darkidol")  # Uncensored model for evil mode
+JAPANESE_TEXT_MODEL = os.getenv("JAPANESE_TEXT_MODEL", "swallow")  # Llama 3.1 Swallow model for Japanese
 OWNER_USER_ID = int(os.getenv("OWNER_USER_ID", "209381657369772032"))  # Bot owner's Discord user ID for reports
 
+# Language mode for Miku (english or japanese)
+LANGUAGE_MODE = "english"  # Can be "english" or "japanese"
+
 # Fish.audio TTS settings
 FISH_API_KEY = os.getenv("FISH_API_KEY", "478d263d8c094e0c8993aae3e9cf9159")
 MIKU_VOICE_ID = os.getenv("MIKU_VOICE_ID", "b28b79555e8c4904ac4d048c36e716b7")

bot/static/index.html

@@ -660,10 +660,11 @@
       <button class="tab-button active" onclick="switchTab('tab1')">Server Management</button>
       <button class="tab-button" onclick="switchTab('tab2')">Actions</button>
       <button class="tab-button" onclick="switchTab('tab3')">Status</button>
-      <button class="tab-button" onclick="switchTab('tab4')">🎨 Image Generation</button>
+      <button class="tab-button" onclick="switchTab('tab4')">⚙️ LLM Settings</button>
-      <button class="tab-button" onclick="switchTab('tab5')">📊 Autonomous Stats</button>
+      <button class="tab-button" onclick="switchTab('tab5')">🎨 Image Generation</button>
-      <button class="tab-button" onclick="switchTab('tab6')">💬 Chat with LLM</button>
+      <button class="tab-button" onclick="switchTab('tab6')">📊 Autonomous Stats</button>
-      <button class="tab-button" onclick="switchTab('tab7')">📞 Voice Call</button>
+      <button class="tab-button" onclick="switchTab('tab7')">💬 Chat with LLM</button>
+      <button class="tab-button" onclick="switchTab('tab8')">📞 Voice Call</button>
       <button class="tab-button" onclick="window.location.href='/static/system.html'">🎛️ System Settings</button>
     </div>
 
