Fix cache permissions: use /tmp/hf with proper Dockerfile

Dockerfile

@@ -1,11 +1,26 @@
 FROM pytorch/pytorch:2.1.2-cuda11.8-cudnn8-runtime
+
 WORKDIR /app
+
+# Create cache directory with proper permissions
+RUN mkdir -p /tmp/hf && chmod 777 /tmp/hf
+
+# Set environment variables for model caching
+ENV HF_HOME=/tmp/hf
+ENV TRANSFORMERS_CACHE=/tmp/hf
+ENV HUGGINGFACE_HUB_CACHE=/tmp/hf
+
+# Install Python dependencies
 RUN pip install --no-cache-dir --upgrade pip
 COPY requirements.txt /app/requirements.txt
 RUN pip install --no-cache-dir -r requirements.txt
+
+# Copy application code
 COPY app.py /app/app.py
+
+# Expose port
 ENV PORT=7860
 EXPOSE 7860
-RUN mkdir -p /data/hf
-ENV HF_HOME=/data/hf TRANSFORMERS_CACHE=/data/hf HUGGINGFACE_HUB_CACHE=/data/hf
+
+# Run the application
 CMD ["uvicorn", "app:app", "--host", "0.0.0.0", "--port", "7860"]

app.py

@@ -14,6 +14,15 @@ DEVICE = torch.device("cuda" if torch.cuda.is_available() else "cpu")
 DEFAULT_MAX_LEN = int(os.getenv("MAX_LEN", "256"))
 DEFAULT_BATCH_SIZE = int(os.getenv("BATCH_SIZE", "16"))
 
+# Use /tmp for model cache (always writable in containers)
+HF_CACHE_DIR = "/tmp/hf"
+os.makedirs(HF_CACHE_DIR, exist_ok=True)
+
+# Also set env vars commonly used by transformers and huggingface_hub
+os.environ.setdefault("HF_HOME", HF_CACHE_DIR)
+os.environ.setdefault("TRANSFORMERS_CACHE", HF_CACHE_DIR)
+os.environ.setdefault("HUGGINGFACE_HUB_CACHE", HF_CACHE_DIR)
+
 
 class DesklibAIDetectionModel(PreTrainedModel):
     config_class = AutoConfig
@@ -36,8 +45,8 @@ class DesklibAIDetectionModel(PreTrainedModel):
 
 
 def load_model():
-    tokenizer = AutoTokenizer.from_pretrained(MODEL_ID)
-    model = DesklibAIDetectionModel.from_pretrained(MODEL_ID)
+    tokenizer = AutoTokenizer.from_pretrained(MODEL_ID, cache_dir=HF_CACHE_DIR)
+    model = DesklibAIDetectionModel.from_pretrained(MODEL_ID, cache_dir=HF_CACHE_DIR)
     model.to(DEVICE)
     model.eval()