enable 8bit and fsdp tests for xpu

Signed-off-by: jiqing-feng <jiqing.feng@intel.com>

Author: jiqing-feng

tests/test_linear8bitlt.py

@@ -172,8 +172,9 @@ def test_linear_serialization(
     assert torch.allclose(x_first.grad, x_third.grad, atol=1e-5)
 
 
-@pytest.fixture
-def linear8bit(requires_cuda):
+@pytest.fixture(params=get_available_devices(no_cpu=True))
+def linear8bit(request):
+    device = request.param
     linear = torch.nn.Linear(32, 96)
     linear_custom = Linear8bitLt(
         linear.in_features,
@@ -188,7 +189,7 @@ def linear8bit(requires_cuda):
         has_fp16_weights=False,
     )
     linear_custom.bias = linear.bias
-    linear_custom = linear_custom.cuda()
+    linear_custom = linear_custom.to(device)
     return linear_custom
 
 

tests/test_modules.py

@@ -448,34 +448,36 @@ def test_4bit_embedding_warnings(device, caplog):
     assert any("inference" in msg for msg in caplog.messages)
 
 
-def test_4bit_embedding_weight_fsdp_fix(requires_cuda):
+@pytest.mark.parametrize("device", get_available_devices(no_cpu=True))
+def test_4bit_embedding_weight_fsdp_fix(device):
     num_embeddings = 64
     embedding_dim = 32
 
     module = bnb.nn.Embedding4bit(num_embeddings=num_embeddings, embedding_dim=embedding_dim)
 
-    module.cuda()
+    module.to(device)
 
     module.weight.quant_state = None
 
-    input_tokens = torch.randint(low=0, high=num_embeddings, size=(1,), device="cuda")
+    input_tokens = torch.randint(low=0, high=num_embeddings, size=(1,), device=device)
 
     module(input_tokens)
 
     assert module.weight.quant_state is not None
 
 
-def test_4bit_linear_weight_fsdp_fix(requires_cuda):
+@pytest.mark.parametrize("device", get_available_devices(no_cpu=True))
+def test_4bit_linear_weight_fsdp_fix(device):
     inp_size = 64
     out_size = 32
 
     module = bnb.nn.Linear4bit(inp_size, out_size)
 
-    module.cuda()
+    module.to(device)
 
     module.weight.quant_state = None
 
-    input_tensor = torch.randn((1, inp_size), device="cuda")
+    input_tensor = torch.randn((1, inp_size), device=device)
 
     module(input_tensor)