@@ -1173,8 +1174,70 @@
       </div>
     </div>
 
-    <!-- Image Generation Tab Content -->
+    <!-- LLM Settings Tab Content -->
     <div id="tab4" class="tab-content">
+      <div class="section">
+        <h3>⚙️ Language Model Settings</h3>
+        <p>Configure language model behavior and language mode.</p>
+        
+        <!-- Language Mode Section -->
+        <div style="margin-bottom: 1.5rem; padding: 1rem; background: #2a2a2a; border-radius: 4px; border: 2px solid #4a7bc9;">
+          <h4 style="margin-top: 0; color: #61dafb;">🌐 Language Mode</h4>
+          <p style="margin: 0.5rem 0; color: #aaa;">Switch Miku between English and Japanese responses.</p>
+          
+          <div style="margin: 1rem 0;">
+            <div style="margin-bottom: 1rem;">
+              <strong>Current Language:</strong> <span id="current-language-display" style="color: #61dafb; font-weight: bold;">English</span>
+            </div>
+            <button onclick="toggleLanguageMode()" style="background: #4a7bc9; color: #fff; padding: 0.6rem 1.2rem; border: 2px solid #61dafb; border-radius: 4px; cursor: pointer; font-weight: bold; font-size: 1rem;">
+              🔄 Toggle Language (English ↔ Japanese)
+            </button>
+          </div>
+          
+          <div style="margin-top: 1rem; padding: 1rem; background: #1a1a1a; border-radius: 4px; border-left: 3px solid #4a7bc9;">
+            <div style="font-size: 0.9rem;">
+              <div style="margin-bottom: 0.5rem;"><strong>English Mode:</strong></div>
+              <ul style="margin: 0 0 0.5rem 0; padding-left: 1.5rem; color: #aaa;">
+                <li>Uses standard Llama 3.1 model</li>
+                <li>Responds in English only</li>
+              </ul>
+              
+              <div style="margin-bottom: 0.5rem;"><strong>Japanese Mode (日本語):</strong></div>
+              <ul style="margin: 0 0 0; padding-left: 1.5rem; color: #aaa;">
+                <li>Uses Llama 3.1 Swallow model (trained for Japanese)</li>
+                <li>Responds entirely in Japanese</li>
+              </ul>
+            </div>
+          </div>
+        </div>
+
+        <!-- Language Mode Status Section -->
+        <div style="margin-bottom: 1.5rem; padding: 1rem; background: #2a2a2a; border-radius: 4px;">
+          <h4 style="margin-top: 0;">📊 Current Status</h4>
+          <div id="language-status-display" style="background: #1a1a1a; padding: 1rem; border-radius: 4px; font-family: monospace; font-size: 0.9rem;">
+            <p style="margin: 0.5rem 0;"><strong>Language Mode:</strong> <span id="status-language">English</span></p>
+            <p style="margin: 0.5rem 0;"><strong>Active Model:</strong> <span id="status-model">llama3.1</span></p>
+            <p style="margin: 0.5rem 0;"><strong>Available Languages:</strong> English, 日本語 (Japanese)</p>
+          </div>
+          <button onclick="refreshLanguageStatus()" style="margin-top: 1rem;">🔄 Refresh Status</button>
+        </div>
+
+        <!-- Information Section -->
+        <div style="padding: 1rem; background: #1a1a1a; border-radius: 4px; border-left: 3px solid #ff9800;">
+          <h4 style="margin-top: 0; color: #ff9800;">ℹ️ How Language Mode Works</h4>
+          <ul style="margin: 0.5rem 0; padding-left: 1.5rem; font-size: 0.9rem; color: #aaa;">
+            <li>English mode uses your default text model for English responses</li>
+            <li>Japanese mode switches to Swallow and responds only in 日本語</li>
+            <li>All personality traits, mood system, and features work in both modes</li>
+            <li>Language mode is global - affects all servers and DMs</li>
+            <li>Conversation history is preserved across language switches</li>
+          </ul>
+        </div>
+      </div>
+    </div>
+
+    <!-- Image Generation Tab Content -->
+    <div id="tab5" class="tab-content">
       <div class="section">
         <h3>🎨 Image Generation System</h3>
         <p>Natural language image generation powered by ComfyUI. Users can ask Miku to create images naturally without commands!</p>
@@ -1232,7 +1295,7 @@
     </div>
 
     <!-- Autonomous Stats Tab Content -->
-    <div id="tab5" class="tab-content">
+    <div id="tab6" class="tab-content">
       <div class="section">
         <h3>📊 Autonomous V2 Decision Engine Stats</h3>
         <p>Real-time monitoring of Miku's autonomous decision-making context and mood-based personality stats.</p>
@@ -1250,7 +1313,7 @@
     </div>
 
     <!-- Chat with LLM Tab Content -->
-    <div id="tab6" class="tab-content">
+    <div id="tab7" class="tab-content">
       <div class="section">
         <h3>💬 Chat with LLM</h3>
         <p>Direct chat interface with the language models. Test responses, experiment with prompts, or just chat with Miku!</p>
@@ -1375,8 +1438,8 @@
       </div>
     </div>
 
-    <!-- Tab 7: Voice Call Management -->
+    <!-- Tab 8: Voice Call Management -->
-    <div id="tab7" class="tab-content">
+    <div id="tab8" class="tab-content">
       <div class="section">
         <h3>📞 Initiate Voice Call</h3>
         <p>Start an automated voice chat session with a user. Miku will automatically manage containers, join voice chat, and send an invitation DM.</p>
@@ -1559,6 +1622,7 @@ document.addEventListener('DOMContentLoaded', function() {
   checkEvilModeStatus();  // Check evil mode on load
   checkBipolarModeStatus();  // Check bipolar mode on load
   checkGPUStatus();  // Check GPU selection on load
+  refreshLanguageStatus();  // Check language mode on load
   console.log('🚀 DOMContentLoaded - initializing figurine subscribers list');
   refreshFigurineSubscribers();
   loadProfilePictureMetadata();
@@ -2251,6 +2315,43 @@ async function calmMiku() {
   }
 }
 
