Enable Whisper medium cache using /data persistent storage (HF Pro)

Dockerfile

@@ -21,9 +21,11 @@ RUN pip install --no-cache-dir -r requirements.txt
 # Create cache directory for models BEFORE copying code
 # This ensures model downloads are cached even when code changes
 RUN mkdir -p /.cache && chmod -R 777 /.cache
+RUN mkdir -p /data/.cache && chmod -R 777 /data/.cache
 ENV HF_HOME=/.cache
-ENV TORCH_HOME=/.cache
+ENV TORCH_HOME=/data/.cache
 ENV XDG_CACHE_HOME=/.cache
+ENV WHISPER_CACHE=/data/.cache
 
 # Pre-download models during build (HF Pro with persistent storage)
 # These layers will be CACHED and won't rebuild when only code changes
@@ -35,13 +37,15 @@ RUN python -c "from transformers import AutoTokenizer, AutoModelForSequenceClass
     AutoModelForSequenceClassification.from_pretrained('Cyberlace/swara-structure-model', cache_dir='/.cache'); \
     print('✅ Structure Model cached!')" && chmod -R 777 /.cache
 
-# 2. Whisper medium will be downloaded on FIRST REQUEST (lazy loading)
-# Build OOM with 1.5GB download - HF Space has limited build memory
-# First request will take ~2-3 min extra for one-time download
-# RUN python -c "import whisper; \
-#     print('📥 Downloading Whisper medium model (1.5GB)...'); \
-#     whisper.load_model('medium', download_root='/.cache'); \
-#     print('✅ Whisper medium cached!')" && chmod -R 777 /.cache
+# 2. Download Whisper medium model (~1.5GB) 
+# Using /data for HF Pro Persistent Storage (survives restarts)
+RUN mkdir -p /data/.cache && \
+    python -c "import whisper, os; \
+    os.environ['TORCH_HOME'] = '/data/.cache'; \
+    print('📥 Downloading Whisper medium to persistent storage...'); \
+    whisper.load_model('medium', download_root='/data/.cache'); \
+    print('✅ Whisper medium cached!')" && \
+    chmod -R 777 /data/.cache
 
 # 3. Download Sentence Transformer for Keywords (~420MB)
 RUN python -c "from sentence_transformers import SentenceTransformer; \

app/services/speech_to_text.py

@@ -29,7 +29,8 @@ class SpeechToTextService:
             print(f"💻 Using device: {self.device}")
         
         # Check if model is already cached
-        cache_dir = os.environ.get('XDG_CACHE_HOME', '/.cache')
+        # Use /data/.cache for Whisper (persistent storage on HF Pro)
+        cache_dir = os.environ.get('WHISPER_CACHE', '/data/.cache')
         model_cache_path = os.path.join(cache_dir, f'{model_name}.pt')
         
         # Load Whisper model

start.sh

@@ -6,6 +6,7 @@ echo "=========================================="
 
 # Fix cache permissions on startup (in case of permission issues)
 chmod -R 777 /.cache 2>/dev/null || true
+chmod -R 777 /data/.cache 2>/dev/null || true
 
 # Fix OpenMP warning - set proper thread count
 export OMP_NUM_THREADS=4