Fix macOS MPS compatibility in test server startup

Add PYTORCH_ENABLE_MPS_FALLBACK=1 environment variable when starting
test servers to allow PyTorch operations to fall back to CPU when
MPS (Metal Performance Shaders) doesn't support them.

This fixes test failures on macOS with Apple Silicon where operations
like aten::linalg_cholesky_ex.L are not implemented for MPS device.

Author: claude

tests/test_conversation_logging_server.py

@@ -94,6 +94,8 @@ def _start_server_with_logging(cls):
         env["OPTILLM_API_KEY"] = "optillm"
         env["OPTILLM_LOG_CONVERSATIONS"] = "true"
         env["OPTILLM_CONVERSATION_LOG_DIR"] = str(cls.temp_log_dir)
+        # Enable MPS fallback to CPU for unsupported operations (fixes macOS compatibility)
+        env["PYTORCH_ENABLE_MPS_FALLBACK"] = "1"
 
         # Get the project root directory (parent of tests directory)
         project_root = os.path.dirname(os.path.dirname(os.path.abspath(__file__)))

tests/test_utils.py

@@ -42,6 +42,8 @@ def start_test_server(model: str = TEST_MODEL, port: int = 8000) -> subprocess.P
     # Set environment for local inference
     env = os.environ.copy()
     env["OPTILLM_API_KEY"] = "optillm"
+    # Enable MPS fallback to CPU for unsupported operations (fixes macOS compatibility)
+    env["PYTORCH_ENABLE_MPS_FALLBACK"] = "1"
 
     # Get the project root directory (parent of tests directory)
     project_root = os.path.dirname(os.path.dirname(os.path.abspath(__file__)))