diff --git a/.github/workflows/build-upload-conda.yml b/.github/workflows/build-upload-conda.yml
index 3f3fca308fe0..75a35cc9cadd 100644
--- a/.github/workflows/build-upload-conda.yml
+++ b/.github/workflows/build-upload-conda.yml
@@ -42,10 +42,14 @@ jobs:
         with:
           fetch-depth: 0
 
-      - name: Create and activate conda env
-        uses: conda-incubator/setup-miniconda@fc2d68f6413eb2d87b895e92f8584b5b94a10167
+      - name: Setup and activate base conda env
+        uses: conda-incubator/setup-miniconda@8ee1f361103df19b6f8c8655fd3967a8ecb162d5 # v4.0.1
         with:
-          python-version: "3.12"
+          # Keep Conda and conda-build on conda-forge so its build plugin loads.
+          miniforge-version: "26.3.2-3"
+          conda-remove-defaults: true
+          activate-environment: ""
+          auto-activate: true
           auto-update-conda: false
           show-channel-urls: true
 
@@ -53,10 +57,10 @@ jobs:
         # This ensures conda env is active.
         shell: bash -l {0}
         run: |
-          # Setup the anaconda environment.
-          conda remove --name base conda-anaconda-telemetry
-          conda install -q -y -c conda-forge conda-build conda-verify anaconda-client
+          # Setup the conda-forge environment.
+          conda install -n base -q -y --override-channels -c conda-forge conda-build anaconda-client
           conda config --set anaconda_upload no
+          conda build --version
 
           # Build the package.
           conda build \
@@ -70,7 +74,7 @@ jobs:
             buildscripts/conda-recipes/pyomp)
 
           # Upload the package.
-          anaconda -t ${{ secrets.ANACONDA_TOKEN }} upload \
+          conda run -n base anaconda -t ${{ secrets.ANACONDA_TOKEN }} upload \
             --user python-for-hpc --label ${{ env.CONDA_LABEL}} \
             --force "$OUTPUT"
 
diff --git a/.github/workflows/build-upload-wheels.yml b/.github/workflows/build-upload-wheels.yml
index af7376c7f15e..9cb26f23168e 100644
--- a/.github/workflows/build-upload-wheels.yml
+++ b/.github/workflows/build-upload-wheels.yml
@@ -6,9 +6,20 @@ permissions:
 on:
   release:
     types: [published]
+  push:
+    branches: [main]
+    paths:
+      - "buildscripts/cibuildwheel/**"
+      - "buildscripts/modal/**"
+      - ".github/workflows/build-upload-wheels.yml"
+      - "src/**"
+      - setup.py
+      - MANIFEST.in
+      - pyproject.toml
   pull_request:
     paths:
       - "buildscripts/cibuildwheel/**"
+      - "buildscripts/modal/**"
       - ".github/workflows/build-upload-wheels.yml"
       - "src/**"
       - setup.py
@@ -110,9 +121,61 @@ jobs:
           OMP_TARGET_OFFLOAD=mandatory TEST_DEVICE=host RUN_TARGET=1 \
             python -m numba.runtests -v -- numba.openmp.tests.test_openmp.TestOpenmpTarget
 
+  # Test Linux wheels on a real NVIDIA GPU. Fork PRs run after merging to main,
+  # where Modal credentials are available to trusted code.
+  test-modal-gpu:
+    needs: build-wheels
+    if: >-
+      github.event_name != 'pull_request' ||
+      github.event.pull_request.head.repo.full_name == github.repository
+    runs-on: ubuntu-latest
+    concurrency:
+      group: modal-gpu-${{ github.workflow }}-${{ github.ref }}
+      cancel-in-progress: true
+    env:
+      MODAL_TOKEN_ID: ${{ secrets.MODAL_TOKEN_ID }}
+      MODAL_TOKEN_SECRET: ${{ secrets.MODAL_TOKEN_SECRET }}
+    steps:
+      - uses: actions/checkout@34e114876b0b11c390a56381ad16ebd13914f8d5
+
+      - name: Download Linux wheels
+        uses: actions/download-artifact@634f93cb2916e3fdff6788551b99b062d0335ce0
+        with:
+          pattern: cibw-wheels-ubuntu-latest-*
+          path: dist
+          merge-multiple: true
+
+      - name: Setup Python
+        uses: actions/setup-python@a26af69be951a213d495a4c3e4e4022e16d87065
+        with:
+          python-version: "3.12"
+
+      - name: Validate Modal credentials
+        shell: bash
+        run: |
+          if [[ -z "$MODAL_TOKEN_ID" || -z "$MODAL_TOKEN_SECRET" ]]; then
+            echo "::error::MODAL_TOKEN_ID and MODAL_TOKEN_SECRET must be configured"
+            exit 1
+          fi
+
+      - name: Run GPU tests on Modal
+        run: |
+          python -m pip install modal==1.5.3
+          set +e
+          timeout --signal=TERM --kill-after=30s 10m \
+            modal run --timestamps buildscripts/modal/test_gpu.py
+          modal_status=$?