+// ===== Language Mode Functions =====
+async function refreshLanguageStatus() {
+  try {
+    const result = await apiCall('/language');
+    document.getElementById('current-language-display').textContent = 
+      result.language_mode === 'japanese' ? '日本語 (Japanese)' : 'English';
+    document.getElementById('status-language').textContent = 
+      result.language_mode === 'japanese' ? '日本語 (Japanese)' : 'English';
+    document.getElementById('status-model').textContent = result.current_model;
+    
+    console.log('Language status:', result);
+  } catch (error) {
+    console.error('Failed to get language status:', error);
+    showNotification('Failed to load language status', 'error');
+  }
+}
+
+async function toggleLanguageMode() {
+  try {
+    const result = await apiCall('/language/toggle', 'POST');
+    
+    // Update UI
+    document.getElementById('current-language-display').textContent = 
+      result.language_mode === 'japanese' ? '日本語 (Japanese)' : 'English';
+    document.getElementById('status-language').textContent = 
+      result.language_mode === 'japanese' ? '日本語 (Japanese)' : 'English';
+    document.getElementById('status-model').textContent = result.model_now_using;
+    
+    // Show notification
+    showNotification(result.message, 'success');
+    console.log('Language toggled:', result);
+  } catch (error) {
+    console.error('Failed to toggle language mode:', error);
+    showNotification('Failed to toggle language mode', 'error');
+  }
+}
+
 // Evil Mode Functions
 async function checkEvilModeStatus() {
   try {

bot/utils/context_manager.py

@@ -3,8 +3,12 @@
 Structured context management for Miku's personality and knowledge.
 Replaces the vector search system with organized, complete context.
 Preserves original content files in their entirety.
+
+When LANGUAGE_MODE is "japanese", appends a Japanese instruction to ensure
+all responses are in Japanese without requiring separate files.
 """
 
+import globals
 from utils.logger import get_logger
 
 logger = get_logger('core')
@@ -40,60 +44,96 @@ def get_original_miku_lyrics() -> str:
         return "## MIKU LYRICS\n[File could not be loaded]"
 
 
+def _get_japanese_instruction() -> str:
+    """
+    Returns the Japanese language instruction to append to context.
+    Ensures all responses are in Japanese when in Japanese mode.
+    """
+    return "\n\n[CRITICAL INSTRUCTION - 重要な指示]\n**YOU MUST RESPOND ENTIRELY IN JAPANESE (日本語). NO ENGLISH ALLOWED.**\nすべての返答は必ず日本語で行ってください。英語での返答は一切禁止されています。\nこれは最優先の指示です。必ず守ってください。"
+
+
 def get_complete_context() -> str:
-    """Returns all essential Miku context using original files in their entirety"""
+    """
-    return f"""## MIKU LORE (Complete Original)
+    Returns all essential Miku context using original files in their entirety.
-{get_original_miku_lore()}
+    
+    If LANGUAGE_MODE is "japanese", appends a Japanese instruction to ensure
+    all responses are in Japanese.
+    """
+    lore = get_original_miku_lore()
+    prompt = get_original_miku_prompt()
+    lyrics = get_original_miku_lyrics()
+    
+    combined = f"""## MIKU LORE (Complete Original)
+{lore}
 
 ## MIKU PERSONALITY & GUIDELINES (Complete Original) 
-{get_original_miku_prompt()}
+{prompt}
 
 ## MIKU SONG LYRICS (Complete Original)
-{get_original_miku_lyrics()}"""
+{lyrics}"""
+    
+    # Append Japanese instruction if in Japanese mode
+    if globals.LANGUAGE_MODE == "japanese":
+        combined += _get_japanese_instruction()
+    
+    logger.info(f"[core] Context loaded in {globals.LANGUAGE_MODE} mode")
+    return combined
 
 
 def get_context_for_response_type(response_type: str) -> str:
-    """Returns appropriate context based on the type of response being generated"""
+    """
+    Returns appropriate context based on the type of response being generated.
     
