From 77db26723842adc657c17906e89540ebeea46220 Mon Sep 17 00:00:00 2001
From: Matthew Douglas <38992547+matthewdouglas@users.noreply.github.com>
Date: Mon, 28 Apr 2025 17:33:07 -0400
Subject: [PATCH 01/10] Run unit tests on GH Actions

---
 .github/scripts/build-cuda.sh |  13 ++-
 .github/workflows/tests.yml   | 172 ++++++++++++++++++++++++++++++++++
 2 files changed, 181 insertions(+), 4 deletions(-)
 create mode 100644 .github/workflows/tests.yml

diff --git a/.github/scripts/build-cuda.sh b/.github/scripts/build-cuda.sh
index 9c92e0d01..315bcf60d 100644
--- a/.github/scripts/build-cuda.sh
+++ b/.github/scripts/build-cuda.sh
@@ -2,14 +2,19 @@
 declare build_arch
 declare build_os
 declare cuda_version
+declare cuda_targets
 
 set -xeuo pipefail
 
-# By default, target Maxwell through Hopper.
-build_capability="50;52;60;61;70;75;80;86;89;90"
+if [ -n "${cuda_targets}" ]; then
+    # By default, target Maxwell through Hopper.
+    build_capability="50;52;60;61;70;75;80;86;89;90"
 
-# CUDA 12.8: Add sm100 and sm120; remove < sm75 to align with PyTorch 2.7+cu128 minimum
-[[ "${cuda_version}" == 12.8.* ]] && build_capability="75;80;86;89;90;100;120"
+    # CUDA 12.8: Add sm100 and sm120; remove < sm75 to align with PyTorch 2.7+cu128 minimum
+    [[ "${cuda_version}" == 12.8.* ]] && build_capability="75;80;86;89;90;100;120"
+else
+    build_capability="${cuda_targets}"
+fi
 
