ci(cuda): build only for CI runner GPU arch and raise parallelism

.github/workflows/cuda.yml

@@ -46,8 +46,24 @@ jobs:
           nvidia-smi
           source toolchain/install/setup
           rm -rf build
-          cmake -B build -G Ninja -DUSE_CUDA=ON -DBUILD_TESTING=ON -DENABLE_FLOAT_FFTW=ON
-          cmake --build build -j4
+          # Pin the CUDA architecture to sm_70 (Tesla V100) instead of building
+          # for the default 7 architectures (60/70/75/80/86/89/90). The CI GPU
+          # pool is a homogeneous set of V100 pods (verified by sampling 5+
+          # runner IDs, all on the gpu-runner-6qqd8-* K8s deployment with
+          # identical test step timings across runs). Building for a single
+          # arch cuts nvcc work to ~1/7 and keeps the ccache key uniform so
+          # cache_warmer (#7756) can populate every runner with the same
+          # entries.
+          #
+          # NOTE: this is an explicit pin rather than auto-detection. On a
+          # cluster, configuring on a login node whose GPU differs from the
+          # compute node would silently produce a binary that fails with
+          # "no kernel image is available" at run time; HPC convention is to
+          # build on the compute node, where the workflow's hardcoded -D
+          # matches the actual hardware.
+          cmake -B build -G Ninja -DUSE_CUDA=ON -DBUILD_TESTING=ON -DENABLE_FLOAT_FFTW=ON \
+            -DCMAKE_CUDA_ARCHITECTURES=70
+          cmake --build build -j "$(nproc)"
           cmake --install build
 
       - name: Module_LCAO CUDA Unittests