SplitLayer

[Semyon1104]

No description provided.

[aobolensk]

Consider using std::accumulate

size_t outer_size = std::accumulate(
    shape.begin(), shape.begin() + axis, 
    static_cast<size_t>(1), std::multiplies<size_t>());

size_t inner_size = std::accumulate(
    shape.begin() + axis + 1, shape.end(), 
    static_cast<size_t>(1), std::multiplies<size_t>());

[Semyon1104]

we do not have access to the vector<size_t> dims_ inside the Shape class, as it is a private field. and therefore we cannot use iterators

[aobolensk]

OK, let's leave as is

[aobolensk]

Suggested change

	int num_outputs_ = 0;
	int num_outputs_;

[aobolensk]

This still can have underflow. For example: axis = -5, rank = 2

[aobolensk]

What is the purpose of this?

[Semyon1104]

this is for the ability to split into equal parts if splits_ are not specified

[aobolensk]

Please, add negative axis tests

[allnes]

Use block copies where memory is contiguous. For each output slice at fixed outer/part, the data region is contiguous (output_axis_size * inner_size elements). Replace the inner element-wise loops with a single std::copy_n (or memcpy when types allow). This will significantly reduce overhead on larger tensors.

[allnes]

Make the mode explicit. splits_ vs num_outputs_ are mutually exclusive. Consider:
• std::variant<std::vector, int> split_spec_; or
• std::optional num_outputs_;
This encodes invariants in the type system and simplifies validation.

[allnes]

Tests to add
• Uneven split (remainder) test. Verify that the entire remainder lands in the last output (per the chosen policy).
• num_outputs_ > size test. Either expect an exception (if you forbid it) or assert the exact shapes including zero-sized outputs (if you allow it).
• Integer dtype path. Add at least one test covering Type::kInt to ensure both code paths behave identically.

[allnes]

Behavior when num_outputs_ > size_along_axis is undefined.
Today this yields multiple zero-sized outputs. Decide and enforce one policy:
• either reject with a clear exception in validation, or
• allow but document it and add tests asserting exact output shapes (including zeros).
Leaving this implicit can easily break downstream code that doesn’t expect zero-length tensors.