 [[ "${build_os}" = windows-* ]] && python3 -m pip install ninja
 
diff --git a/.github/workflows/tests.yml b/.github/workflows/tests.yml
new file mode 100644
index 000000000..d322d0006
--- /dev/null
+++ b/.github/workflows/tests.yml
@@ -0,0 +1,172 @@
+name: Unit tests
+
+on:
+  workflow_dispatch:
+
+
+jobs:
+
+  build-cpu:
+    strategy:
+      matrix:
+        os: [ubuntu-22.04, windows-2025]
+        arch: [x86_64]
+    runs-on: ${{ matrix.os }}
+    steps:
+      - uses: actions/checkout@v4
+
+      - name: Setup MSVC
+        if: startsWith(matrix.os, 'windows')
+        uses: ilammy/msvc-dev-cmd@v1.13.0 # to use cl
+
+      - name: Build C++
+        run: bash .github/scripts/build-cpu.sh
+        env:
+          build_os: ${{ matrix.os }}
+          build_arch: ${{ matrix.arch }}
+
+      - name: Upload build artifact
+        uses: actions/upload-artifact@v4
+        with:
+          name: lib_cpu_${{ matrix.os }}_${{ matrix.arch }}
+          path: output/*
+          retention-days: 7
+
+  build-cuda:
+    strategy:
+      matrix:
+        cuda_version: ["11.8.0", "12.8.1"]
+        include:
+          - os: ubuntu-22.04
+            arch: x86_64
+          - os: windows-2025
+            arch: x86_64
+
+    runs-on: ${{ matrix.os }}
+
+    steps:
+      - uses: actions/checkout@v4
+
+      - name: Install CUDA Toolkit
+        uses: Jimver/cuda-toolkit@v0.2.23
+        if: startsWith(matrix.os, 'windows')
+        id: cuda-toolkit
+        with:
+          cuda: ${{ matrix.cuda_version }}
+          method: "network"
+          sub-packages: '["nvcc","cudart","cusparse","cublas","thrust","nvrtc_dev","cublas_dev","cusparse_dev"]'
+
+      - name: Setup MSVC
+        if: startsWith(matrix.os, 'windows')
+        uses: ilammy/msvc-dev-cmd@v1.13.0 # to use cl
+
+      # We're running on T4 only for now, so we only target sm75.
+      - name: Build C++ / CUDA
+        run: bash .github/scripts/build-cuda.sh
+        env:
+          build_os: ${{ matrix.os }}
+          build_arch: x86_64
+          cuda_version: ${{ matrix.cuda_version }}
+          cuda_targets: "75"
+
+      - name: Upload build artifact
+        uses: actions/upload-artifact@v4
+        with:
+          name: lib_cuda_${{matrix.cuda_version}}_${{ matrix.os }}_${{ matrix.arch }}
+          path: output/*
+          retention-days: 7
+
+  cpu-tests:
+    needs: build-cpu
+    strategy:
+      fail-fast: false
+      matrix:
+        os: [ubuntu-22.04, windows-2025]
+        arch: [x86_64]
+        torch_version: ["2.4.1", "2.7.0"]
+    runs-on: ${{ matrix.os }}
+    env:
+      BNB_TEST_DEVICE: cpu
+    steps:
+      - name: Show CPU Information
+        run: |
+          if [[ $RUNNER_OS == 'Linux' ]]; then
+            lscpu
+          else
+            systeminfo | findstr /C:"Processor"
+          fi
+
+      - uses: actions/checkout@v4
+
+      - name: Download build artifact
+        uses: actions/download-artifact@v4
+        with:
+          name: lib_cpu_${{ matrix.os }}_${{ matrix.arch }}
+          path: bitsandbytes/
+          merge-multiple: true
+
+      - name: Setup Python
+        uses: actions/setup-python@v5
+        with:
+          python-version: 3.9
+
+      - name: Install dependencies
+        run: |
+          pip install torch==${{ matrix.torch_version }} --index-url https://download.pytorch.org/whl/cpu
+          pip install -e ".[test]"
+          pip install pytest-cov
+
+      - name: Show installed packages
+        run: pip list
+
+      - name: Run tests
+        run: pytest
+
+  cuda-tests:
+    needs: build-cuda
+    strategy:
+      fail-fast: false
+      matrix:
+        os: [ubuntu-22.04, windows-2025]
+        arch: [x86_64]
+        torch_version: ["2.4.1", "2.7.0"]
+        include:
+          - torch_version: "2.4.1"
+            cuda_version: "11.8.0"
+            pypi_index: "https://download.pytorch.org/whl/cu118"
+          - torch_version: "2.7.0"
+            cuda_version: "12.8.1"
+            pypi_index: "https://download.pytorch.org/whl/cu128"
+    runs-on:
+      labels: ${{ contains(matrix.os, 'windows') && 'CUDA-Windows-x64' || 'CUDA-Linux-x64' }}
+    env:
+      BNB_TEST_DEVICE: cuda
+    steps:
+      - name: Show GPU Information
+        run: nvidia-smi
+
+      - uses: actions/checkout@v4
+
+      - name: Download build artifact
+        uses: actions/download-artifact@v4
+        with:
+          name: lib_cuda_${{ matrix.cuda_version }}_${{ matrix.os }}_${{ matrix.arch }}
+          path: bitsandbytes/
+          merge-multiple: true
+
+      - name: Setup Python
+        uses: actions/setup-python@v5
+        with:
+          python-version: 3.9
+
+      - name: Install dependencies
+        run: |
+          pip install torch==${{ matrix.torch_version }} --index-url ${{ matrix.pypi_index }}
+          pip install -e ".[test]"
+          pip install pytest-cov
+
+      - name: Show installed packages
+        run: pip list
+
+      - name: Run tests
+        run: pytest

From 18c07bdff7138bad4a4c5f143f80bd363db9d59c Mon Sep 17 00:00:00 2001
From: Matthew Douglas <38992547+matthewdouglas@users.noreply.github.com>
Date: Mon, 28 Apr 2025 17:36:28 -0400
Subject: [PATCH 02/10] fix

---
 .github/scripts/build-cuda.sh | 6 +++---
 1 file changed, 3 insertions(+), 3 deletions(-)

diff --git a/.github/scripts/build-cuda.sh b/.github/scripts/build-cuda.sh
index 315bcf60d..430058285 100644
--- a/.github/scripts/build-cuda.sh
+++ b/.github/scripts/build-cuda.sh
@@ -6,14 +6,14 @@ declare cuda_targets
 
 set -xeuo pipefail
 
-if [ -n "${cuda_targets}" ]; then
+if [[ -v cuda_targets ]; then
+    build_capability="${cuda_targets}"
+else
     # By default, target Maxwell through Hopper.
     build_capability="50;52;60;61;70;75;80;86;89;90"
 
     # CUDA 12.8: Add sm100 and sm120; remove < sm75 to align with PyTorch 2.7+cu128 minimum
     [[ "${cuda_version}" == 12.8.* ]] && build_capability="75;80;86;89;90;100;120"
-else
-    build_capability="${cuda_targets}"
 fi
 
