From 6ed8688bb64ae08b70aff65ea6d3ec73fd6b1e71 Mon Sep 17 00:00:00 2001 From: Ian Zhang Date: Sun, 22 Sep 2024 22:32:23 -0700 Subject: [PATCH] Fix typo(s) and CUDA extensions URLs in README.md --- README.md | 8 ++++---- 1 file changed, 4 insertions(+), 4 deletions(-) diff --git a/README.md b/README.md index a6d8b72..80fc671 100644 --- a/README.md +++ b/README.md @@ -2,7 +2,7 @@ [OCP MX Formats Specification](https://www.opencompute.org/documents/ocp-microscaling-formats-mx-v1-0-spec-final-pdf) -This library provides the capability to emulate MX-compatble formats +This library provides the capability to emulate MX-compatible formats and bfloat quantization in pytorch, enabling data science exploration for DNNs with different MX formats. The underlying computations are done in float32/bfloat16/fp16 but with values restricted to @@ -272,9 +272,9 @@ MX library functions. These are automatically JIT-compiled via custom\_extensions.py The following are some references for creating custom extensions for PyTorch: -* Custom C++ and CUDA Extension: https://pytorch.org/tutorials/advanced/cpp\_extension.html -* Tensor class: https://pytorch.org/cppdocs/api/classat\_1\_1\_tensor.html -* Tensor creation API: https://pytorch.org/cppdocs/notes/tensor\_creation.html +* Custom C++ and CUDA Extension: https://pytorch.org/tutorials/advanced/cpp_extension.html +* Tensor class: https://pytorch.org/cppdocs/api/classat_1_1_tensor.html +* Tensor creation API: https://pytorch.org/cppdocs/notes/tensor_creation.html In the CUDA files, we subsitute the following MX terminology as "block_size" already has a different meaning in CUDA: