ci: cap parallelism (MAX_JOBS=2), add job timeout + concurrency auto-cancel

Author: GeophyAI

.github/workflows/wheels.yml

@@ -27,10 +27,17 @@ on:
     branches: ["feat/pypi-cuda-wheels"]
     tags: ["v*"]
 
+# A new push to the same ref cancels the in-flight run — prevents pile-ups
+# (and auto-kills a wedged run instead of waiting out the 6 h job limit).
+concurrency:
+  group: wheels-${{ github.ref }}
+  cancel-in-progress: true
+
 jobs:
   build_wheels:
     name: cp${{ matrix.python }} CUDA wheel
     runs-on: ubuntu-latest
+    timeout-minutes: 60          # bound any compile hang (no more 6 h wedges)
     strategy:
       fail-fast: false
       matrix:
@@ -63,9 +70,13 @@ jobs:
           # NB: quote the arch list — cibuildwheel parses CIBW_ENVIRONMENT with
           # bashlex, so unquoted ';' would be read as a command separator and
           # error out ("Malformed environment") before any build starts.
+          # MAX_JOBS=2 caps ninja file-parallelism: the default (all 4 vCPUs ×
+          # nvcc --threads across 6 arches) oversubscribed RAM and wedged the
+          # runner. 2 files in flight keeps peak memory well under 16 GB.
           CIBW_ENVIRONMENT_LINUX: >
             SWEEP_BUILD_CUDA=1
             TORCH_CUDA_ARCH_LIST="7.0;7.5;8.0;8.6;8.9;9.0"
+            MAX_JOBS=2
             CUDA_HOME=/usr/local/cuda
             PATH=/usr/local/cuda/bin:$PATH
           # Build against the exact torch users resolve at install time.