 [[ "${build_os}" = windows-* ]] && python3 -m pip install ninja

From ad90a8429c0aca535af92b682644c1964fc5f1e0 Mon Sep 17 00:00:00 2001
From: Matthew Douglas <38992547+matthewdouglas@users.noreply.github.com>
Date: Mon, 28 Apr 2025 17:37:24 -0400
Subject: [PATCH 03/10] fix

---
 .github/scripts/build-cuda.sh | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/.github/scripts/build-cuda.sh b/.github/scripts/build-cuda.sh
index 430058285..be8e98704 100644
--- a/.github/scripts/build-cuda.sh
+++ b/.github/scripts/build-cuda.sh
@@ -6,7 +6,7 @@ declare cuda_targets
 
 set -xeuo pipefail
 
-if [[ -v cuda_targets ]; then
+if [[ -v cuda_targets ]]; then
     build_capability="${cuda_targets}"
 else
     # By default, target Maxwell through Hopper.

From f1bd41bc03f31662b400ad542482f978e703b733 Mon Sep 17 00:00:00 2001
From: Matthew Douglas <38992547+matthewdouglas@users.noreply.github.com>
Date: Mon, 28 Apr 2025 17:53:32 -0400
Subject: [PATCH 04/10] trigger workflow

---
 .github/workflows/tests.yml | 2 ++
 1 file changed, 2 insertions(+)

diff --git a/.github/workflows/tests.yml b/.github/workflows/tests.yml
index d322d0006..4219d75ba 100644
--- a/.github/workflows/tests.yml
+++ b/.github/workflows/tests.yml
@@ -2,6 +2,8 @@ name: Unit tests
 
 on:
   workflow_dispatch:
+  push:
+    branches: [testing-ci]
 
 
 jobs:

From 03a27b0d7356fdc4e2e70cce36466037d1bcef71 Mon Sep 17 00:00:00 2001
From: Matthew Douglas <38992547+matthewdouglas@users.noreply.github.com>
Date: Mon, 28 Apr 2025 18:00:00 -0400
Subject: [PATCH 05/10] Update

---
 .github/workflows/tests.yml | 10 ++++------
 1 file changed, 4 insertions(+), 6 deletions(-)

diff --git a/.github/workflows/tests.yml b/.github/workflows/tests.yml
index 4219d75ba..85ac1faad 100644
--- a/.github/workflows/tests.yml
+++ b/.github/workflows/tests.yml
@@ -38,12 +38,8 @@ jobs:
     strategy:
       matrix:
         cuda_version: ["11.8.0", "12.8.1"]
-        include:
-          - os: ubuntu-22.04
-            arch: x86_64
-          - os: windows-2025
-            arch: x86_64
-
+        os: [ubuntu-22.04, windows-2025]
+        arch: [x86_64]
     runs-on: ${{ matrix.os }}
 
     steps:
@@ -57,6 +53,7 @@ jobs:
           cuda: ${{ matrix.cuda_version }}
           method: "network"
           sub-packages: '["nvcc","cudart","cusparse","cublas","thrust","nvrtc_dev","cublas_dev","cusparse_dev"]'
+          use-github-cache: false
 
       - name: Setup MSVC
         if: startsWith(matrix.os, 'windows')
@@ -91,6 +88,7 @@ jobs:
       BNB_TEST_DEVICE: cpu
     steps:
       - name: Show CPU Information
+        shell: bash
         run: |
           if [[ $RUNNER_OS == 'Linux' ]]; then
             lscpu

From 2df730c623c1d4080f891cfcb9f6d4018cb0f69c Mon Sep 17 00:00:00 2001
From: Matthew Douglas <38992547+matthewdouglas@users.noreply.github.com>
Date: Mon, 28 Apr 2025 18:38:48 -0400
Subject: [PATCH 06/10] Update

---
 .github/workflows/tests.yml | 20 +++++++++-----------
 1 file changed, 9 insertions(+), 11 deletions(-)

diff --git a/.github/workflows/tests.yml b/.github/workflows/tests.yml
index 85ac1faad..373895327 100644
--- a/.github/workflows/tests.yml
+++ b/.github/workflows/tests.yml
@@ -5,6 +5,9 @@ on:
   push:
     branches: [testing-ci]
 
+concurrency:
+  group: ${{ github.workflow }}-${{ github.event.pull_request.number || github.ref }}
+  cancel-in-progress: true
 
 jobs:
 
