RuntimeWarning: nanobind: type '_Globals' was already registered!

[banach-space]

Hey,

I sadly I am unable to run examples from Lighthous, e.g. mlp_from_file.py. Here's a diff I need to apply for things to work on my AArch64 machine (*):

diff --git a/pyproject.toml b/pyproject.toml
index ba0cba7..b78aef5 100644
--- a/pyproject.toml
+++ b/pyproject.toml
@@ -10,7 +10,7 @@ dependencies = [
 # Additional "targets" which pull in optional dependencies -- use `uv sync --extra TARGET`
 ingress_torch_cpu = [
     "torch==v2.8.0+cpu",
-    "torch-mlir==20251003.589"
+    "torch-mlir==20251116.633"
 ]
 ingress_torch_nvidia = [
     "torch==2.8.0",  # Nvidia-enabled version of torch

Here's a mini Python repro:

from mlir import ir
from torch_mlir  import fx
print(ir.Context)

Copy that into Lighthouse checkout. Then,

$ uv venv
$ uv sync
$ uv sync --extra ingress-torch-cpu
$ uv run python temp.py
<frozen importlib._bootstrap>:488: RuntimeWarning: nanobind: type '_Globals' was already registered!

<frozen importlib._bootstrap>:488: RuntimeWarning: nanobind: type 'DiagnosticSeverity' was already registered!

Critical nanobind error: refusing to add duplicate key "ERROR" to enumeration "mlir._mlir_libs._mlir.ir.DiagnosticSeverity"!

IIUC, this error caused by some global state from MLIR Python bindings being registered twice.

Kind request, could somebody try this on X86? Thanks!

-Andrzej

(*) Updating torch-mlir version as the AArch64 releases are a fairly recent addition.

[rolfmorel]

Cannot reproduce (with torch-mlir==20251003.589 it's fine for me) right now as the torch-mlir-release repo does not have the x86_64 linux packages for the latest releases at the moment: llvm/torch-mlir-release#24

[banach-space]

@rolfmorel Looks like a generic error (as opposed to being AArch64 specific):

• https://github.com/llvm/lighthouse/actions/runs/19565684835/job/56027057471

[rolfmorel]

#22 's CI demonstrates that tests now succeed on x86 and Arm while making use of (close to) latest torch-mlir and mlir-python-bindings.