2025-11 monthly performance summary for pytorch/pytorch focusing on key developer achievements and business impact. This month delivered major CUDA kernel optimizations for sorting, unique, and EmbeddingBag, achieving substantial speedups across NVIDIA GPUs while maintaining API compatibility. The work targeted critical data-paths used by common ML workloads, improving end-to-end throughput and reducing GPU resource pressure. Cross-GPU validation (H100/H20) and multiple CUDA versions (12.x–13.x) confirmed robustness and scalability.

October 2025 ROCm/pytorch performance optimization: fixed a major GPU performance regression in torch.topk by introducing a dedicated histogram+cumsum kernel (computeDigitCumSum) and refactoring the top-k path to use it. This eliminated redundant global memory reads and improved large-input throughput. The changes include loop unrolling in computeDigitCumSum and updating computeBlockwiseWithinKCounts to rely on the new kernel, while preserving correctness across inputs. Key commit: 3cc8af2d67f42bf2a933796290446c5ab8978aac; PR164459 merged with approvals from core maintainers ngimel and Skylion007. Benchmarks on NVIDIA H20 show substantial gains for large tensors: for example, 1B input top-100 now runs in ~25.6 ms vs 36.6 ms on 2.6.0 and 1564.1 ms on 2.8.0, illustrating the regression fix and throughput improvement; 100M input improves from 17.4 ms (2.8.0) to ~2.54 ms with the PR. The PR also reports 1,000,000 and 512x128000 scales with tight performance, and confirms correctness across varied shapes.