-    # Core context always includes the complete original files
+    If LANGUAGE_MODE is "japanese", appends Japanese instruction to all contexts
+    to ensure responses are in Japanese.
+    """
+    
+    lore = get_original_miku_lore()
+    prompt = get_original_miku_prompt()
+    lyrics = get_original_miku_lyrics()
+    
+    # Build core context (always in English source files)
     core_context = f"""## MIKU LORE (Complete Original)
-{get_original_miku_lore()}
+{lore}
 
 ## MIKU PERSONALITY & GUIDELINES (Complete Original) 
-{get_original_miku_prompt()}"""
+{prompt}"""
     
+    # Return context based on response type
     if response_type == "autonomous_general":
-        # For general autonomous messages, include everything
+        context = f"""{core_context}
-        return f"""{core_context}
 
 ## MIKU SONG LYRICS (Complete Original)
-{get_original_miku_lyrics()}"""
+{lyrics}"""
         
     elif response_type == "autonomous_tweet":
-        # For tweet responses, include lyrics for musical context
+        context = f"""{core_context}
-        return f"""{core_context}
 
 ## MIKU SONG LYRICS (Complete Original)
-{get_original_miku_lyrics()}"""
+{lyrics}"""
         
     elif response_type == "dm_response" or response_type == "server_response":
-        # For conversational responses, include everything
+        context = f"""{core_context}
-        return f"""{core_context}
 
 ## MIKU SONG LYRICS (Complete Original)
-{get_original_miku_lyrics()}"""
+{lyrics}"""
         
     elif response_type == "conversation_join":
-        # For joining conversations, include everything
+        context = f"""{core_context}
-        return f"""{core_context}
 
 ## MIKU SONG LYRICS (Complete Original)
-{get_original_miku_lyrics()}"""
+{lyrics}"""
 
     elif response_type == "emoji_selection":
-        # For emoji reactions, no context needed - the prompt has everything
+        # For emoji reactions, minimal context needed
-        return ""
+        context = ""
             
     else:
         # Default: comprehensive context
-        return get_complete_context()
+        context = get_complete_context()
+    
+    # Append Japanese instruction if in Japanese mode
+    if globals.LANGUAGE_MODE == "japanese" and context:
+        context += _get_japanese_instruction()
+    
+    return context

bot/utils/image_handling.py

@@ -239,7 +239,13 @@ async def analyze_image_with_vision(base64_img):
     Uses OpenAI-compatible chat completions API with image_url.
     Always uses NVIDIA GPU for vision model.
     """
-    from utils.llm import get_vision_gpu_url
+    from utils.llm import get_vision_gpu_url, check_vision_endpoint_health
+    
+    # Check if vision endpoint is healthy before attempting request
+    is_healthy, error = await check_vision_endpoint_health()
+    if not is_healthy:
+        logger.warning(f"Vision endpoint unhealthy: {error}")
+        return f"Vision service currently unavailable: {error}"
     
