Fix CPU TinyLlama PyTorch initialization

app.py

@@ -1,4 +1,7 @@
 import os
+# Keep all model inference on CPU; this must be set before importing torch.
+os.environ["CUDA_VISIBLE_DEVICES"] = ""
+os.environ["CUDA_DEVICE_ORDER"] = "PCI_BUS_ID"
 import platform
 import sys
 import time
@@ -549,8 +552,11 @@ def _load_tinyllama():
         if _TINYLLAMA is not None:
             return _TINYLLAMA_TOKENIZER, _TINYLLAMA
         from transformers import AutoModelForCausalLM, AutoTokenizer
-        torch.set_num_threads(min(_parse_cpu_count() or 1, 16))
-        torch.set_num_interop_threads(1)
+        try:
+            torch.set_num_threads(min(_parse_cpu_count() or 1, 16))
+        except RuntimeError:
+            # Gradio may already have started PyTorch work during app startup.
+            pass
         tokenizer = AutoTokenizer.from_pretrained(_TINYLLAMA_REPO)
         model = AutoModelForCausalLM.from_pretrained(
             _TINYLLAMA_REPO,