|
24 | 24 | # wheel.py-api = "py3" makes the CUDA wheel interpreter-agnostic, |
25 | 25 | # so one builder per toolkit version is sufficient. |
26 | 26 | 'pyver' = @("3.9") |
27 | | - 'cuda' = @("12.1.1", "12.2.2", "12.3.2", "12.4.1", "12.5.1") |
| 27 | + 'cuda' = @("12.1.1", "12.2.2", "12.3.2", "12.4.1", "12.5.1", "13.0.2", "13.2.1") |
28 | 28 | 'releasetag' = @("basic") |
29 | 29 | 'exclude' = @( |
30 | 30 | @{ 'os' = 'windows-2022'; 'cuda' = '12.1.1' }, |
@@ -169,10 +169,15 @@ jobs: |
169 | 169 | } |
170 | 170 | $cudaTagVersion = $nvccVersion.Replace('.','') |
171 | 171 | $env:VERBOSE = '1' |
172 | | - # Build real cubins for the supported GPUs, including Pascal, and keep |
| 172 | + $cudaArchs = "60-real;61-real;70-real;75-real;80-real;86-real;89-real;90-real;90-virtual" |
| 173 | + if ([version]$nvccVersion -ge [version]"13.0") { |
| 174 | + # CUDA 13 dropped offline compilation support for pre-Turing targets. |
| 175 | + $cudaArchs = "75-real;80-real;86-real;89-real;90-real;90-virtual" |
| 176 | + } |
| 177 | + # Build real cubins for the supported GPUs and keep |
173 | 178 | # one forward-compatible PTX target instead of embedding PTX for every |
174 | 179 | # SM. This keeps the wheel under GitHub's 2 GiB release-asset limit. |
175 | | - $env:CMAKE_ARGS = "-DGGML_CUDA_FORCE_MMQ=ON -DGGML_CUDA=on -DCMAKE_CUDA_ARCHITECTURES=60-real;61-real;70-real;75-real;80-real;86-real;89-real;90-real;90-virtual -DCMAKE_CUDA_FLAGS=-allow-unsupported-compiler -DCMAKE_CUDA_FLAGS_INIT=-allow-unsupported-compiler $env:CMAKE_ARGS" |
| 180 | + $env:CMAKE_ARGS = "-DGGML_CUDA_FORCE_MMQ=ON -DGGML_CUDA=on -DCMAKE_CUDA_ARCHITECTURES=$cudaArchs -DCMAKE_CUDA_FLAGS=-allow-unsupported-compiler -DCMAKE_CUDA_FLAGS_INIT=-allow-unsupported-compiler $env:CMAKE_ARGS" |
176 | 181 | $env:CMAKE_ARGS = $env:CMAKE_ARGS + ' -DGGML_AVX2=off -DGGML_FMA=off -DGGML_F16C=off' |
177 | 182 | python -m build --wheel |
178 | 183 | # Publish tags that reflect the actual installed toolkit version. |
|
0 commit comments