[AMD CI] Add 2-GPU sgl-kernel Tests (#17555)

Co-authored-by: YC Tseng <yctseng@amd.com>

sgl-kernel/tests/test_amd_deterministic_custom_allreduce.py

@@ -12,12 +12,13 @@ This test compares:
 2. Deterministic kernel (different batch size)
 
 Usage:
-    python test_amd_deterministic_custom_allreduce.py
+    pytest test_amd_deterministic_custom_allreduce.py
 """
 
 import multiprocessing as mp
 import socket
 
+import pytest
 import torch
 import torch.distributed as dist
 
@@ -266,5 +267,14 @@ def main():
         p.join()
 
 
+@pytest.mark.skipif(
+    not torch.cuda.is_available() or torch.cuda.device_count() < 2,
+    reason="Requires at least 2 CUDA GPUs",
+)
+def test_deterministic_custom_allreduce():
+    """Test that deterministic custom all-reduce produces consistent results."""
+    main()
+
+
 if __name__ == "__main__":
     main()

sgl-kernel/tests/test_amd_nccl_allreduce_determinism.py

@@ -10,12 +10,13 @@ This test compares:
 2. Default all-reduce (different batch size) - typically NON-DETERMINISTIC for bfloat16
 
 Usage:
-    python test_ar.py
+    pytest test_amd_nccl_allreduce_determinism.py
 """
 
 import multiprocessing as mp
 import socket
 
+import pytest
 import torch
 import torch.distributed as dist
 
@@ -194,5 +195,14 @@ def main():
         p.join()
 
 
+@pytest.mark.skipif(
+    not torch.cuda.is_available() or torch.cuda.device_count() < 2,
+    reason="Requires at least 2 CUDA GPUs",
+)
+def test_nccl_allreduce_determinism():
+    """Test NCCL all-reduce determinism behavior with varying batch sizes."""
+    main()
+
+
 if __name__ == "__main__":
     main()