updated docker file

Dockerfile

@@ -1,4 +1,5 @@
 # Hugging Face Spaces - MrrrMe with Coqui XTTS v2 + MODULAR BACKEND
+# FIXED: Pre-downloads emotion2vec during build for fast startup
 FROM nvidia/cuda:11.8.0-cudnn8-runtime-ubuntu22.04
 
 # Install system dependencies
@@ -54,9 +55,25 @@ ENV COQUI_TOS_AGREED=1
 COPY requirements_docker.txt ./
 RUN python3.11 -m pip install --no-cache-dir -r requirements_docker.txt
 
+# ============================================================
+# PRE-DOWNLOAD MODELS DURING BUILD (for fast startup!)
+# ============================================================
+
 # PRE-DOWNLOAD XTTS V2 MODEL (Memory Safe)
 # Uses ModelManager to download without loading to RAM (Fixes Exit 137)
-RUN python3.11 -c "from TTS.utils.manage import ModelManager; print('⏳ Downloading XTTS v2 model...'); ModelManager().download_model('tts_models/multilingual/multi-dataset/xtts_v2'); print('✅ Download complete.')"
+RUN python3.11 -c "from TTS.utils.manage import ModelManager; print('⏳ Downloading XTTS v2 model...'); ModelManager().download_model('tts_models/multilingual/multi-dataset/xtts_v2'); print('✅ XTTS v2 downloaded.')"
+
+# PRE-DOWNLOAD emotion2vec+ Large (1.8GB) - FIXES SLOW STARTUP!
+# Without this, the model downloads on every restart (5+ minutes)
+RUN python3.11 -c "\
+import os; \
+os.makedirs('/root/.cache/modelscope', exist_ok=True); \
+os.environ['MODELSCOPE_CACHE'] = '/root/.cache/modelscope'; \
+os.environ['MS_CACHE_HOME'] = '/root/.cache/modelscope'; \
+from funasr import AutoModel; \
+print('⏳ Downloading emotion2vec+ Large (1.8GB)...'); \
+model = AutoModel(model='iic/emotion2vec_plus_large', hub='ms'); \
+print('✅ emotion2vec+ Large downloaded.')"
 
 # Install avatar dependencies
 RUN python3.11 -m pip install --no-cache-dir \
@@ -98,17 +115,20 @@ RUN mkdir -p /etc/nginx/certs && \
     -days 365 \
     -subj "/CN=mrrrme.hf.space"
 
+# ============================================================
+# COPY CACHED MODELS TO USER HOME (runtime user is 1000)
+# ============================================================
+RUN mkdir -p /home/user/.cache && \
+    cp -r /root/.cache/modelscope /home/user/.cache/ 2>/dev/null || true && \
+    cp -r /root/.local /home/user/ 2>/dev/null || true && \
+    chown -R 1000:1000 /home/user
+
 # ✅ Create startup script with MODULAR BACKEND + MODEL CACHING
-RUN printf '#!/bin/bash\nset -e\nexport HOME=/home/user\nmkdir -p /tmp\n\n# Agree to TOS\nexport COQUI_TOS_AGREED=1\n\n# ============================================================\n# MODEL CACHING - emotion2vec downloads once, then cached\n# ============================================================\nexport MODELSCOPE_CACHE=/data/modelscope\nexport MS_CACHE_HOME=/data/modelscope\nmkdir -p /data/modelscope 2>/dev/null || true\n\nif [ -d "/data" ] && [ -w "/data" ]; then\n    echo "Persistent storage: /data"\n    echo "Model cache: /data/modelscope"\n    chmod 777 /data 2>/dev/null || true\n    chmod 777 /data/modelscope 2>/dev/null || true\nelse\n    echo "Ephemeral storage: /tmp"\n    export MODELSCOPE_CACHE=/tmp/modelscope\n    export MS_CACHE_HOME=/tmp/modelscope\n    mkdir -p /tmp/modelscope\nfi\n\npkill -f "backend_new.py" 2>/dev/null || true\npkill -f "speak_server.py" 2>/dev/null || true\npkill -f "node server.js" 2>/dev/null || true\npkill -f "nginx" 2>/dev/null || true\n\nsleep 2\necho "Starting MrrrMe (XTTS v2 + Modular Backend v2.0)..."\n\n# Start NEW modular backend\ncd /app && python3.11 mrrrme/backend_new.py &\n\n# Start avatar TTS\ncd /app/avatar && python3.11 speak_server.py &\n\n# Start Next.js frontend\ncd /app/frontend/.next/standalone && HOSTNAME=0.0.0.0 PORT=3001 node server.js &\n\nsleep 10\nnginx -g "daemon off;" &\necho "Ready!"\nwait\n' > /app/start.sh && chmod +x /app/start.sh
+RUN printf '#!/bin/bash\nset -e\nexport HOME=/home/user\nmkdir -p /tmp\n\n# Agree to TOS\nexport COQUI_TOS_AGREED=1\n\n# ============================================================\n# MODEL CACHING - emotion2vec pre-downloaded in Docker build\n# ============================================================\n# Model is at /home/user/.cache/modelscope (from build)\n# Also try /data/modelscope for persistent storage\nexport MODELSCOPE_CACHE=/home/user/.cache/modelscope\nexport MS_CACHE_HOME=/home/user/.cache/modelscope\n\n# Create /data dirs if writable (for other runtime caching)\nif [ -d "/data" ] && [ -w "/data" ]; then\n    echo "Persistent storage available: /data"\n    mkdir -p /data/modelscope 2>/dev/null || true\nfi\n\npkill -f "backend_new.py" 2>/dev/null || true\npkill -f "speak_server.py" 2>/dev/null || true\npkill -f "node server.js" 2>/dev/null || true\npkill -f "nginx" 2>/dev/null || true\n\nsleep 2\necho "Starting MrrrMe (XTTS v2 + Modular Backend v2.0)..."\necho "Models pre-loaded from Docker build - fast startup!"\n\n# Start NEW modular backend\ncd /app && python3.11 mrrrme/backend_new.py &\n\n# Start avatar TTS\ncd /app/avatar && python3.11 speak_server.py &\n\n# Start Next.js frontend\ncd /app/frontend/.next/standalone && HOSTNAME=0.0.0.0 PORT=3001 node server.js &\n\nsleep 10\nnginx -g "daemon off;" &\necho "Ready!"\nwait\n' > /app/start.sh && chmod +x /app/start.sh
 