+          set -e
+
+          if [[ $modal_status -eq 124 ]]; then
+            echo "::error::Modal GPU tests exceeded the 10-minute timeout"
+            exit 1
+          fi
+          exit "$modal_status"
+
   # Only on pre-release: Publish to TestPyPI for testing.
   publish-testpypi:
-    needs: [build-wheels, test-wheels, build-sdist]
+    needs: [build-wheels, test-wheels, test-modal-gpu, build-sdist]
     if: github.event.release.prerelease
     runs-on: ubuntu-latest
     environment: testpypi
@@ -133,7 +196,7 @@ jobs:
 
   # Only on full release: Publish to production PyPI.
   publish-pypi:
-    needs: [build-wheels, test-wheels, build-sdist]
+    needs: [build-wheels, test-wheels, test-modal-gpu, build-sdist]
     if: github.event_name == 'release' && !github.event.release.prerelease
     runs-on: ubuntu-latest
     environment: pypi
diff --git a/.gitlab-ci.yml b/.gitlab-ci.yml
deleted file mode 100644
index 5ef586c16b75..000000000000
--- a/.gitlab-ci.yml
+++ /dev/null
@@ -1,98 +0,0 @@
-###############################################################################
-# Copyright (c) 2022-23, Lawrence Livermore National Security, LLC and RADIUSS
-# project contributors. See the COPYRIGHT file for details.
-#
-# SPDX-License-Identifier: (MIT)
-###############################################################################
-
-# DESCRIPTION:
-###############################################################################
-# General GitLab pipelines configurations for supercomputers and Linux clusters
-# at Lawrence Livermore National Laboratory (LLNL).
-# This entire pipeline is LLNL-specific
-#
-# Important note: This file is a template provided by llnl/radiuss-shared-ci.
-# Remains to set variable values, change the reference to the radiuss-shared-ci
-# repo, opt-in and out optional features. The project can then extend it with
-# additional stages.
-#
-# In addition, each project should copy over and complete:
-# - .gitlab/custom-jobs-and-variables.yml
-# - .gitlab/subscribed-pipelines.yml
-#
-# The jobs should be specified in a file local to the project,
-# - .gitlab/jobs/${CI_MACHINE}.yml
-# or generated (see LLNL/Umpire for an example).
-###############################################################################
-
-# We define the following GitLab pipeline variables:
-variables:
-##### LC GITLAB CONFIGURATION
-# Use an LLNL service user to run CI. This prevents from running pipelines as
-# an actual user.
-  LLNL_SERVICE_USER: ""
-# Use the service user workspace. Solves permission issues, stores everything
-# at the same location whoever triggers a pipeline.
-  CUSTOM_CI_BUILDS_DIR: "/usr/workspace/ggeorgak/projects/pyomp/gitlab-runner"
-# Tells Gitlab to recursively update the submodules when cloning the project.
-#  GIT_SUBMODULE_STRATEGY: recursive
-
-##### PROJECT VARIABLES
-# We build the projects in the CI clone directory.
-# Used in script/gitlab/build_and_test.sh script.
-# TODO: add a clean-up mechanism.
-  BUILD_ROOT: ${CI_PROJECT_DIR}
-
-##### SHARED_CI CONFIGURATION
-# Required information about GitHub repository
-  GITHUB_PROJECT_NAME: "PyOMP"
-  GITHUB_PROJECT_ORG: "Python-for-HPC"
-# Set the build-and-test command.
-# Nested variables are allowed and useful to customize the job command. We
-# prevent variable expansion so that you can define them at job level.
-  JOB_CMD:
-    value: "buildscripts/gitlab/build-and-test.sh"
-    expand: false
-# Override the pattern describing branches that will skip the "draft PR filter
-# test".  Add protected branches here. See default value in
-# preliminary-ignore-draft-pr.yml.
-#  ALWAYS_RUN_PATTERN: ""
-
-# We organize the build-and-test stage with sub-pipelines. Each sub-pipeline
-# corresponds to a test batch on a given machine.
