diff --git a/.github/workflows/build-wheel.yml b/.github/workflows/build-wheel.yml
index 09a8b00e952..d1a6d156d91 100644
--- a/.github/workflows/build-wheel.yml
+++ b/.github/workflows/build-wheel.yml
@@ -227,16 +227,18 @@ jobs:
         with:
           package-dir: ./cuda_bindings/
           output-dir: ${{ env.CUDA_BINDINGS_ARTIFACTS_DIR }}
+          extras: uv
         env:
           CIBW_BUILD: ${{ env.CIBW_BUILD }}
-          CIBW_BEFORE_BUILD_LINUX: 'python -m pip install --upgrade "pip>=25.3"'
-          CIBW_BEFORE_BUILD_WINDOWS: 'python -m pip install --upgrade "pip>=25.3" delvewheel'
+          # build[uv]: same PEP 517 orchestration, with uv instead of pip.
+          # extras: uv puts uv on PATH (required on Windows).
+          CIBW_BUILD_FRONTEND: "build[uv]"
           # CIBW mounts the host filesystem under /host
           CIBW_ENVIRONMENT_LINUX: >
             CUDA_PATH=/host/${{ env.CUDA_PATH }}
             CUDA_PYTHON_PARALLEL_LEVEL=${{ env.CUDA_PYTHON_PARALLEL_LEVEL }}
-            PIP_BUILD_CONSTRAINT=/host/${{ github.workspace }}/wheel-constraints/cuda-bindings.txt
-            PIP_CONSTRAINT=/host/${{ github.workspace }}/wheel-constraints/cuda-bindings.txt
+            UV_BUILD_CONSTRAINT=/host/${{ github.workspace }}/wheel-constraints/cuda-bindings.txt
+            UV_CONSTRAINT=/host/${{ github.workspace }}/wheel-constraints/cuda-bindings.txt
             CC="/host/${{ env.SCCACHE_PATH }} cc"
             CXX="/host/${{ env.SCCACHE_PATH }} c++"
             SCCACHE_GHA_ENABLED=true
@@ -248,17 +250,13 @@ jobs:
           CIBW_ENVIRONMENT_WINDOWS: >
             CUDA_PATH="$(cygpath -w ${{ env.CUDA_PATH }})"
             CUDA_PYTHON_PARALLEL_LEVEL=${{ env.CUDA_PYTHON_PARALLEL_LEVEL }}
-            PIP_BUILD_CONSTRAINT="$(cygpath -w ./wheel-constraints/cuda-bindings.txt)"
-            PIP_CONSTRAINT="$(cygpath -w ./wheel-constraints/cuda-bindings.txt)"
-          # check cache stats before leaving cibuildwheel
-          CIBW_BEFORE_TEST_LINUX: >
+            UV_BUILD_CONSTRAINT="$(cygpath -w ./wheel-constraints/cuda-bindings.txt)"
+            UV_CONSTRAINT="$(cygpath -w ./wheel-constraints/cuda-bindings.txt)"
+          # Dump sccache stats during repair; no dummy test stage.
+          CIBW_REPAIR_WHEEL_COMMAND_LINUX: >
             "/host/${{ env.SCCACHE_PATH }}" --show-adv-stats &&
-            "/host/${{ env.SCCACHE_PATH }}" --show-stats --stats-format=json > /host/${{ github.workspace }}/sccache_bindings.json
-          # force the test stage to be run (so that before-test is not skipped)
-          # TODO: we might want to think twice on adding this, it does a lot of
-          # things before reaching this command.
-          CIBW_TEST_COMMAND: >
-            echo "ok!"
+            "/host/${{ env.SCCACHE_PATH }}" --show-stats --stats-format=json > /host/${{ github.workspace }}/sccache_bindings.json &&
+            auditwheel repair -w {dest_dir} {wheel}
 
