pin numpy to < 2.0 until tensorboard cuts a release

Summary: The latest TB version hasn't incorporated numpy2.0-compatible changes (tensorflow/tensorboard#6869 (comment)), and recently unpinned botorch (D60735276) means Ax CI is now failing. This pins numpy to < 2.0 to prevent issues in CI.

Differential Revision: D60917879

ax/metrics/tensorboard.py

@@ -228,3 +228,14 @@ def _get_event_multiplexer_for_trial(
         "TensorboardMetric, please install tensorboard."
     )
     pass
+except AttributeError as e:
+    if (
+        "`np.string_` was removed in the NumPy 2.0 release. Use `np.bytes_` instead."
+        in repr(e)
+    ):
+        logger.warning(
+            "tensorboard package not compatible with numpy 2.0+. "
+            "Please install numpy < 2.0."
+        )
+    else:
+        raise e

setup.py

@@ -53,6 +53,9 @@
 NOTEBOOK_REQUIRES = ["jupyter"]
 
 UNITTEST_MINIMAL_REQUIRES = [
+    # Replace with `tensorboard >= x.x` once tb cuts a release.
+    # https://github.com/tensorflow/tensorboard/issues/6869#issuecomment-2273718763
+    "numpy<2.0",
     "tensorboard",  # For tensorboard unit tests.
     "torchvision",  # For torchvision unit tests.
     "torchx",  # For torchx unit tests.