-
-# High level stages
-stages:
-  - prerequisites
-  - build-and-test
-
-# Template for jobs triggering a build-and-test sub-pipeline:
-.build-and-test:
-  stage: build-and-test
-  trigger:
-    include:
-      - local: '.gitlab/custom-jobs-and-variables.yml'
-      - project: 'radiuss/radiuss-shared-ci'
-        ref: 'v2024.07.0'
-        file: 'pipelines/${CI_MACHINE}.yml'
-      # Add your jobs
-      # you can use a local file
-      - local: '.gitlab/jobs/${CI_MACHINE}.yml'
-      # or a file generated in the previous steps
-      # - artifact: '${CI_MACHINE}-jobs.yml'
-      #   job: 'generate-job-file'
-      # (See Umpire CI setup for an example).
-    strategy: depend
-    forward:
-      pipeline_variables: true
-
-include:
-  # Sets ID tokens for every job using `default:`
-  - project: 'lc-templates/id_tokens'
-    file: 'id_tokens.yml'
-  # [Optional] checks preliminary to running the actual CI test
-  #- project: 'radiuss/radiuss-shared-ci'
-  #  ref: 'v2024.07.0'
-  #  file: 'utilities/preliminary-ignore-draft-pr.yml'
-  # pipelines subscribed by the project
-  - local: '.gitlab/subscribed-pipelines.yml'
diff --git a/.gitlab/custom-jobs-and-variables.yml b/.gitlab/custom-jobs-and-variables.yml
deleted file mode 100644
index 42f4f55703fb..000000000000
--- a/.gitlab/custom-jobs-and-variables.yml
+++ /dev/null
@@ -1,30 +0,0 @@
-###############################################################################
-# Copyright (c) 2022-23, Lawrence Livermore National Security, LLC and RADIUSS
-# project contributors. See the COPYRIGHT file for details.
-#
-# SPDX-License-Identifier: (MIT)
-###############################################################################
-
-# We define the following GitLab pipeline variables:
-variables:
-# In some pipelines we create only one allocation shared among jobs in
-# order to save time and resources. This allocation has to be uniquely
-# named so that we are sure to retrieve it and avoid collisions.
-  ALLOC_NAME: ${CI_PROJECT_NAME}_ci_${CI_PIPELINE_ID}
-
-# Tioga
-# Arguments for top level allocation
-# OPTIONAL: "-o per-resource.count=2" allows to get 2 jobs running on each node.
-  TIOGA_SHARED_ALLOC: "--queue=pci --exclusive --time-limit=30m --nodes=1"
-# Arguments for job level allocation
-  TIOGA_JOB_ALLOC: "--nodes=1 --begin-time=+5s"
-# Add variables that should apply to all the jobs on a machine:
-#  TIOGA_MY_VAR: "..."
-
-# Configuration shared by build and test jobs specific to this project.
-# Not all configuration can be shared. Here projects can fine tune the
-# CI behavior.
-# See Umpire for an example (export junit test reports).
-.custom_job:
-  variables:
-    JOB_TEMPLATE_CANNOT_BE_EMPTY: "True"
diff --git a/.gitlab/jobs/tioga.yml b/.gitlab/jobs/tioga.yml
deleted file mode 100644
index 3914eb5c310f..000000000000
--- a/.gitlab/jobs/tioga.yml
+++ /dev/null
@@ -1,62 +0,0 @@
-###############################################################################
-# Copyright (c) 2022-23, Lawrence Livermore National Security, LLC and RADIUSS
-# project contributors. See the COPYRIGHT file for details.
-#
-# SPDX-License-Identifier: (MIT)
-###############################################################################
-
-# We require project to define their job command using a variable (JOB_CMD).
-# In customization/gitlab-ci.yml, we encourage to define this variable as
-# non-expandable, so that project can use nested variables to configure the job
-# command. The caveat is that the reproducer here cannot capture the
-# definition of these variables in a generic fashion. By overriding the
-# following section, projects can specify the variables to define in the
-# reproducer to exactly reproduce the CI build.
-.tioga_reproducer_vars:
-  script:
-    - echo -e "Running on tioga\n"
-
-# With GitLab CI, included files cannot be empty.
-# TODO: remove when you have at least on job defined.
-variables:
-  INCLUDED_FILE_CANNOT_BE_EMPTY: "True"
-
-###############
-# Explanations:
-###############
-# RADIUSS Shared CI provides a pipeline for each machine, where a template job
-# is provided. Each of your jobs must extend this template to be added to the
-# list of jobs running on the associated machine.
-#
-# The job template then expects you to define the "JOB_CMD" variable with the
-# one line command used to trigger the build and test of your project.
