[deep1blue]

What does this PR do?

When ModelSummary (or DeepSpeedSummary, which inherits total_layer_params) is initialized with max_depth > 1 or max_depth == -1, self._layer_summary contains both top-level child modules (e.g. net) and their nested submodules (e.g. net.0, net.1).

Because each LayerSummary.num_parameters calls self._module.parameters() recursively, a parent module's num_parameters already includes all parameters of its nested submodules. Summing self.param_nums across all entries in self._layer_summary in ModelSummary.total_layer_params therefore double/triple-counts nested submodule parameters whenever max_depth > 1 or max_depth == -1. As a consequence:

1. summary.total_layer_params exceeds summary.total_parameters on any nested model with max_depth != 1.
2. In _get_summary_data(), total_leftover_params = self.total_parameters - self.total_layer_params becomes <= 0, silently dropping the "other params" (LEFTOVER_PARAMS_NAME) row for direct nn.Parameter attributes on the LightningModule whenever max_depth > 1 or max_depth == -1.

Changes

• Compute ModelSummary.total_layer_params by summing layer.num_parameters only across top-level child modules ("." not in name) in self._layer_summary.
• Parameterize test_summary_data_with_non_layer_params in tests/tests_pytorch/utilities/test_model_summary.py over max_depth in [-1, 1, 2] with a nested submodule (nn.Sequential(nn.Linear(2, 2))) to verify that total_layer_params and the "other params" row remain accurate at all depths.

Acknowledgments

Credit to @awaelchli, @Borda, @carmocca, @SkafteNicki, and @GdoongMathew for their work on ModelSummary and DeepSpeedSummary.