     payload = {
         "model": globals.VISION_MODEL,
@@ -269,17 +275,20 @@ async def analyze_image_with_vision(base64_img):
     async with aiohttp.ClientSession() as session:
         try:
             vision_url = get_vision_gpu_url()
-            async with session.post(f"{vision_url}/v1/chat/completions", json=payload, headers=headers) as response:
+            logger.info(f"Sending vision request to {vision_url} using model: {globals.VISION_MODEL}")
+            
+            async with session.post(f"{vision_url}/v1/chat/completions", json=payload, headers=headers, timeout=aiohttp.ClientTimeout(total=60)) as response:
                 if response.status == 200:
                     data = await response.json()
-                    return data.get("choices", [{}])[0].get("message", {}).get("content", "No description.")
+                    result = data.get("choices", [{}])[0].get("message", {}).get("content", "No description.")
+                    logger.info(f"Vision analysis completed successfully")
+                    return result
                 else:
                     error_text = await response.text()
-                    logger.error(f"Vision API error: {response.status} - {error_text}")
+                    logger.error(f"Vision API error: {response.status} - {error_text} (endpoint: {vision_url})")
                     return f"Error analyzing image: {response.status}"
         except Exception as e:
-            logger.error(f"Error in analyze_image_with_vision: {e}")
+            logger.error(f"Error in analyze_image_with_vision: {e}", exc_info=True)
-            return f"Error analyzing image: {str(e)}"
 
 
 async def analyze_video_with_vision(video_frames, media_type="video"):
@@ -288,6 +297,13 @@ async def analyze_video_with_vision(video_frames, media_type="video"):
     video_frames: list of base64-encoded frames
     media_type: "video", "gif", or "tenor_gif" to customize the analysis prompt
     """
+    from utils.llm import get_vision_gpu_url, check_vision_endpoint_health
+    
+    # Check if vision endpoint is healthy before attempting request
+    is_healthy, error = await check_vision_endpoint_health()
+    if not is_healthy:
+        logger.warning(f"Vision endpoint unhealthy: {error}")
+        return f"Vision service currently unavailable: {error}"
     
     # Customize prompt based on media type
     if media_type == "gif":
@@ -331,16 +347,20 @@ async def analyze_video_with_vision(video_frames, media_type="video"):
     async with aiohttp.ClientSession() as session:
         try:
             vision_url = get_vision_gpu_url()
-            async with session.post(f"{vision_url}/v1/chat/completions", json=payload, headers=headers) as response:
+            logger.info(f"Sending video analysis request to {vision_url} using model: {globals.VISION_MODEL} (media_type: {media_type}, frames: {len(video_frames)})")
+            
+            async with session.post(f"{vision_url}/v1/chat/completions", json=payload, headers=headers, timeout=aiohttp.ClientTimeout(total=120)) as response:
                 if response.status == 200:
                     data = await response.json()
-                    return data.get("choices", [{}])[0].get("message", {}).get("content", "No description.")
+                    result = data.get("choices", [{}])[0].get("message", {}).get("content", "No description.")
+                    logger.info(f"Video analysis completed successfully")
+                    return result
                 else:
                     error_text = await response.text()
-                    logger.error(f"Vision API error: {response.status} - {error_text}")
+                    logger.error(f"Vision API error: {response.status} - {error_text} (endpoint: {vision_url})")
                     return f"Error analyzing video: {response.status}"
         except Exception as e:
-            logger.error(f"Error in analyze_video_with_vision: {e}")
+            logger.error(f"Error in analyze_video_with_vision: {e}", exc_info=True)
             return f"Error analyzing video: {str(e)}"
 
 

