Multiple updates and refactorings (#280)

2026-01-16 17:06:52 +08:00
parent 3ccf40c53a
commit 0f5f266202
55 changed files with 2706 additions and 891 deletions
--- a/csrc/utils/math.hpp
+++ b/csrc/utils/math.hpp
@@ -6,6 +6,9 @@

 namespace deep_gemm {

+// TODO: Use `torch::kFloat4_e2m1fn_x2`
+constexpr auto kPackedFP4 = torch::kUInt8;
+
 template <typename T>
 static T ceil_div(const T& a, const T& b) {
    return (a + b - 1) / b;