cuda: declare the qwen3_5 backend

.github/workflows/ci.yml

@@ -42,3 +42,42 @@ jobs:
       - run: python -m pip install -r benchmarks/requirements.txt
       - run: python benchmarks/bench.py validate benchmarks/smoke.jsonl
       - run: python -m unittest discover -s tests/benchmarks -p 'test_*.py' -v
+
+  cuda-contract:
+    # The CUDA backend contract and its checker. No CUDA toolkit and no GPU: both
+    # read files as text.
+    #
+    # The checker and its tests arrive in #25, so on a tree without them those
+    # two steps are a skip, not a failure: this job has to be green on the branch
+    # that carries it before #25 lands, and it has to start running by itself
+    # once #25 does, in either merge order. Failing instead would redden whichever
+    # PR arrived second. The backend tests come from the repository tests/ tree,
+    # which is already in main, so they run either way.
+    runs-on: ubuntu-latest
+    timeout-minutes: 5
+    steps:
+      - uses: actions/checkout@v5
+      - uses: actions/setup-python@v5
+        with:
+          python-version: "3.11"
+      - name: Check for the checker
+        id: present
+        run: |
+          if [ -f src/backends/cuda/check_contract.py ] && [ -d src/backends/cuda/tests ]; then
+            echo "ready=true" >> "$GITHUB_OUTPUT"
+          else
+            echo "ready=false" >> "$GITHUB_OUTPUT"
+            echo "::notice::src/backends/cuda/check_contract.py and src/backends/cuda/tests are not in this tree; they arrive in #25. Skipping."
+          fi
+      - name: Contract checker tests
+        if: steps.present.outputs.ready == 'true'
+        run: python -m unittest discover -s src/backends/cuda/tests -t . -v
+      - name: Backend tests
+        # Tests live under the repository-level tests/ tree, not beside the
+        # kernels they cover (CONTRIBUTING.md, "Adding a model"). Same
+        # invocation as the benchmarks job: tests/cuda is not a package, so
+        # discovery has to use the start directory as its own top level.
+        run: python -m unittest discover -s tests/cuda -p 'test_*.py' -v
+      - name: Contract check
+        if: steps.present.outputs.ready == 'true'
+        run: python src/backends/cuda/check_contract.py --repo-root .

src/backends/cuda/qwen3_5/qwen3_5.backend.json

@@ -0,0 +1,25 @@
+{
+  "name": "qwen3_5",
+  "abi_version": 4,
+  "status": "experimental",
+  "sources": [
+    "ops.h",
+    "common.cuh",
+    "mma.cuh",
+    "norm.cu",
+    "elementwise.cu",
+    "attention.cu",
+    "gdn_prefill.cu",
+    "gemm.cu",
+    "runtime.cu"
+  ],
+  "build": {
+    "script": "build.sh",
+    "output": "libqwen3_5_cuda.so",
+    "default_arch": 89,
+    "architectures": [
+      89
+    ],
+    "min_capability": 80
+  }
+}