CUDA: Blackwell features for non-native builds

[JohannesGaessler]

This PR adds architectures to enable the recent Blackwell-specific MXFP4 optimizations for non-native builds. The problem with 120f-virtual which we were using in the initial PR is that it doesn't match some regex that CMake was using to validate CUDA architectures. But the same regex seems to be compatible with 120a-real and 121a-real so I would suggest that we for now simply build those since there is no other hardware to cover. Newer CMake versions come with a bugfix for the regex so presumably this will be less problematic for us to handle in the future. @CISC is there a way to run the Windows CUDA release CI without merging a PR?

[CISC]

is there a way to run the Windows CUDA release CI without merging a PR?

Yes, you can run the Release CI manually against a specified branch, or on your own fork.

Edit: I have already tested with 120a-real though:
https://github.com/CISC/llama.cpp/actions/runs/20543537454/job/59011054386#step:6:52

[CISC]

Together with #18441 we should be in business...

[am17an]

I think there is another device with sm 122? Not sure There isn't according to https://developer.nvidia.com/cuda/gpus

[thomasjfox]

@JohannesGaessler: This is the cmake output using a fresh build dir using commit 0c89864 on master:

$ cmake -DGGML_CUDA=on -DCMAKE_CUDA_COMPILER=/usr/local/cuda-13.0/bin/nvcc ../
-- The C compiler identification is GNU 15.2.1
-- The CXX compiler identification is GNU 15.2.1
-- Detecting C compiler ABI info
-- Detecting C compiler ABI info - done
-- Check for working C compiler: /usr/lib64/ccache/cc - skipped
-- Detecting C compile features
-- Detecting C compile features - done
-- Detecting CXX compiler ABI info
-- Detecting CXX compiler ABI info - done
-- Check for working CXX compiler: /usr/lib64/ccache/c++ - skipped
-- Detecting CXX compile features
-- Detecting CXX compile features - done
CMAKE_BUILD_TYPE=Release
-- Found Git: /usr/bin/git (found version "2.52.0")
-- The ASM compiler identification is GNU
-- Found assembler: /usr/lib64/ccache/cc
-- Performing Test CMAKE_HAVE_LIBC_PTHREAD
-- Performing Test CMAKE_HAVE_LIBC_PTHREAD - Success
-- Found Threads: TRUE
-- ccache found, compilation results will be cached. Disable with GGML_CCACHE=OFF.
-- CMAKE_SYSTEM_PROCESSOR: x86_64
-- GGML_SYSTEM_ARCH: x86
-- Including CPU backend
-- Found OpenMP_C: -fopenmp (found version "4.5")
-- Found OpenMP_CXX: -fopenmp (found version "4.5")
-- Found OpenMP: TRUE (found version "4.5")
-- x86 detected
-- Adding CPU backend variant ggml-cpu: -march=native 
-- Found CUDAToolkit: /usr/local/cuda-13.0/targets/x86_64-linux/include (found version "13.0.88")
-- CUDA Toolkit found
-- Using CUDA architectures: native
-- The CUDA compiler identification is NVIDIA 13.0.88 with host compiler GNU 15.2.1
-- Detecting CUDA compiler ABI info
-- Detecting CUDA compiler ABI info - done
-- Check for working CUDA compiler: /usr/local/cuda-13.0/bin/nvcc - skipped
-- Detecting CUDA compile features
-- Detecting CUDA compile features - done
-- Replacing 120-real with 120a-real
-- CUDA host compiler is GNU 15.2.1
-- Including CUDA backend
-- ggml version: 0.9.4
-- ggml commit:  0c8986403
-- Found CURL: /usr/lib64/libcurl.so (found version "8.11.1")
-- Configuring done (3.7s)
-- Generating done (0.1s)

All compiled just fine. cmake version is 3.31.6.

I will test up next your change in PR #18457