[BUG] CU_TENSOR_MAP_DATA_TYPE_16U6_ALIGN16B isnt defiend in CUDA toolkit 12.4

[dtometzki]

Describe the bug
CU_TENSOR_MAP_DATA_TYPE_16U6_ALIGN16B is undefined in CUDA toolkit 12.4. See Screenshot

Steps/Code to reproduce bug
mkdir cutlass/build
cd build
cmake .. -DCUTLASS_NVCC_ARCHS="89"

Expected behavior
see screenshot

Environment details (please complete the following information):
NVIDIA GeForce RTX 4070 Ti

• runpod.io with cuda12.4 devel on Ubuntu

Additional context
workaround ist update to cuda toolkit 12.8

[LANSHANGH]

same issue，according to the reply, version v3.7.0-2-g389e4930 was used，even compiling any example of running example will return this same information.

[Skylion007]

This is blocking the upgrade of CUTLASS in PyTorch to 3.8 as well, which is in turn blocking full PyTorch Blackwell support.

[thakkarV]

Noted. We are working on a fix ASAP

[hwu36]

hopefully it is fixed in #2066

[dtometzki]

Hello together,

many thanks i works now.