Contributed a memory-optimized backward pass for the fused linear cross entropy kernel in the linkedin/Liger-Kernel repository, focusing on reducing memory usage during backpropagation for large vocabulary models. The solution replaced the traditional grad_weight accumulation pattern with an in-place torch.addmm operation supporting out_dtype, which eliminated large intermediate allocations and reduced peak memory footprint on CUDA-enabled PyTorch 2.8 and newer. This approach maintained compatibility with older PyTorch versions, ensuring broad usability. The work involved performance benchmarking, targeted testing with pytest, and code style validation, demonstrating proficiency in CUDA, PyTorch, and Python for performance optimization and memory-efficient deep learning workflows.