alrojo/fp4attn is a repository that handles the forward and backward directions of FP4 attention under the name CroissAttn. The latest publicly released version is not listed in FACTS, so the version cannot be confirmed. The repository description lists not only the operation implementation but also stochastic rounding, the property that the sum of softmax gradients is zero, and sampled centering as its topics. The scope of this repository therefore covers FP4 attention operations together with the numerical issues worth considering in those computations.
Stochastic rounding is the rounding method explicitly named in the description, and zero-sum softmax gradients and sampled centering are presented as analysis targets. These topics show which numerical characteristics one needs to examine when handling forward computation and backward computation together in low-precision operations. What the repository introduction particularly clarifies is the direction of the operations and the analysis items. Since specific performance figures, comparison results, and detailed implementation methods do not appear in the given material, they must not be interpreted as confirmed results or specific runtime characteristics.
On the default branch, a single related PDF paper file is confirmed instead of an informational README. The primary language and license are not present in the provided information, and no public release is listed either. There are 8 stars. Therefore, with only the currently verifiable material, it is difficult to judge the conditions for using the code, how to run it, or the actual performance of the kernel. It can be viewed as a reference topic for low-precision inference kernels, but asserting more than that in terms of implementation completeness or practicality would be premature.