       - name: Report sccache stats (cuda.bindings)
         if: ${{ env.BUILD_BINDINGS == 'true' && !startsWith(inputs.host-platform, 'win') }}
@@ -327,17 +325,18 @@ jobs:
         with:
           package-dir: ./cuda_core/
           output-dir: ${{ env.CUDA_CORE_ARTIFACTS_DIR }}
+          extras: uv
         env:
           CIBW_BUILD: ${{ env.CIBW_BUILD }}
-          CIBW_BEFORE_BUILD_LINUX: 'python -m pip install --upgrade "pip>=25.3"'
-          CIBW_BEFORE_BUILD_WINDOWS: 'python -m pip install --upgrade "pip>=25.3" delvewheel'
+          # See the cuda.bindings step above for UV_* vs PIP_*.
+          CIBW_BUILD_FRONTEND: "build[uv]"
           # CIBW mounts the host filesystem under /host
           CIBW_ENVIRONMENT_LINUX: >
             CUDA_PATH=/host/${{ env.CUDA_PATH }}
             CUDA_PYTHON_PARALLEL_LEVEL=${{ env.CUDA_PYTHON_PARALLEL_LEVEL }}
             CUDA_CORE_BUILD_MAJOR=${{ env.BUILD_CUDA_MAJOR }}
-            PIP_BUILD_CONSTRAINT=/host/${{ github.workspace }}/wheel-constraints/cuda-core.txt
-            PIP_CONSTRAINT=/host/${{ github.workspace }}/wheel-constraints/cuda-core.txt
+            UV_BUILD_CONSTRAINT=/host/${{ github.workspace }}/wheel-constraints/cuda-core.txt
+            UV_CONSTRAINT=/host/${{ github.workspace }}/wheel-constraints/cuda-core.txt
             CC="/host/${{ env.SCCACHE_PATH }} cc"
             CXX="/host/${{ env.SCCACHE_PATH }} c++"
             SCCACHE_GHA_ENABLED=true
@@ -350,17 +349,13 @@ jobs:
             CUDA_PATH="$(cygpath -w ${{ env.CUDA_PATH }})"
             CUDA_PYTHON_PARALLEL_LEVEL=${{ env.CUDA_PYTHON_PARALLEL_LEVEL }}
             CUDA_CORE_BUILD_MAJOR=${{ env.BUILD_CUDA_MAJOR }}
-            PIP_BUILD_CONSTRAINT="$(cygpath -w ./wheel-constraints/cuda-core.txt)"
-            PIP_CONSTRAINT="$(cygpath -w ./wheel-constraints/cuda-core.txt)"
-          # check cache stats before leaving cibuildwheel
-          CIBW_BEFORE_TEST_LINUX: >
-            "/host${{ env.SCCACHE_PATH }}" --show-adv-stats &&
-            "/host${{ env.SCCACHE_PATH }}" --show-stats --stats-format=json > /host/${{ github.workspace }}/sccache_core.json
-          # force the test stage to be run (so that before-test is not skipped)
-          # TODO: we might want to think twice on adding this, it does a lot of
-          # things before reaching this command.
-          CIBW_TEST_COMMAND: >
-            echo "ok!"
+            UV_BUILD_CONSTRAINT="$(cygpath -w ./wheel-constraints/cuda-core.txt)"
+            UV_CONSTRAINT="$(cygpath -w ./wheel-constraints/cuda-core.txt)"
+          # Dump sccache stats during repair; no dummy test stage.
+          CIBW_REPAIR_WHEEL_COMMAND_LINUX: >
+            "/host/${{ env.SCCACHE_PATH }}" --show-adv-stats &&
+            "/host/${{ env.SCCACHE_PATH }}" --show-stats --stats-format=json > /host/${{ github.workspace }}/sccache_core.json &&
+            auditwheel repair -w {dest_dir} {wheel}
 
