diff --git a/.github/workflows/turboquant-ci.yml b/.github/workflows/turboquant-ci.yml
index bd01bff84..844c1a8b5 100644
--- a/.github/workflows/turboquant-ci.yml
+++ b/.github/workflows/turboquant-ci.yml
@@ -35,6 +35,12 @@ jobs:
           sudo apt-get update
           sudo apt-get install -y build-essential cmake ccache
 
+      - uses: ggml-org/ccache-action@v1.2.21
+        with:
+          key: turboquant-ci-cpu
+          evict-old-files: 1d
+          save: ${{ github.event_name == 'push' && github.ref == 'refs/heads/master' }}
+
       - name: Configure
         run: |
           cmake -B build \
@@ -59,22 +65,32 @@ jobs:
   cuda:
     name: CUDA compile check
     runs-on: ubuntu-latest
+    # Same approach as the upstream ubuntu-latest-cuda job. Installing the toolkit
+    # onto a bare runner does not work: the cuda-12-6 meta-package depends on the
+    # NVIDIA kernel driver, which cannot build there, and dpkg then fails the whole
+    # install. A devel image ships nvcc and cuBLAS and needs no driver at all —
+    # correct for a compile-only check.
+    container: nvidia/cuda:12.6.2-devel-ubuntu24.04
     steps:
       - uses: actions/checkout@v4
 
-      - uses: Jimver/cuda-toolkit@v0.2.19
-        id: cuda-toolkit
-        with:
-          cuda: "12.6.2"
-          method: network
-          # No sub-packages filter: ggml-cuda needs cuBLAS in addition to nvcc and
-          # the runtime, and guessing package names cost two failed runs. A full
-          # install is slower but does not break on naming.
-
       - name: Install dependencies
+        env:
+          DEBIAN_FRONTEND: noninteractive
+        # No sudo: the container runs as root.
         run: |
-          sudo apt-get update
-          sudo apt-get install -y build-essential cmake
+          apt-get update
+          apt-get install -y build-essential cmake git libgomp1
+
+      # ggml's CMake picks ccache up automatically once the binary exists, so this
+      # only has to provide the cache. It matters here more than anywhere else:
+      # ggml-cuda is 131 translation units, and nvcc compiles each one several
+      # times over (PTX, SASS, host). Must run before Configure.
+      - uses: ggml-org/ccache-action@v1.2.21
+        with:
+          key: turboquant-ci-cuda
+          evict-old-files: 1d
+          save: ${{ github.event_name == 'push' && github.ref == 'refs/heads/master' }}
 
       - name: Configure
         run: |
@@ -82,6 +98,8 @@ jobs:
             -DCMAKE_BUILD_TYPE=Release \
             -DGGML_CUDA=ON \
             -DCMAKE_CUDA_ARCHITECTURES=75 \
+            -DGGML_NATIVE=OFF \
+            -DLLAMA_FATAL_WARNINGS=ON \
             -DLLAMA_BUILD_TESTS=OFF \
             -DLLAMA_BUILD_EXAMPLES=OFF \
             -DLLAMA_BUILD_SERVER=OFF