chore(build-prerelease): reduce build jobs and nvcc threads

optimize resource usage by reducing parallelism in builds to avoid
overloading runners. applies to all gpu flavours (ampere, ada,
blackwell).

.gitea/workflows/build-prerelease.yml

@@ -36,24 +36,24 @@ jobs:
             cuda_home: /usr/local/cuda-13.0
             cargo_features: "cuda cudnn flash-attn nccl"
             compute_caps: "86"
-            build_jobs: 12
-            nvcc_threads: 4
+            build_jobs: 4
+            nvcc_threads: 2
           - name: ada
             fedora_version: "43"
             runner: cuda-13.0
             cuda_home: /usr/local/cuda-13.0
             cargo_features: "cuda cudnn flash-attn nccl"
             compute_caps: "89"
-            build_jobs: 12
-            nvcc_threads: 4
+            build_jobs: 4
+            nvcc_threads: 2
           - name: blackwell
             fedora_version: "43"
             runner: cuda-13.0
             cuda_home: /usr/local/cuda-13.0
             cargo_features: "cuda cudnn flash-attn nccl"
             compute_caps: "120"
-            build_jobs: 12
-            nvcc_threads: 4
+            build_jobs: 4
+            nvcc_threads: 2
     runs-on: ${{ matrix.runner }}
     steps:
       - uses: actions/checkout@v4

.gitea/workflows/build-release.yml

@@ -24,24 +24,24 @@ jobs:
             cuda_home: /usr/local/cuda-13.0
             cargo_features: "cuda cudnn flash-attn nccl"
             compute_caps: "86"
-            build_jobs: 12
-            nvcc_threads: 4
+            build_jobs: 4
+            nvcc_threads: 2
           - name: ada
             fedora_version: "43"
             runner: cuda-13.0
             cuda_home: /usr/local/cuda-13.0
             cargo_features: "cuda cudnn flash-attn nccl"
             compute_caps: "89"
-            build_jobs: 12
-            nvcc_threads: 4
+            build_jobs: 4
+            nvcc_threads: 2
           - name: blackwell
             fedora_version: "43"
             runner: cuda-13.0
             cuda_home: /usr/local/cuda-13.0
             cargo_features: "cuda cudnn flash-attn nccl"
             compute_caps: "120"
-            build_jobs: 12
-            nvcc_threads: 4
+            build_jobs: 4
+            nvcc_threads: 2
     runs-on: ${{ matrix.runner }}
     steps:
       - uses: actions/checkout@v4