@@ -31,7 +34,7 @@ jobs:
         uses: actions/upload-artifact@v4
         with:
           name: lib_cpu_${{ matrix.os }}_${{ matrix.arch }}
-          path: output/*
+          path: output/${{ matrix.os }}/${{ matrix.arch }}/*
           retention-days: 7
 
   build-cuda:
@@ -72,7 +75,7 @@ jobs:
         uses: actions/upload-artifact@v4
         with:
           name: lib_cuda_${{matrix.cuda_version}}_${{ matrix.os }}_${{ matrix.arch }}
-          path: output/*
+          path: output/${{ matrix.os }}/${{ matrix.arch }}/*
           retention-days: 7
 
   cpu-tests:
@@ -87,15 +90,6 @@ jobs:
     env:
       BNB_TEST_DEVICE: cpu
     steps:
-      - name: Show CPU Information
-        shell: bash
-        run: |
-          if [[ $RUNNER_OS == 'Linux' ]]; then
-            lscpu
-          else
-            systeminfo | findstr /C:"Processor"
-          fi
-
       - uses: actions/checkout@v4
 
       - name: Download build artifact
@@ -154,6 +148,10 @@ jobs:
           path: bitsandbytes/
           merge-multiple: true
 
+      - name: Show files
+        run: ls -lR .
+        shell: bash
+
       - name: Setup Python
         uses: actions/setup-python@v5
         with:

From 08871a42f1857e78f2d16c3137c35526359fab6c Mon Sep 17 00:00:00 2001
From: Matthew Douglas <38992547+matthewdouglas@users.noreply.github.com>
Date: Mon, 28 Apr 2025 20:22:11 -0400
Subject: [PATCH 07/10] Update

---
 .github/workflows/tests.yml |  6 +-----
 tests/test_functional.py    | 16 +++++++++++++++-
 2 files changed, 16 insertions(+), 6 deletions(-)

diff --git a/.github/workflows/tests.yml b/.github/workflows/tests.yml
index 373895327..de3c7b2af 100644
--- a/.github/workflows/tests.yml
+++ b/.github/workflows/tests.yml
@@ -85,7 +85,7 @@ jobs:
       matrix:
         os: [ubuntu-22.04, windows-2025]
         arch: [x86_64]
-        torch_version: ["2.4.1", "2.7.0"]
+        torch_version: ["2.7.0"]
     runs-on: ${{ matrix.os }}
     env:
       BNB_TEST_DEVICE: cpu
@@ -148,10 +148,6 @@ jobs:
           path: bitsandbytes/
           merge-multiple: true
 
-      - name: Show files
-        run: ls -lR .
-        shell: bash
-
       - name: Setup Python
         uses: actions/setup-python@v5
         with:
diff --git a/tests/test_functional.py b/tests/test_functional.py
index ee2b52429..c8a390733 100644
--- a/tests/test_functional.py
+++ b/tests/test_functional.py
@@ -728,6 +728,9 @@ def test_int8_double_quant(self, dim1, dim2):
         ),
     )
     def test_integrated_int8_linear_matmul(self, device, dim1, dim4, inner):
+        if device == "cpu" and inner > 2048:
+            pytest.skip("Slow on CPU")
+
         for i in range(k):
             A = torch.randn(dim1, inner, device=device).half()
             B = torch.randn(dim4, inner, device=device).half()
@@ -1316,7 +1319,18 @@ def test_gemv_4bit(self, device, dim, dtype, storage_type, quant_storage, double
         if dtype == torch.float16:
             if dim <= 512:
                 assert err1 < 7e-5
-                assert relerr1 < 0.0008
+
+                # TODO(matthewdouglas): On T4, dim=128-fp16-fc2-fp4-DQ will have relerror ~ 0.00092727
+                if (
+                    device == "cuda"
+                    and double_quant
+                    and storage_type == "fp4"
+                    and kind == "fc2"
+                    and torch.cuda.get_device_capability() == (7, 5)
+                ):
+                    assert relerr1 < 0.00093
+                else:
+                    assert relerr1 < 0.0008
             else:
                 assert err1 < 6e-5
                 assert relerr1 < 2e-4

From 5b6a9b98cc82feb7628bde5452873de44ed19624 Mon Sep 17 00:00:00 2001
From: Matthew Douglas <38992547+matthewdouglas@users.noreply.github.com>
Date: Tue, 29 Apr 2025 12:34:04 -0400
Subject: [PATCH 08/10] Run tests nightly

