diff --git a/.github/actions/workflow-build/build-workflow.py b/.github/actions/workflow-build/build-workflow.py
index 58b818ffe30..598f8b64b42 100755
--- a/.github/actions/workflow-build/build-workflow.py
+++ b/.github/actions/workflow-build/build-workflow.py
@@ -269,6 +269,8 @@ def get_job_type_info(job):
         result['name'] = job.capitalize()
     if not 'gpu' in result:
         result['gpu'] = False
+    if not 'cuda_ext' in result:
+        result['cuda_ext'] = False
     if not 'needs' in result:
         result['needs'] = None
     if not 'invoke' in result:
@@ -413,13 +415,16 @@ def generate_dispatch_job_image(matrix_job, job_type):
     ctk = matrix_job['ctk']
     host_compiler = generate_dispatch_job_host_compiler(matrix_job, job_type)
 
+    job_info = get_job_type_info(job_type)
+    ctk_suffix = "ext" if job_info['cuda_ext'] else ""
+
     if is_windows(matrix_job):
-        return f"rapidsai/devcontainers:{devcontainer_version}-cuda{ctk}-{host_compiler}"
+        return f"rapidsai/devcontainers:{devcontainer_version}-cuda{ctk}{ctk_suffix}-{host_compiler}"
 
     if is_nvhpc(matrix_job):
         return f"rapidsai/devcontainers:{devcontainer_version}-cpp-{host_compiler}"
 
-    return f"rapidsai/devcontainers:{devcontainer_version}-cpp-{host_compiler}-cuda{ctk}"
+    return f"rapidsai/devcontainers:{devcontainer_version}-cpp-{host_compiler}-cuda{ctk}{ctk_suffix}"
 
 
 def generate_dispatch_job_command(matrix_job, job_type):
diff --git a/ci/matrix.yaml b/ci/matrix.yaml
index d4f19f5b2a3..0fbc24906dd 100644
--- a/ci/matrix.yaml
+++ b/ci/matrix.yaml
@@ -8,6 +8,10 @@ workflows:
   #   - {jobs: ['test'], project: 'thrust', std: 17, ctk: 'curr', cxx: ['gcc12', 'llvm16']}
   #
   override:
+    - {jobs: ['test_ext'],  project: 'cudax', ctk: ['12.0'        ], std: 'min', cxx: ['gcc12'],   cmake_options: '-Dcudax_ENABLE_CUDASTF_MATHLIBS=ON'}
+    - {jobs: ['test_ext'],  project: 'cudax', ctk: [        'curr'], std: 'all', cxx: ['gcc12'],   cmake_options: '-Dcudax_ENABLE_CUDASTF_MATHLIBS=ON'}
+    - {jobs: ['test_ext'],  project: 'cudax', ctk: ['12.0'        ], std: 'max', cxx: ['clang14'], cmake_options: '-Dcudax_ENABLE_CUDASTF_MATHLIBS=ON'}
+    - {jobs: ['test_ext'],  project: 'cudax', ctk: [        'curr'], std: 'max', cxx: ['clang18'], cmake_options: '-Dcudax_ENABLE_CUDASTF_MATHLIBS=ON'}
 
   pull_request:
     # Old CTK
@@ -46,10 +50,10 @@ workflows:
     - {jobs: ['build'], project: 'cudax', ctk: ['12.0'        ], std: 17,    cxx: ['gcc12'], sm: "90"}
     - {jobs: ['build'], project: 'cudax', ctk: [        'curr'], std: 17,    cxx: ['gcc13'], sm: "90a"}
     - {jobs: ['build'], project: 'cudax', ctk: [        'curr'], std: 'all', cxx: ['gcc13', 'clang16'], cpu: 'arm64'}
-    - {jobs: ['test'],  project: 'cudax', ctk: ['12.0'        ], std: 'min', cxx: ['gcc12']}
-    - {jobs: ['test'],  project: 'cudax', ctk: [        'curr'], std: 'all', cxx: ['gcc12']}
-    - {jobs: ['test'],  project: 'cudax', ctk: ['12.0'        ], std: 'max', cxx: ['clang14']}
-    - {jobs: ['test'],  project: 'cudax', ctk: [        'curr'], std: 'max', cxx: ['clang18']}
+    - {jobs: ['test_ext'],  project: 'cudax', ctk: ['12.0'        ], std: 'min', cxx: ['gcc12'],   cmake_options: '-Dcudax_ENABLE_CUDASTF_MATHLIBS=ON'}
+    - {jobs: ['test_ext'],  project: 'cudax', ctk: [        'curr'], std: 'all', cxx: ['gcc12'],   cmake_options: '-Dcudax_ENABLE_CUDASTF_MATHLIBS=ON'}
+    - {jobs: ['test_ext'],  project: 'cudax', ctk: ['12.0'        ], std: 'max', cxx: ['clang14'], cmake_options: '-Dcudax_ENABLE_CUDASTF_MATHLIBS=ON'}
+    - {jobs: ['test_ext'],  project: 'cudax', ctk: [        'curr'], std: 'max', cxx: ['clang18'], cmake_options: '-Dcudax_ENABLE_CUDASTF_MATHLIBS=ON'}
     # Python and c/parallel jobs:
     - {jobs: ['test'], project: ['cccl_c_parallel', 'pycuda'], ctk: '12.6'}
     # cccl-infra:
@@ -169,8 +173,9 @@ host_compilers:
 
 # Jobs support the following properties:
 #
-# - gpu: Whether the job requires a GPU runner. Default is false.
 # - name: The human-readable name of the job. Default is the capitalized job key.
+# - gpu: Whether the job requires a GPU runner. Default is false.
+# - cuda_ext: Whether the job requires a devcontainer with extra CUDA libraries. Default is false.
 # - needs:
 #   - A list of jobs that must be completed before this job can run. Default is an empty list.
 #   - These are automatically added if needed:
@@ -188,6 +193,10 @@ jobs:
   test:         { gpu: true, needs: 'build' }
   test_nobuild: { gpu: true, name: 'Test', invoke: { prefix: 'test' } }
 
+  # Use images with extra CUDA libs:
+  build_ext:    { name: "Build (extra CTK libs)", gpu: false, cuda_ext: true }
+  test_ext:     { name: "Test (extra CTK libs)",  gpu: true,  cuda_ext: true, needs: 'build_ext' }
+
   # CCCL:
   infra: { gpu: true } # example project launches a kernel