bot/utils/llm.py

@@ -38,8 +38,47 @@ def get_vision_gpu_url():
     Strategy: Always use NVIDIA GPU for vision to avoid unloading/reloading.
     - When NVIDIA is primary: Use NVIDIA for both text and vision
     - When AMD is primary: Use AMD for text, NVIDIA for vision (keeps vision loaded)
+    
+    Important: Vision model (MiniCPM-V) is ONLY configured on NVIDIA GPU.
+    This ensures vision inference is always fast and doesn't interfere with 
+    AMD text model inference.
     """
-    return globals.LLAMA_URL  # Always use NVIDIA for vision
+    current_text_gpu = get_current_gpu_url()
+    nvidia_vision_url = globals.LLAMA_URL
+    
+    # Vision ALWAYS uses NVIDIA, regardless of which GPU is primary for text
+    # Log this decision when GPU switching is active (primary text GPU is AMD)
+    if current_text_gpu == globals.LLAMA_AMD_URL:
+        logger.debug(f"Primary GPU is AMD for text, but using NVIDIA for vision model")
+    
+    return nvidia_vision_url  # Always use NVIDIA for vision
+
+async def check_vision_endpoint_health():
+    """
+    Check if NVIDIA GPU vision endpoint is healthy and responsive.
+    This is important when AMD is the primary GPU to ensure vision still works.
+    
+    Returns:
+        Tuple of (is_healthy: bool, error_message: Optional[str])
+    """
+    import aiohttp
+    vision_url = get_vision_gpu_url()
+    
+    try:
+        async with aiohttp.ClientSession() as session:
+            async with session.get(f"{vision_url}/health", timeout=aiohttp.ClientTimeout(total=5)) as response:
+                is_healthy = response.status == 200
+                if is_healthy:
+                    logger.info(f"Vision endpoint ({vision_url}) health check: OK")
+                else:
+                    logger.warning(f"Vision endpoint ({vision_url}) health check failed: status {response.status}")
+                return is_healthy, None if is_healthy else f"Status {response.status}"
+    except asyncio.TimeoutError:
+        logger.error(f"Vision endpoint ({vision_url}) health check: timeout")
+        return False, "Endpoint timeout"
+    except Exception as e:
+        logger.error(f"Vision endpoint ({vision_url}) health check error: {e}")
+        return False, str(e)
 
 def _strip_surrounding_quotes(text):
     """
@@ -108,8 +147,12 @@ async def query_llama(user_prompt, user_id, guild_id=None, response_type="dm_res
         if evil_mode:
             model = globals.EVIL_TEXT_MODEL  # Use DarkIdol uncensored model
             logger.info(f"Using evil model: {model}")
+        elif globals.LANGUAGE_MODE == "japanese":
+            model = globals.JAPANESE_TEXT_MODEL  # Use Swallow for Japanese
+            logger.info(f"Using Japanese model: {model}")
         else:
             model = globals.TEXT_MODEL
+            logger.info(f"Using default model: {model}")
     
     # Determine channel_id for conversation history
     # For servers, use guild_id; for DMs, use user_id

llama-swap-config.yaml

@@ -18,6 +18,15 @@ models:
       - darkidol
       - evil-model
       - uncensored
+  
+  # Japanese language model (Llama 3.1 Swallow - Japanese optimized)
+  swallow:
+    cmd: /app/llama-server --port ${PORT} --model /models/Llama-3.1-Swallow-8B-Instruct-v0.5-Q4_K_M.gguf -ngl 99 -nkvo -c 16384 --host 0.0.0.0 --no-warmup
+    ttl: 1800  # Unload after 30 minutes of inactivity
+    aliases:
+      - swallow
+      - japanese
+      - japanese-model
     
   # Vision/Multimodal model (MiniCPM-V-4.5 - supports images, video, and GIFs)
   vision:

llama-swap-rocm-config.yaml

@@ -19,6 +19,15 @@ models:
       - darkidol
       - evil-model
       - uncensored
+  
+  # Japanese language model (Llama 3.1 Swallow - Japanese optimized)
+  swallow:
+    cmd: /app/llama-server --port ${PORT} --model /models/Llama-3.1-Swallow-8B-Instruct-v0.5-Q4_K_M.gguf -ngl 99 -nkvo -c 16384 --host 0.0.0.0 --no-warmup
+    ttl: 1800  # Unload after 30 minutes of inactivity
+    aliases:
+      - swallow
+      - japanese
+      - japanese-model
 
 # Server configuration
 # llama-swap will listen on this address