fix(qwen-tts): limit CUDA compiler threads

The CUDA 13 FlashAttention build still exhausts hosted-runner memory with a single ninja worker because nvcc can compile multiple threads internally. Limit nvcc to one thread for that profile and guard the setting in the backend test script.

Assisted-by: Codex:gpt-5

backend/python/qwen-tts/install.sh

@@ -10,4 +10,11 @@ else
     source $backend_dir/../common/libbackend.sh
 fi
 
+# CUDA 13 has no prebuilt FlashAttention wheel, so the fallback source build
+# exceeds the CI runner's memory when ninja compiles multiple units at once.
+if [ "x${BUILD_PROFILE}" = "xcublas13" ]; then
+    export MAX_JOBS="${MAX_JOBS:-1}"
+    export NVCC_THREADS="${NVCC_THREADS:-1}"
+fi
+
 installRequirements

backend/python/qwen-tts/requirements-cublas13-after.txt

@@ -0,0 +1 @@
+flash-attn

backend/python/qwen-tts/test.sh

@@ -2,6 +2,15 @@
 set -e
 
 backend_dir=$(dirname $0)
+
+for cuda_version in 12 13; do
+    grep -qx "flash-attn" "$backend_dir/requirements-cublas${cuda_version}-after.txt"
+done
+
+grep -q 'BUILD_PROFILE.*cublas13' "$backend_dir/install.sh"
+grep -q 'MAX_JOBS.*1' "$backend_dir/install.sh"
+grep -q 'NVCC_THREADS.*1' "$backend_dir/install.sh"
+
 if [ -d $backend_dir/common ]; then
     source $backend_dir/common/libbackend.sh
 else