fix: pass HF_TOKEN to pipeline calls, add retry for rate limits

Dockerfile

@@ -16,7 +16,6 @@ COPY . .
 # Pre-create HF / torch cache dirs writable by appuser
 RUN mkdir -p /home/appuser/.cache && chown -R appuser:appuser /home/appuser /app
 ENV HF_HOME=/home/appuser/.cache/huggingface
-ENV TRANSFORMERS_CACHE=/home/appuser/.cache/huggingface/hub
 
 USER appuser
 

main.py

@@ -77,20 +77,41 @@ arabic_classifier = None        # dedicated Arabic hate-speech model
 llm_client: InferenceClient | None = None  # HF Inference API client
 
 
+def _load_pipeline_with_retry(task: str, model: str, max_retries: int = 5):
+    """Load a HF pipeline with exponential backoff for rate-limit errors."""
+    import time
+
+    for attempt in range(1, max_retries + 1):
+        try:
+            return pipeline(
+                task,
+                model=model,
+                tokenizer=model,
+                token=HF_TOKEN,
+            )
+        except (OSError, ValueError) as exc:
+            if "429" in str(exc) and attempt < max_retries:
+                wait = 2 ** attempt          # 2, 4, 8, 16, 32 s
+                logger.warning(
+                    "Rate-limited loading %s (attempt %d/%d). "
+                    "Retrying in %ds …",
+                    model, attempt, max_retries, wait,
+                )
+                time.sleep(wait)
+            else:
+                raise
+
+
 @asynccontextmanager
 async def lifespan(app: FastAPI):
     """Load both AI models when the server starts; release on shutdown."""
     global toxicity_classifier, arabic_classifier, llm_client
 
-    toxicity_classifier = pipeline(
-        "text-classification",
-        model=MODEL_NAME,
-        tokenizer=MODEL_NAME,
+    toxicity_classifier = _load_pipeline_with_retry(
+        "text-classification", MODEL_NAME,
     )
-    arabic_classifier = pipeline(
-        "text-classification",
-        model=ARABIC_MODEL_NAME,
-        tokenizer=ARABIC_MODEL_NAME,
+    arabic_classifier = _load_pipeline_with_retry(
+        "text-classification", ARABIC_MODEL_NAME,
     )
 
     # Set up the LLM client (optional – works without a token at lower rate)