-#
-# We suggest that you set your command in such a way that you can then
-# customize it per job with variables. E.g.:
-# "./path/to/my_ci_script ${A_VARIABLE}"
-
-## Adding jobs defined by the project.
-## Note: placing the extends section first allows you to override part of the
-## shared implementation if needed (and if you know what you are doing).
-#<job-name (typically build target description)>:
-#  extends: .job_on_tioga
-#  variables:
-#    <A_VARIABLE>: "<with job specific value>"
-
-.base-job:
-  extends: .job_on_tioga
-  after_script:
-    - rm -rf ${CI_BUILDS_DIR} ${CI_PROJECT_DIR}
-
-.python-variants:
-  parallel:
-    matrix:
-      - PYOMP_CI_PYTHON_VERSION:
-        - "3.10"
-        - "3.11"
-        - "3.12"
-        - "3.13"
-        - "3.14"
-
-build-and-test-tioga:
-  extends: [.base-job, .python-variants]
diff --git a/.gitlab/subscribed-pipelines.yml b/.gitlab/subscribed-pipelines.yml
deleted file mode 100644
index 714e5618290c..000000000000
--- a/.gitlab/subscribed-pipelines.yml
+++ /dev/null
@@ -1,43 +0,0 @@
-###############################################################################
-# Copyright (c) 2022-23, Lawrence Livermore National Security, LLC and RADIUSS
-# project contributors. See the COPYRIGHT file for details.
-#
-# SPDX-License-Identifier: (MIT)
-###############################################################################
-
-# The template job to test whether a machine is up.
-# Expects CI_MACHINE defined to machine name.
-.machine-check:
-  stage: prerequisites
-  tags: [shell, oslic]
-  variables:
-    GIT_STRATEGY: none
-  script:
-    - |
-      if [[ $(jq '.[env.CI_MACHINE].total_nodes_up' /usr/global/tools/lorenz/data/loginnodeStatus) == 0 ]]
-      then
-        echo -e "\e[31mNo node available on ${CI_MACHINE}\e[0m"
-        curl --url "https://api.github.com/repos/${GITHUB_PROJECT_ORG}/${GITHUB_PROJECT_NAME}/statuses/${CI_COMMIT_SHA}" \
-             --header 'Content-Type: application/json' \
-             --header "authorization: Bearer ${GITHUB_TOKEN}" \
-             --data "{ \"state\": \"failure\", \"target_url\": \"${CI_PIPELINE_URL}\", \"description\": \"GitLab ${CI_MACHINE} down\", \"context\": \"ci/gitlab/${CI_MACHINE}\" }"
-        exit 1
-      fi
-
-###
-# Trigger a build-and-test pipeline for a machine.
-# Comment the jobs for machines you don’t need.
-###
-
-
-## TIOGA
-tioga-up-check:
-  variables:
-    CI_MACHINE: "tioga"
-  extends: [.machine-check]
-
-tioga-build-and-test:
-  variables:
-    CI_MACHINE: "tioga"
-  needs: [tioga-up-check]
-  extends: [.build-and-test]
diff --git a/buildscripts/cibuildwheel/setup-miniconda3.sh b/buildscripts/cibuildwheel/setup-miniconda3.sh
deleted file mode 100644
index b50cae6f3647..000000000000
--- a/buildscripts/cibuildwheel/setup-miniconda3.sh
+++ /dev/null
@@ -1,28 +0,0 @@
-#!/usr/bin/env bash
-
-set -euxo pipefail
-
-# Read LLVM_VERSION from environment and error if not set
-if [ -z "${LLVM_VERSION:-}" ]; then
-    echo "Error: LLVM_VERSION environment variable is not set." >&2
-    exit 1
-fi
-
-if [ "$(uname)" = "Darwin" ]; then
-    OS_NAME="MacOSX"
-else
-    OS_NAME="Linux"
-fi
-
-echo "Installing miniconda3..."
-mkdir -p _downloads
-curl -L https://repo.anaconda.com/miniconda/Miniconda3-py311_25.5.1-1-${OS_NAME}-$(uname -m).sh -o _downloads/mini3.sh
-mkdir -p _stage
-bash _downloads/mini3.sh -b -f -p "_stage/miniconda3"
-echo "Miniconda installed"
-source "_stage/miniconda3/bin/activate" base
-export CONDA_PLUGINS_AUTO_ACCEPT_TOS=true
-
-# Create conda environment with tools and libraries for the LLVM_VERSION.
-echo "Installing llvmdev ${LLVM_VERSION}..."
