Tests: skip torch.compile for cuda on windows

Author: matthewdouglas

.github/workflows/tests.yml

@@ -205,27 +205,40 @@ jobs:
             torch_version: "2.7.0"
             pypi_index: "https://download.pytorch.org/whl/cu128"
 
-          # Add torch 2.7+cu118 for Windows.
-          - os: windows-2025
-            arch: x86_64
-            gpu: T4
-            runner: CUDA-Windows-x64
-            cuda_version: "11.8.0"
-            torch_version: "2.7.0"
-            pypi_index: "https://download.pytorch.org/whl/cu118"
 
-          # L40S runners
+          # Linux L40S runners
           - os: ubuntu-22.04
             gpu: L40S
             runner: bandb-aws-g6e-4xlarge-plus-use1-public-80
 
-          # T4 runners
+          # Linux T4 runners
           - os: ubuntu-22.04
             gpu: T4
             runner: bandb-aws-g4dn-4xlarge-plus-use1-public-80
+
+          # Specific Windows runners using cu118
+          - os: windows-2025
+            arch: x86_64
+            gpu: T4
+            runner: CUDA-Windows-x64
+            cuda_version: "11.8.0"
+            torch_version: "2.2.0"
+            pypi_index: "https://download.pytorch.org/whl/cu118"
           - os: windows-2025
+            arch: x86_64
             gpu: T4
             runner: CUDA-Windows-x64
+            cuda_version: "11.8.0"
+            torch_version: "2.6.0"
+            pypi_index: "https://download.pytorch.org/whl/cu118"
+          - os: windows-2025
+            arch: x86_64
+            gpu: T4
+            runner: CUDA-Windows-x64
+            cuda_version: "11.8.0"
+            torch_version: "2.7.0"
+            pypi_index: "https://download.pytorch.org/whl/cu118"
+
         exclude:
           # Our current T4 Windows runner has a driver too old (471.11)
           # and cannot support CUDA 12+. Skip for now.

tests/test_linear4bit.py

@@ -300,6 +300,9 @@ def test_linear4bit_torch_compile(device, quant_type, compute_dtype, compress_st
     if fullgraph and torch.__version__ < (2, 8):
         pytest.skip("fullgraph mode requires torch 2.8 or higher")
 
+    if device == "cuda" and platform.system() == "Windows":
+        pytest.skip("Triton is not officially supported on Windows")
+
     # Has a strange regression on Linux aarch64 CPU in torch==2.6.0 when fullgraph=False.
     if (
         not fullgraph

tests/test_linear8bitlt.py

@@ -234,6 +234,9 @@ def test_linear8bit_serialization(linear8bit):
 @pytest.mark.parametrize("mode", ["default", "reduce-overhead"], ids=id_formatter("mode"))
 @pytest.mark.skipif(torch.__version__ < (2, 4), reason="Not supported in torch < 2.4")
 def test_linear8bitlt_torch_compile(device, threshold, bias, fullgraph, mode):
+    if device == "cuda" and platform.system() == "Windows":
+        pytest.skip("Triton is not officially supported on Windows")
+
     dim = 256
     batch_size = 16