CI Updates

.github/workflows/test-runner.yml

@@ -32,11 +32,6 @@ on:
         type: string
         default: ""
         description: "CPU architecture for testing: icelake, cascadelake (default: platform default runner)"
-      torch_nightly:
-        type: boolean
-        default: false
-        description: "Install nightly torch (--pre, no version pin) instead of a stable release"
-
 env:
   BNB_SKIP_CMAKE: 1
 
@@ -210,16 +205,18 @@ jobs:
         if: inputs.platform == 'windows'
         uses: ilammy/msvc-dev-cmd@v1.13.0
 
+      - name: Install torch (nightly)
+        if: inputs.torch_version == 'nightly'
+        run: pip install --pre torch --index-url ${{ inputs.pypi_index }}
+
+      - name: Install torch
+        if: inputs.torch_version != 'nightly'
+        run: pip install --pre torch==${{ inputs.torch_version }} --index-url ${{ inputs.pypi_index }}
+
       - name: Install dependencies
         run: |
-          if [[ "${{ inputs.torch_nightly }}" == "true" ]]; then
-            pip install --pre torch --index-url ${{ inputs.pypi_index }}
-          else
-            pip install torch==${{ inputs.torch_version }} --index-url ${{ inputs.pypi_index }}
-          fi
           pip install -e ".[test]" -v
           pip install pytest-cov
-        shell: bash
 
       # Windows: Downgrade NumPy for torch<2.4.1 compatibility
       # See: https://github.com/pytorch/pytorch/issues/131668

.github/workflows/tests-nightly.yml

@@ -49,7 +49,7 @@ jobs:
       cpu_type: ${{ matrix.cpu_type }}
 
   test-cuda:
-    name: CUDA
+    name: CUDA ${{ matrix.cuda_version }} (${{ matrix.platform }}, ${{ matrix.gpu_type }}, ${{ matrix.torch_version }})
     if: github.repository == 'bitsandbytes-foundation/bitsandbytes'
     strategy:
       fail-fast: false
@@ -65,17 +65,16 @@ jobs:
             torch_version: "2.3.1"
             pypi_index: "https://download.pytorch.org/whl/cu118"
           - cuda_version: "12.6.3"
-            torch_version: "2.8.0"
+            torch_version: "2.11.0"
             pypi_index: "https://download.pytorch.org/whl/cu126"
           - cuda_version: "12.8.1"
-            torch_version: "2.9.1"
+            torch_version: "2.10.0"
             pypi_index: "https://download.pytorch.org/whl/cu128"
           - cuda_version: "13.0.2"
-            torch_version: "2.11.0"
+            torch_version: "2.12.0"
             pypi_index: "https://download.pytorch.org/whl/cu130"
           - cuda_version: "13.2.0"
-            torch_version: "2.12.0"
-            torch_nightly: true
+            torch_version: "nightly"
             pypi_index: "https://download.pytorch.org/whl/nightly/cu132"
 
           # Windows CUDA Tests - T4 GPU (CUDA 11.8 only, multiple torch versions)
@@ -102,5 +101,4 @@ jobs:
       cuda_version: ${{ matrix.cuda_version }}
       gpu_type: ${{ matrix.gpu_type }}
       torch_version: ${{ matrix.torch_version }}
-      torch_nightly: ${{ matrix.torch_nightly || false }}
       pypi_index: ${{ matrix.pypi_index }}

.github/workflows/tests-pr.yml

@@ -58,25 +58,25 @@ jobs:
       cpu_type: ${{ matrix.cpu_type }}
 
   test-cuda:
-    name: CUDA
+    name: CUDA ${{ matrix.cuda_version }} (${{ matrix.platform }}, ${{ matrix.gpu_type }}, ${{ matrix.torch_version }})
     if: github.repository == 'bitsandbytes-foundation/bitsandbytes'
     strategy:
       fail-fast: false
       matrix:
         platform: [linux-x64]
         gpu_type: [T4, A10, L40S]
-        cuda_version: ["11.8.0", "12.8.1", "13.0.2"]
+        cuda_version: ["11.8.0", "12.6.3", "13.0.2"]
 
         include:
           # Map CUDA version to torch version and PyPI index
           - cuda_version: "11.8.0"
             torch_version: "2.3.1"
             pypi_index: "https://download.pytorch.org/whl/cu118"
-          - cuda_version: "12.8.1"
-            torch_version: "2.9.1"
-            pypi_index: "https://download.pytorch.org/whl/cu128"
-          - cuda_version: "13.0.2"
+          - cuda_version: "12.6.3"
             torch_version: "2.11.0"
+            pypi_index: "https://download.pytorch.org/whl/cu126"
+          - cuda_version: "13.0.2"
+            torch_version: "2.12.0"
             pypi_index: "https://download.pytorch.org/whl/cu130"
 
           # Windows CUDA test - single configuration

tests/test_linear4bit.py

@@ -453,6 +453,13 @@ def test_linear4bit_torch_compile_activation_checkpointing(device, quant_type, c
         pytest.skip("This configuration is not supported on HPU.")
     if device == "cuda" and platform.system() == "Windows":
         pytest.skip("Triton is not officially supported on Windows")
+    if (
+        device == "cuda"
+        and torch.cuda.is_available()
+        and torch.cuda.get_device_capability() < (8, 0)
+        and torch.__version__ >= (2, 13, 0, "dev")
+    ):
+        pytest.xfail("no compiled frames error on SM75")
     dim = 256
     batch_size = 16
     compute_dtype = torch.bfloat16