[Tests] Use requires_gpu, fix missing gpu test skip, add explicit test for gpu from gha (#1264)

## Purpose ##
* Update all tests to use `requires_gpu` decorator
* Add GPU mark skip for `test_compressor_stacking`, which requires a GPU
* Add an explicit GPU test for GHA, so as to unambiguously catch
situations where CUDA is not properly installed on a runner

---------

Signed-off-by: Kyle Sayers <[email protected]>

Author: kylesayrs

tests/llmcompressor/pytorch/utils/test_helpers.py

@@ -14,6 +14,7 @@
     tensors_to_device,
     tensors_to_precision,
 )
+from tests.testing_utils import requires_gpu
 
 
 @pytest.mark.unit
@@ -55,6 +56,7 @@ def test_tensors_to_device_cpu(tensors):
     os.getenv("NM_ML_SKIP_PYTORCH_TESTS", False),
     reason="Skipping pytorch tests",
 )
+@requires_gpu
 @pytest.mark.parametrize(
     "tensors",
     [
@@ -69,7 +71,6 @@ def test_tensors_to_device_cpu(tensors):
         [[torch.randn(1, 8)], torch.randn(8, 8)],
     ],
 )
-@pytest.mark.skipif(not torch.cuda.is_available(), reason="requires cuda availability")
 def test_tensors_to_device_cuda(tensors):
     out = tensors_to_device(tensors, "cuda")
 
@@ -364,6 +365,7 @@ def test_tensors_module_forward(module, tensors, check_feat_lab_inp):
     os.getenv("NM_ML_SKIP_PYTORCH_TESTS", False),
     reason="Skipping pytorch tests",
 )
+@requires_gpu
 @pytest.mark.parametrize(
     "module,tensors,check_feat_lab_inp",
     [
@@ -417,7 +419,6 @@ def test_tensors_module_forward(module, tensors, check_feat_lab_inp):
         ),
     ],
 )
-@pytest.mark.skipif(not torch.cuda.is_available(), reason="requires cuda availability")
 def test_tensors_module_forward_cuda(module, tensors, check_feat_lab_inp):
     module = module.to("cuda")
     tensors = tensors_to_device(tensors, "cuda")
@@ -471,6 +472,7 @@ def test_tensor_sparsity(tensor, dim, expected_sparsity):
     os.getenv("NM_ML_SKIP_PYTORCH_TESTS", False),
     reason="Skipping pytorch tests",
 )
+@requires_gpu
 @pytest.mark.parametrize(
     "tensor,dim,expected_sparsity",
     [
@@ -490,7 +492,6 @@ def test_tensor_sparsity(tensor, dim, expected_sparsity):
         ),
     ],
 )
-@pytest.mark.skipif(not torch.cuda.is_available(), reason="requires cuda availability")
 def test_tensor_sparsity_cuda(tensor, dim, expected_sparsity):
     tensor = tensor.to("cuda")
     sparsity = tensor_sparsity(tensor, dim)

tests/llmcompressor/transformers/compression/test_has_gpu.py

@@ -0,0 +1,13 @@
+import os
+
+import pytest
+import torch
+
+
+@pytest.mark.skipif(os.getenv("GITHUB_ACTIONS") != "true", reason="Only run for GHA")
+def test_has_gpu():
+    """
+    This test exists purely to raise an error if
+    a runner performs transformers tests without a GPU
+    """
+    assert torch.cuda.is_available()

tests/llmcompressor/transformers/sparsification/test_compress_tensor_utils.py

@@ -30,6 +30,7 @@
     modify_save_pretrained,
     patch_tied_tensors_bug,
 )
+from tests.testing_utils import requires_gpu
 
 
 @pytest.mark.parametrize(
@@ -275,7 +276,7 @@ def test_model_reload(offload, torch_dtype, tie_word_embeddings, device_map, tmp
         assert torch.equal(model_dict[key].cpu(), reloaded_dict[key].cpu())
 
 
-@pytest.mark.skipif(not torch.cuda.is_available(), reason="requires gpu")
+@requires_gpu
 @pytest.mark.parametrize(
     "offload,torch_dtype,tie_word_embeddings,device_map",
     [
@@ -340,7 +341,7 @@ def test_model_shared_tensors(
         assert not torch.equal(lm_head, embed_tokens)
 
 
-@pytest.mark.skipif(not torch.cuda.is_available(), reason="requires gpu")
+@requires_gpu
 @pytest.mark.parametrize(
     "offload,torch_dtype,tie_word_embeddings,device_map",
     [
@@ -356,6 +357,7 @@ def test_model_shared_tensors_gpu(
     )
 
 
+@requires_gpu
 @pytest.mark.parametrize(
     "model_stub, recipe, sparse_format, quant_format",
     [