---
 .github/workflows/tests.yml | 15 +++++++++------
 1 file changed, 9 insertions(+), 6 deletions(-)

diff --git a/.github/workflows/tests.yml b/.github/workflows/tests.yml
index de3c7b2af..9aa2acaf0 100644
--- a/.github/workflows/tests.yml
+++ b/.github/workflows/tests.yml
@@ -2,11 +2,14 @@ name: Unit tests
 
 on:
   workflow_dispatch:
+  schedule:
+    # Every day at 02:15 AM UTC
+    - cron: "15 2 * * *"
   push:
     branches: [testing-ci]
 
 concurrency:
-  group: ${{ github.workflow }}-${{ github.event.pull_request.number || github.ref }}
+  group: ${{ github.workflow }}-${{ github.ref }}
   cancel-in-progress: true
 
 jobs:
@@ -123,13 +126,13 @@ jobs:
       matrix:
         os: [ubuntu-22.04, windows-2025]
         arch: [x86_64]
-        torch_version: ["2.4.1", "2.7.0"]
+        cuda_version: ["11.8.0", "12.8.1"]
         include:
-          - torch_version: "2.4.1"
-            cuda_version: "11.8.0"
+          - cuda_version: "11.8.0"
+            torch_version: "2.4.1"
             pypi_index: "https://download.pytorch.org/whl/cu118"
-          - torch_version: "2.7.0"
-            cuda_version: "12.8.1"
+          - cuda_version: "12.8.1"
+            torch_version: "2.7.0"
             pypi_index: "https://download.pytorch.org/whl/cu128"
     runs-on:
       labels: ${{ contains(matrix.os, 'windows') && 'CUDA-Windows-x64' || 'CUDA-Linux-x64' }}

From fb85e9ce5592c055c622fad91f51b9d04fe9d2e3 Mon Sep 17 00:00:00 2001
From: Matthew Douglas <38992547+matthewdouglas@users.noreply.github.com>
Date: Tue, 29 Apr 2025 12:42:43 -0400
Subject: [PATCH 09/10] Disable paged optimizer test on Windows

---
 tests/test_optim.py | 4 ++++
 1 file changed, 4 insertions(+)

diff --git a/tests/test_optim.py b/tests/test_optim.py
index 9358a2e9b..0d86da7d8 100644
--- a/tests/test_optim.py
+++ b/tests/test_optim.py
@@ -1,6 +1,7 @@
 import os
 from os.path import join
 import shutil
+import sys
 import time
 import uuid
 
@@ -168,6 +169,9 @@ def rm_path(path):
 @pytest.mark.parametrize("dim1", [1024], ids=id_formatter("dim1"))
 @pytest.mark.parametrize("dim2", [32, 1024, 4097, 1], ids=id_formatter("dim2"))
 def test_optimizer32bit(requires_cuda, dim1, dim2, gtype, optim_name):
+    if optim_name.startswith("paged_") and sys.platform == "win32":
+        pytest.skip("Paged optimizers can have issues on Windows.")
+
     if gtype == torch.bfloat16 and optim_name in ["momentum", "rmsprop"]:
         pytest.skip()
     if dim1 == 1 and dim2 == 1:

From 818144b85284e2f76f37b7f4eb2e47a988fa6fc4 Mon Sep 17 00:00:00 2001
From: Matthew Douglas <38992547+matthewdouglas@users.noreply.github.com>
Date: Tue, 29 Apr 2025 13:17:54 -0400
Subject: [PATCH 10/10] Skip unit tests on Windows for CUDA 12.x (driver on
 runner is too old)

---
 .github/workflows/tests.yml | 5 +++++
 1 file changed, 5 insertions(+)

diff --git a/.github/workflows/tests.yml b/.github/workflows/tests.yml
index 9aa2acaf0..a2917b9bb 100644
--- a/.github/workflows/tests.yml
+++ b/.github/workflows/tests.yml
@@ -134,6 +134,11 @@ jobs:
           - cuda_version: "12.8.1"
             torch_version: "2.7.0"
             pypi_index: "https://download.pytorch.org/whl/cu128"
+        exclude:
+          # Our current T4 Windows runner has a driver too old (471.11)
+          # and cannot support CUDA 12+. Skip for now.
+          - os: windows-2025
+            cuda_version: "12.8.1"
     runs-on:
       labels: ${{ contains(matrix.os, 'windows') && 'CUDA-Windows-x64' || 'CUDA-Linux-x64' }}
     env: