On large multiclass datasets such as dionis with C=355 classes, out-of-memory errors can occur in the following cases:

1. Fitting stacker models with N base models.
2. Fitting weighted ensembles with N base models.

In both cases this results in X having an extra N*C features, or an extra N*C*4 bytes per sample S (4 because of use of float32). This can lead to cases where AutoGluon works fine on lower time limits, but on higher time limits it fits more base models and then gets an out-of-memory error due to too many features in X in the stack layers.

For example, given S=1,000,000 rows of data and C=355 classes, each base model would increase memory usage of X by 1.42 GB.
On a system with 32 GB of memory, if 25 base models were used, this would require 35.5 GB of memory, leading to an out-of-memory error.

• Logic should be added that computes this expected memory usage via S*C*4 bytes.
• Next, the logic will define N where N is the number of base models allowed for memory safety. (The logic could be something like N*S*C*4 bytes + X_og_mem_usage < 25% of memory).
• Then, base models can be sorted by validation score, and the top N are used, with the rest discarded.
• The logic could mimic the logic in AbstractTrainer when infer_limit is specified.