-conda create -n llvmdev-${LLVM_VERSION} -c conda-forge -q -y clang=${LLVM_VERSION} clangxx=${LLVM_VERSION} clang-tools=${LLVM_VERSION} llvmdev=${LLVM_VERSION} zstd
diff --git a/buildscripts/cibuildwheel/setup-miniforge3.sh b/buildscripts/cibuildwheel/setup-miniforge3.sh
new file mode 100644
index 000000000000..2142c212429f
--- /dev/null
+++ b/buildscripts/cibuildwheel/setup-miniforge3.sh
@@ -0,0 +1,63 @@
+#!/usr/bin/env bash
+
+set -euxo pipefail
+
+MINIFORGE_VERSION="26.3.2-3"
+
+# Read LLVM_VERSION from environment and error if not set
+if [ -z "${LLVM_VERSION:-}" ]; then
+    echo "Error: LLVM_VERSION environment variable is not set." >&2
+    exit 1
+fi
+
+if [ "$(uname)" = "Darwin" ]; then
+    OS_NAME="MacOSX"
+else
+    OS_NAME="Linux"
+fi
+ARCH_NAME="$(uname -m)"
+
+case "${OS_NAME}-${ARCH_NAME}" in
+    Linux-x86_64)
+        MINIFORGE_SHA256="848194851a98903134187fbb4ab50efe87b003e0c0f808f97644b7524a62bf2c"
+        ;;
+    Linux-aarch64)
+        MINIFORGE_SHA256="2c113a69297e612b01ca0f320c22a3107a11f2ab9b573d79ac868a175945ce29"
+        ;;
+    Linux-ppc64le)
+        MINIFORGE_SHA256="df7e80ee070ccc6031e2710eb4cf81ee0012264b587306b5aa3890d3d89edd97"
+        ;;
+    MacOSX-x86_64)
+        MINIFORGE_SHA256="39273e4c89a0a1af4538010615d44ae8f44e1af41007e02def593d20f316b003"
+        ;;
+    MacOSX-arm64)
+        MINIFORGE_SHA256="59168f1e24d0a4ad9932021170809fca836cd240e183eeeb331d5bcfc0098168"
+        ;;
+    *)
+        echo "Unsupported Miniforge platform: ${OS_NAME}-${ARCH_NAME}" >&2
+        exit 1
+        ;;
+esac
+
+echo "Installing Miniforge..."
+MINIFORGE_INSTALLER="Miniforge3-${MINIFORGE_VERSION}-${OS_NAME}-${ARCH_NAME}.sh"
+MINIFORGE_RELEASE_URL="https://github.com/conda-forge/miniforge/releases/download/${MINIFORGE_VERSION}"
+MINIFORGE_URL="${MINIFORGE_RELEASE_URL}/${MINIFORGE_INSTALLER}"
+mkdir -p _downloads
+curl -fsSL "${MINIFORGE_URL}" -o _downloads/miniforge3.sh
+if command -v sha256sum >/dev/null 2>&1; then
+    printf '%s  %s\n' "${MINIFORGE_SHA256}" _downloads/miniforge3.sh | sha256sum --check -
+elif command -v shasum >/dev/null 2>&1; then
+    printf '%s  %s\n' "${MINIFORGE_SHA256}" _downloads/miniforge3.sh | shasum -a 256 --check -
+else
+    echo "No SHA256 verification command is available" >&2
+    exit 1
+fi
+mkdir -p _stage
+bash _downloads/miniforge3.sh -b -f -p "_stage/miniforge3"
+echo "Miniforge installed"
+source "_stage/miniforge3/bin/activate" base
+
+# Create conda environment with tools and libraries for the LLVM_VERSION.
+echo "Installing llvmdev ${LLVM_VERSION}..."
+conda create -n llvmdev-${LLVM_VERSION} --override-channels -c conda-forge -q -y clang=${LLVM_VERSION} clangxx=${LLVM_VERSION} clang-tools=${LLVM_VERSION} llvmdev=${LLVM_VERSION} zstd
diff --git a/buildscripts/conda-recipes/pyomp/meta.yaml b/buildscripts/conda-recipes/pyomp/meta.yaml
index b46ea3b357d3..0c689c90fea6 100644
--- a/buildscripts/conda-recipes/pyomp/meta.yaml
+++ b/buildscripts/conda-recipes/pyomp/meta.yaml
@@ -52,7 +52,8 @@ requirements:
