A report published on March 4, 2026 says splitting 256 SIGReg slices into chunks of 32 used eight times less VRAM and finished faster on a 4090 MaxQ GPU.
A report published on March 4, 2026 describes an optimization for LeJEPA: splitting 256 SIGReg slices into chunks of 32. The author says the test used eight times less VRAM and finished faster on a 4090 MaxQ GPU. The text identifies large complex-exponential tensors as the bottleneck; it gives no absolute memory use, timings, or other benchmark results.
The observation may help people working with LeJEPA on memory-limited GPUs, but it is a reported result, not a guarantee for other configurations. Consult the original publication and, before adopting the change, compare the implementation with the unchunked version, recording VRAM and runtime on the same hardware and under the same conditions. If you use AI to study or apply the method, avoid submitting code or internal data without authorization and follow your organization’s data-protection policy.