| About the performance category |  | 0 | 771 | January 22, 2021 |
| Performance of "activation" sparsity |  | 1 | 212 | April 21, 2025 |
| The "Ideal" PyTorch FLOP Counter (with __torch_dispatch__) |      | 20 | 16680 | August 22, 2024 |
| Fast combined C++/Python/TorchScript/Inductor tracebacks |     | 4 | 3106 | August 17, 2023 |
| Estimate theoritical FLOPs of backward pass of a DNN |  | 1 | 1123 | April 16, 2023 |
| Performance gains w/ nanoGPT using SDPA Custom Kernel |  | 0 | 4724 | January 30, 2023 |
| Making Transformer inference faster on GPUs |   | 1 | 3993 | December 16, 2022 |
| Investigation report: what would it cost to optimize c10::intrusive_ptr destruction for refcount == 1? (A: too much) |  | 0 | 718 | May 3, 2022 |
| Working With `c10::IValue` Efficiently | | 0 | 2719 | April 18, 2022 |
| Unionizing for Profit: How to Exploit the Power of Unions in C++ |  | 2 | 3360 | January 7, 2022 |
| CUDA loops case study: code generation vs templates |   | 4 | 2584 | December 12, 2021 |
| Multiple workers for single batch |   | 2 | 1175 | July 14, 2021 |
| Optimizing contiguous() for the case where the Tensor is_contiguous()? |   | 6 | 1755 | May 24, 2021 |
| Converting weights flat buffer |  | 0 | 902 | March 25, 2021 |
| Why `torch::jit::pop` (and sometimes push) is worse for performance than direct `std::vector` access | | 0 | 1122 | March 17, 2021 |
| Pytorch Benchmarks issues with general usability and issues with individual benchmarks |    | 2 | 1334 | March 12, 2021 |
| GPU Overheads and Fused Strassen |  | 0 | 2255 | February 13, 2021 |
| Comparing the performance of 0.4.1 and master | | 0 | 2364 | February 9, 2021 |
| Overhead in `nn.Module` causing massive slowdowns compared to raw CuBLAS or Torchscript | | 0 | 1694 | January 28, 2021 |
| Dispatcher Performance and Inlining: a Report on Two Days Spent on Dispatcher Performance |  | 1 | 1223 | January 28, 2021 |
| We shouldn't feel bad about passing `Tensor` by reference | | 0 | 2368 | January 25, 2021 |