When running locally (even when running in the same singularity image), the networks seem to work reliably.
When run on Alvis, the networks yield plenty of undetermined episodes which run for longer than max_steps
Try to check if the loaded weights/parameters are the same.
Next, try checking the data generation. Many users who reported similar issues were able to trace the problem back to the input data.
When running locally (even when running in the same singularity image), the networks seem to work reliably. When run on Alvis, the networks yield plenty of undetermined episodes which run for longer than max_steps