     - llvmdev {{ LLVM_VERSION }}
     - zlib
     # require llvm-openmp for the openmp cpu runtime.
-    - llvm-openmp {{ LLVM_VERSION }}
+    - llvm-openmp 21.1.8.* # [osx]
+    - llvm-openmp {{ LLVM_VERSION }} # [not osx]
     - elfutils # [linux]
     - libffi # [linux]
   run:
@@ -60,7 +61,8 @@ requirements:
     - setuptools
     - numba >=0.62, <0.64
     # require llvm-openmp for the openmp cpu runtime.
-    - llvm-openmp {{ LLVM_VERSION }}
+    - llvm-openmp 21.1.8.* # [osx]
+    - llvm-openmp {{ LLVM_VERSION }} # [not osx]
     - lark
     - cffi
 
diff --git a/buildscripts/gitlab/build-and-test.sh b/buildscripts/gitlab/build-and-test.sh
deleted file mode 100644
index eda96566c224..000000000000
--- a/buildscripts/gitlab/build-and-test.sh
+++ /dev/null
@@ -1,48 +0,0 @@
-#!/usr/bin/env bash
-
-set -e
-
-# Create a unique temporary directory for this job.
-TMPDIR=/tmp/pyomp/${CI_JOB_ID}
-rm -rf ${TMPDIR}
-mkdir -p ${TMPDIR}
-pushd ${TMPDIR}
-
-# Set the LLVM_VERSION to use.
-export LLVM_VERSION="20.1.8"
-
-# Set the envs directory under the temporary directory.
-export CONDA_ENVS_DIRS="${TMPDIR}/_stage/miniconda3/envs"
-
-# Install miniconda and llvmdev environment.
-source ${CI_PROJECT_DIR}/buildscripts/cibuildwheel/setup-miniconda3.sh
-
-# Export environment variables for building and testing.
-export ENABLE_BUNDLED_LIBOMP="1"
-export ENABLE_BUNDLED_LIBOMPTARGET="1"
-export LLVM_DIR="${CONDA_ENVS_DIRS}/llvmdev-${LLVM_VERSION}"
-export CMAKE_PREFIX_PATH="${CONDA_PREFIX}"
-export USE_CXX11_ABI="1"
-export PIP_NO_INPUT="1"
-
-# Create and activate a conda environment with the desired Python version.
-conda create -n py-${PYOMP_CI_PYTHON_VERSION} -c conda-forge -y python=${PYOMP_CI_PYTHON_VERSION}
-conda activate py-${PYOMP_CI_PYTHON_VERSION}
-# Add extra packages needed to build openmp libraries.
-conda install -c conda-forge -y zstd libffi
-
-# Clone and fetch the commit with history for package versioning.
-git clone https://github.com/${GITHUB_PROJECT_ORG}/${GITHUB_PROJECT_NAME}.git --single-branch
-cd ${GITHUB_PROJECT_NAME}
-git fetch origin ${CI_COMMIT_SHA}
-git checkout ${CI_COMMIT_SHA}
-
-# Install pyomp.
-CC=clang CXX=clang++ python -m pip install -v .
-
-# Run host OpenMP tests.
-RUN_TARGET=0 python -m numba.runtests -v -- numba.openmp.tests.test_openmp
-# Run device (cpu target) OpenMP tests.
-OMP_TARGET_OFFLOAD=mandatory TEST_DEVICE=host RUN_TARGET=1 python -m numba.runtests -v -- numba.openmp.tests.test_openmp.TestOpenmpTarget
-
-popd
diff --git a/buildscripts/gitlab/create-conda-pkgs.sh b/buildscripts/gitlab/create-conda-pkgs.sh
deleted file mode 100755
index ddca46a3d42e..000000000000
--- a/buildscripts/gitlab/create-conda-pkgs.sh
+++ /dev/null
@@ -1,60 +0,0 @@
-#!/usr/bin/env bash
-
-set -e
-
-if [ -n "${CI_COMMIT_TAG}" ]; then
-  LABEL="main"
-else
-  LABEL="test"
-fi
-
-# Create a temporary directory for the build to clone the full repo for package
-# versioning.
-TMPDIR=/tmp/pyomp/${CI_JOB_ID}
-mkdir -p ${TMPDIR}
-pushd ${TMPDIR}
-
-# Clone and fetch the commit with history for package versioning.
-git clone https://github.com/${GITHUB_PROJECT_ORG}/${GITHUB_PROJECT_NAME}.git --single-branch
-cd ${GITHUB_PROJECT_NAME}
-git fetch origin ${CI_COMMIT_SHA}
-git checkout ${CI_COMMIT_SHA}
-
-# Set pkg dir per job to avoid conflicts.
