Replace hardcoded CUDA device with get_device() for XPU support (#13599)

Co-authored-by: Ma Mingfei <mingfei.ma@intel.com>

test/registered/quant/test_fp8_kernel.py

@@ -11,6 +11,13 @@ from sglang.test.test_utils import CustomTestCase
 
 register_cuda_ci(est_time=10, suite="stage-b-test-1-gpu-large")
 
+from sglang.srt.utils import get_device, is_cuda, is_xpu
+
+_is_cuda = is_cuda()
+_is_xpu = is_xpu()
+
+device = get_device()
+
 
 class TestFP8Base(CustomTestCase):
     @classmethod
@@ -26,7 +33,7 @@ class TestFP8Base(CustomTestCase):
     @staticmethod
     def _make_A(M, K, group_size, out_dtype):
         quant_A = torch.rand(
-            M, K // group_size, group_size, dtype=torch.float32, device="cuda"
+            M, K // group_size, group_size, dtype=torch.float32, device=device
         )
         # -1 ~ 1
         quant_A = quant_A * 2 - 1
@@ -38,7 +45,7 @@ class TestFP8Base(CustomTestCase):
         quant_A = quant_A.to(out_dtype).to(torch.float32)
 
         # create scale and A
-        scale = torch.rand(M, K // group_size, dtype=torch.float32, device="cuda")
+        scale = torch.rand(M, K // group_size, dtype=torch.float32, device=device)
         scale /= fmax
         A = quant_A * scale[..., None]
 
@@ -60,7 +67,7 @@ class TestFP8Base(CustomTestCase):
             N_aligned // group_size,
             group_size,
             dtype=torch.float32,
-            device="cuda",
+            device=device,
         )
         quant_B = quant_B * 2 - 1
 
@@ -77,7 +84,7 @@ class TestFP8Base(CustomTestCase):
             N_aligned // group_size,
             1,
             dtype=torch.float32,
-            device="cuda",
+            device=device,
         )
         scale /= fmax
 
@@ -91,8 +98,9 @@ class TestFP8Base(CustomTestCase):
 
 class TestPerTokenGroupQuantFP8(TestFP8Base):
     def test_per_token_group_quant_fp8(self):
-        if torch.cuda.get_device_capability()[0] < 9:
+        if _is_cuda and torch.cuda.get_device_capability()[0] < 9:
             return
+
         A, A_quant_gt, scale_gt = self._make_A(
             M=self.M, K=self.K, group_size=self.group_size, out_dtype=self.quant_type
         )
@@ -107,8 +115,14 @@ class TestPerTokenGroupQuantFP8(TestFP8Base):
 
 class TestW8A8BlockFP8Matmul(TestFP8Base):
     def test_w8a8_block_fp8_matmul(self):
-        if torch.cuda.get_device_capability()[0] < 9:
+        if _is_cuda and torch.cuda.get_device_capability()[0] < 9:
             return
+        elif _is_xpu:
+            # XPU doesn't provide traditional capability info like CUDA
+            pass
+        else:
+            return
+
         A, A_quant_gt, A_scale_gt = self._make_A(
             M=self.M, K=self.K, group_size=self.group_size, out_dtype=self.quant_type
         )