 # Set ownership
 RUN chown -R 1000:1000 /app
 
-# Ensure non-root user has access to models
-RUN mkdir -p /home/user/.local && \
-    cp -r /root/.local/share /home/user/.local/ || true && \
-    chown -R 1000:1000 /home/user
-
 USER 1000
 ENV HOME=/home/user
 

mrrrme/audio/voice_emotion.py

@@ -21,14 +21,27 @@ import webrtcvad
 from ..config import AUDIO_SR, AUDIO_BLOCK, CLIP_SECONDS, VAD_AGGRESSIVENESS, FUSE4
 
 # ============================================================
-# CACHE DIRECTORY - MUST match Dockerfile ENV variables!
+# CACHE DIRECTORY - Check multiple locations
 # ============================================================
-# Dockerfile sets: ENV MODELSCOPE_CACHE=/data/modelscope
-# So we MUST use the same path here!
-if os.path.exists('/data') and os.access('/data', os.W_OK):
-    CACHE_DIR = '/data/modelscope'  # ✅ FIXED: Match Dockerfile
+# Priority order:
+# 1. /home/user/.cache/modelscope (Docker build location - FASTEST)
+# 2. /data/modelscope (HF Spaces persistent storage)
+# 3. /tmp/modelscope (fallback)
+
+# Check where the model actually exists (from Docker build)
+DOCKER_BUILD_CACHE = '/home/user/.cache/modelscope'
+PERSISTENT_CACHE = '/data/modelscope'
+FALLBACK_CACHE = '/tmp/modelscope'
+
+# Use Docker build cache if it exists (pre-downloaded during build)
+if os.path.exists(os.path.join(DOCKER_BUILD_CACHE, 'hub', 'iic')) or \
+   os.path.exists(os.path.join(DOCKER_BUILD_CACHE, 'hub', 'models', 'iic')):
+    CACHE_DIR = DOCKER_BUILD_CACHE
+    print(f"[VoiceEmotion] Using Docker build cache: {CACHE_DIR}")
+elif os.path.exists('/data') and os.access('/data', os.W_OK):
+    CACHE_DIR = PERSISTENT_CACHE
 else:
-    CACHE_DIR = '/tmp/modelscope'   # Fallback for local dev
+    CACHE_DIR = FALLBACK_CACHE
 
 os.makedirs(CACHE_DIR, exist_ok=True)
 
@@ -36,7 +49,7 @@ os.makedirs(CACHE_DIR, exist_ok=True)
 os.environ['MODELSCOPE_CACHE'] = CACHE_DIR
 os.environ['MS_CACHE_HOME'] = CACHE_DIR
 os.environ['FUNASR_CACHE'] = CACHE_DIR
-os.environ['HF_HOME'] = CACHE_DIR  # Some models also use HuggingFace
+os.environ['HF_HOME'] = os.path.dirname(CACHE_DIR)  # Parent dir for HF
 
 
 class VoiceEmotionWorker:
@@ -84,14 +97,18 @@ class VoiceEmotionWorker:
             os.path.join(CACHE_DIR, 'hub', 'iic'),
             os.path.join(CACHE_DIR, 'iic'),
             os.path.join(CACHE_DIR, 'hub', 'models', 'iic'),
+            # Also check default location (Docker build)
+            '/home/user/.cache/modelscope/hub/iic',
+            '/home/user/.cache/modelscope/hub/models/iic',
         ]
         
         model_cached = any(os.path.exists(p) for p in possible_cache_paths)
         
         if model_cached:
-            print("[VoiceEmotion] ✅ Model found in cache - loading from disk")
+            print("[VoiceEmotion] ✅ Model found in cache - loading from disk (FAST)")
         else:
-            print("[VoiceEmotion] ⏳ Model not cached - downloading (one-time, ~1.8GB)")
+            print("[VoiceEmotion] ⚠️ Model not cached - will download (~1.8GB, 5+ min)")
+            print("[VoiceEmotion] 💡 Tip: Add pre-download to Dockerfile for faster startup")
         
         # Load emotion2vec model
         try:
@@ -100,14 +117,14 @@ class VoiceEmotionWorker:
             load_start = time.time()
             
             # ============================================================
-            # FIX: Remove model_revision to use latest compatible version
-            # The v2.0.4 revision was causing decoder weight mismatches
-            # which resulted in uniform 0.25 probability distributions
+            # FIXED: 
+            # - Removed model_revision (was causing decoder weight mismatch)
+            # - disable_update=True since model is pre-downloaded in Docker
             # ============================================================
             self.model = AutoModel(
                 model="iic/emotion2vec_plus_large",
-                # model_revision="v2.0.4",  # ❌ REMOVED - causes decoder mismatch!
-                disable_update=model_cached,  # Only disable updates if already cached
+                # model_revision removed - use latest compatible version
+                disable_update=True,  # Model pre-downloaded in Docker build
                 hub="ms"  # Explicitly use ModelScope hub
             )