-export CONDA_PKGS_DIRS=/tmp/pyomp/conda-pkgs-${CI_JOB_ID}
-mkdir -p "$CONDA_PKGS_DIRS"
-
-function deploy_conda() {
-  PKG="${1}"
-
-  echo "=> Conda deploy ${PKG}"
-
-  set -x
-
-  export CONDA_BLD_PATH="/tmp/pyomp/conda-build-${PYOMP_CI_BUILD_PKG}-${PYOMP_CI_PYTHON_VERSION}"
-  conda build --no-lock --no-locking --user python-for-hpc --label ${LABEL} \
-    -c python-for-hpc/label/${LABEL} -c conda-forge \
-    --python ${PYOMP_CI_PYTHON_VERSION} \
-    buildscripts/conda-recipes/${PKG}
-
-  rm -rf ${CONDA_BLD_PATH}
-  set +x
-}
-
-echo "=> Building ${PYOMP_CI_BUILD_PKG} Python version ${PYOMP_CI_PYTHON_VERSION} Label ${LABEL}"
-
-
-case ${PYOMP_CI_BUILD_PKG} in
-
-  "pyomp")
-    deploy_conda "pyomp"
-    ;;
-
-  *)
-    echo "Unknown package to build ${PYOMP_CI_BUILD_PKG}"
-    exit 1
-    ;;
-
-esac
-
-popd
diff --git a/buildscripts/modal/test_gpu.py b/buildscripts/modal/test_gpu.py
new file mode 100644
index 000000000000..b01a55e94d6a
--- /dev/null
+++ b/buildscripts/modal/test_gpu.py
@@ -0,0 +1,150 @@
+from __future__ import annotations
+
+import os
+from pathlib import Path
+import subprocess
+
+import modal
+
+
+PYTHON_VERSIONS = ("3.10", "3.11", "3.12", "3.13", "3.14")
+NUMBA_VERSION = "0.63.1"
+MINIFORGE_VERSION = "26.3.2-3"
+MINIFORGE_SHA256 = "848194851a98903134187fbb4ab50efe87b003e0c0f808f97644b7524a62bf2c"
+WHEEL_DIRECTORY = (
+    Path(__file__).resolve().parents[2] / "dist"
+    if modal.is_local()
+    else Path("/wheels")
+)
+MINIFORGE_PREFIX = "/opt/miniforge3"
+
+
+def find_linux_wheels() -> dict[str, Path]:
+    if not WHEEL_DIRECTORY.is_dir():
+        raise RuntimeError(f"Linux wheel directory does not exist: {WHEEL_DIRECTORY}")
+
+    wheels = {}
+    for python_version in PYTHON_VERSIONS:
+        abi = python_version.replace(".", "")
+        matches = sorted(WHEEL_DIRECTORY.glob(f"*cp{abi}-*x86_64.whl"))
+        if len(matches) != 1:
+            raise RuntimeError(
+                f"Expected one cp{abi} Linux x86-64 wheel in {WHEEL_DIRECTORY}, "
+                f"found {len(matches)}: {matches}"
+            )
+        wheels[python_version] = matches[0]
+    return wheels
+
+
+LINUX_WHEELS = find_linux_wheels()
+
+
+def image_build_commands() -> tuple[list[str], list[str]]:
+    environment_commands = []
+    wheel_commands = []
+    for python_version, wheel in LINUX_WHEELS.items():
+        environment = f"py{python_version.replace('.', '')}"
+        python = f"{MINIFORGE_PREFIX}/envs/{environment}/bin/python"
+        environment_commands.extend(
+            [
+                (
+                    f"{MINIFORGE_PREFIX}/bin/conda create -q -y "
+                    f"--override-channels -c conda-forge -n {environment} "
+                    f"python={python_version} pip"
+                ),
+                (
+                    f"{python} -m pip install "
+                    f"numba=={NUMBA_VERSION} lark cffi setuptools"
+                ),
+                (
+                    f'{python} -c "import numba; '
+                    f"assert numba.__version__ == '{NUMBA_VERSION}'\""
+                ),
+            ]
+        )
+        wheel_commands.append(
+            f"{python} -m pip install --no-deps /wheels/{wheel.name}"
+        )
+    return environment_commands, wheel_commands
+
+
+ENVIRONMENT_COMMANDS, WHEEL_COMMANDS = image_build_commands()
+
+
+image = (
+    modal.Image.from_registry(
+        "nvidia/cuda:12.8.1-devel-ubuntu22.04", add_python="3.12"
+    )
+    .entrypoint([])