       - name: Report sccache stats (cuda.core)
         if: ${{ env.BUILD_CORE == 'true' && !startsWith(inputs.host-platform, 'win') }}
@@ -614,17 +609,18 @@ jobs:
         with:
           package-dir: ./cuda_core/
           output-dir: ${{ env.CUDA_CORE_ARTIFACTS_DIR }}
+          extras: uv
         env:
           CIBW_BUILD: ${{ env.CIBW_BUILD }}
-          CIBW_BEFORE_BUILD_LINUX: 'python -m pip install --upgrade "pip>=25.3"'
-          CIBW_BEFORE_BUILD_WINDOWS: 'python -m pip install --upgrade "pip>=25.3" delvewheel'
+          # See the cuda.bindings step above for UV_* vs PIP_*.
+          CIBW_BUILD_FRONTEND: "build[uv]"
           # CIBW mounts the host filesystem under /host
           CIBW_ENVIRONMENT_LINUX: >
             CUDA_PATH=/host/${{ env.CUDA_PATH }}
             CUDA_PYTHON_PARALLEL_LEVEL=${{ env.CUDA_PYTHON_PARALLEL_LEVEL }}
             CUDA_CORE_BUILD_MAJOR=${{ env.BUILD_PREV_CUDA_MAJOR }}
-            PIP_BUILD_CONSTRAINT=/host/${{ github.workspace }}/wheel-constraints/cuda-core-prev.txt
-            PIP_CONSTRAINT=/host/${{ github.workspace }}/wheel-constraints/cuda-core-prev.txt
+            UV_BUILD_CONSTRAINT=/host/${{ github.workspace }}/wheel-constraints/cuda-core-prev.txt
+            UV_CONSTRAINT=/host/${{ github.workspace }}/wheel-constraints/cuda-core-prev.txt
             CC="/host/${{ env.SCCACHE_PATH }} cc"
             CXX="/host/${{ env.SCCACHE_PATH }} c++"
             SCCACHE_GHA_ENABLED=true
@@ -637,17 +633,13 @@ jobs:
             CUDA_PATH="$(cygpath -w ${{ env.CUDA_PATH }})"
             CUDA_PYTHON_PARALLEL_LEVEL=${{ env.CUDA_PYTHON_PARALLEL_LEVEL }}
             CUDA_CORE_BUILD_MAJOR=${{ env.BUILD_PREV_CUDA_MAJOR }}
-            PIP_BUILD_CONSTRAINT="$(cygpath -w ./wheel-constraints/cuda-core-prev.txt)"
-            PIP_CONSTRAINT="$(cygpath -w ./wheel-constraints/cuda-core-prev.txt)"
-          # check cache stats before leaving cibuildwheel
-          CIBW_BEFORE_TEST_LINUX: >
-            "/host${{ env.SCCACHE_PATH }}" --show-adv-stats &&
-            "/host${{ env.SCCACHE_PATH }}" --show-stats --stats-format=json > /host/${{ github.workspace }}/sccache_core_prev.json
-          # force the test stage to be run (so that before-test is not skipped)
-          # TODO: we might want to think twice on adding this, it does a lot of
-          # things before reaching this command.
-          CIBW_TEST_COMMAND: >
-            echo "ok!"
+            UV_BUILD_CONSTRAINT="$(cygpath -w ./wheel-constraints/cuda-core-prev.txt)"
+            UV_CONSTRAINT="$(cygpath -w ./wheel-constraints/cuda-core-prev.txt)"
+          # Dump sccache stats during repair; no dummy test stage.
+          CIBW_REPAIR_WHEEL_COMMAND_LINUX: >
+            "/host/${{ env.SCCACHE_PATH }}" --show-adv-stats &&
+            "/host/${{ env.SCCACHE_PATH }}" --show-stats --stats-format=json > /host/${{ github.workspace }}/sccache_core_prev.json &&
+            auditwheel repair -w {dest_dir} {wheel}
 
       - name: Report sccache stats (cuda.core prev)
         if: ${{ !inputs.single-cuda-major && env.BUILD_CORE == 'true' && !startsWith(inputs.host-platform, 'win') }}
diff --git a/cuda_bindings/pyproject.toml b/cuda_bindings/pyproject.toml
index 10e9f5e2fb1..f628f7b2e1a 100644
--- a/cuda_bindings/pyproject.toml
+++ b/cuda_bindings/pyproject.toml
@@ -85,7 +85,6 @@ build-verbosity = 1
 archs = "native"
 
 [tool.cibuildwheel.windows]
-before-build = "pip install delvewheel"
 repair-wheel-command = "delvewheel repair --namespace-pkg cuda -w {dest_dir} {wheel}"
 
 [tool.pytest.ini_options]
diff --git a/cuda_core/pyproject.toml b/cuda_core/pyproject.toml
index 33f72e22dea..9c3173cdfcb 100644
--- a/cuda_core/pyproject.toml
+++ b/cuda_core/pyproject.toml
@@ -152,5 +152,4 @@ build-verbosity = 1
 archs = "native"
 
 [tool.cibuildwheel.windows]
-before-build = "pip install delvewheel"
 repair-wheel-command = "delvewheel repair --namespace-pkg cuda --exclude \"torch_cpu.dll;torch_python.dll\" -w {dest_dir} {wheel}"