+    .apt_install("ca-certificates", "curl")
+    .run_commands(
+        f"curl -fsSL https://github.com/conda-forge/miniforge/releases/download/"
+        f"{MINIFORGE_VERSION}/Miniforge3-{MINIFORGE_VERSION}-Linux-x86_64.sh "
+        "-o /tmp/miniforge3.sh",
+        f"printf '%s  %s\\n' {MINIFORGE_SHA256} /tmp/miniforge3.sh "
+        "| sha256sum --check -",
+        f"bash /tmp/miniforge3.sh -b -p {MINIFORGE_PREFIX}",
+    )
+    .run_commands(*ENVIRONMENT_COMMANDS)
+    .add_local_dir(WHEEL_DIRECTORY, "/wheels", copy=True)
+    .run_commands(*WHEEL_COMMANDS)
+)
+
+app = modal.App("pyomp-gpu-ci")
+
+
+@app.function(
+    image=image,
+    gpu="T4",
+    timeout=60 * 60,
+    restrict_modal_access=True,
+    single_use_containers=True,
+    block_network=True,
+)
+def test_gpu_wheels() -> None:
+    subprocess.run(["nvidia-smi", "--list-gpus"], check=True)
+
+    test_environment = os.environ.copy()
+    test_environment.update(
+        {
+            "NUMBA_DEVELOPER_MODE": "1",
+            "NUMBA_DISABLE_ERROR_MESSAGE_HIGHLIGHTING": "1",
+            "NUMBA_CAPTURED_ERRORS": "new_style",
+            "PYTHONFAULTHANDLER": "1",
+            "OMP_TARGET_OFFLOAD": "mandatory",
+            "TEST_DEVICE": "gpu",
+            "RUN_TARGET": "1",
+        }
+    )
+
+    for python_version in PYTHON_VERSIONS:
+        environment = f"py{python_version.replace('.', '')}"
+        python = f"{MINIFORGE_PREFIX}/envs/{environment}/bin/python"
+        print(f"Running GPU tests with Python {python_version} and Numba {NUMBA_VERSION}")
+        subprocess.run(
+            [
+                python,
+                "-c",
+                "from numba.openmp import print_offloading_info; "
+                "print_offloading_info()",
+            ],
+            check=True,
+            env=test_environment,
+        )
+        subprocess.run(
+            [
+                python,
+                "-m",
+                "numba.runtests",
+                "-v",
+                "--",
+                "numba.openmp.tests.test_openmp.TestOpenmpTarget",
+            ],
+            check=True,
+            env=test_environment,
+        )
+
+
+@app.local_entrypoint()
+def main() -> None:
+    test_gpu_wheels.remote()
diff --git a/pyproject.toml b/pyproject.toml
index aa6c624e2610..5069c16cff2a 100644
--- a/pyproject.toml
+++ b/pyproject.toml
@@ -68,15 +68,15 @@ CC = "/usr/bin/clang"
 CXX = "/usr/bin/clang++"
 
 [tool.cibuildwheel.macos]
-before-all = ["bash buildscripts/cibuildwheel/setup-miniconda3.sh"]
+before-all = ["bash buildscripts/cibuildwheel/setup-miniforge3.sh"]
 
 [tool.cibuildwheel.macos.environment]
 ENABLE_BUNDLED_LIBOMP = "1"
-LLVM_DIR = "${PWD}/_stage/miniconda3/envs/llvmdev-${LLVM_VERSION}/lib/cmake/llvm"
-CC = "${PWD}/_stage/miniconda3/envs/llvmdev-${LLVM_VERSION}/bin/clang"
-CXX = "${PWD}/_stage/miniconda3/envs/llvmdev-${LLVM_VERSION}/bin/clang++"
+LLVM_DIR = "${PWD}/_stage/miniforge3/envs/llvmdev-${LLVM_VERSION}/lib/cmake/llvm"
+CC = "${PWD}/_stage/miniforge3/envs/llvmdev-${LLVM_VERSION}/bin/clang"
+CXX = "${PWD}/_stage/miniforge3/envs/llvmdev-${LLVM_VERSION}/bin/clang++"
 # Set the deplioyment target to macOS 11.0.
 MACOSX_DEPLOYMENT_TARGET = "11.0"
 # Set the cmake prefix path to find libraries in the conda environment which is
 # compatible with the macos target.
-CMAKE_PREFIX_PATH = "${PWD}/_stage/miniconda3/envs/llvmdev-${LLVM_VERSION}/"
+CMAKE_PREFIX_PATH = "${PWD}/_stage/miniforge3/envs/llvmdev-${LLVM_VERSION}/"