[CI] Check B200 NUMA mapping against sysfs numa_node (#40055)

Co-authored-by: Mohammad Angkad <mohammad.angkad@radixark.ai>

test/registered/utils/test_numa_utils.py

@@ -4,6 +4,8 @@ import unittest
 from contextlib import ExitStack
 from unittest.mock import MagicMock, patch
 
+import torch
+
 from sglang.srt.environ import envs
 from sglang.srt.utils.numa_utils import (
     _handle_numa_bind_failure,
@@ -12,6 +14,7 @@ from sglang.srt.utils.numa_utils import (
     _numactl_cpu_mem_args,
     _probe_numactl_args,
     _query_numa_node_for_gpu,
+    _read_pci_numa_node,
     _strip_memory_args,
     configure_subprocess,
     get_numa_node_if_available,
@@ -306,17 +309,19 @@ class TestGraceBlackwellNumaTopology(unittest.TestCase):
     "Requires 4-GPU B200 hardware",
 )
 class TestB200NumaTopology(unittest.TestCase):
-    """Hardware test validating expected NUMA topology on 4-GPU B200."""
+    """Hardware test for the 4-GPU B200 NUMA mapping; layout is BIOS-dependent."""
 
     def test_gpu_numa_mapping(self):
         self.assertEqual(_gpu_count, 4)
-        numa_nodes = {
+        for gpu_id in range(_gpu_count):
-            _query_single_numa_node_for_gpu(gpu_id) for gpu_id in range(_gpu_count)
+            # Not NVML: the index mapping under test must not validate itself.
-        }
+            props = torch.cuda.get_device_properties(gpu_id)
-        self.assertEqual(
+            pci_address = f"{props.pci_domain_id:04x}:{props.pci_bus_id:02x}:{props.pci_device_id:02x}.0"
-            len(numa_nodes),
+            self.assertIn(
-            1,
+                _read_pci_numa_node(pci_address)[0],
-            f"Expected all visible 4-GPU B200 devices on one NUMA node, got {numa_nodes}",
+                _query_numa_node_for_gpu(gpu_id),
+                f"GPU {gpu_id} ({pci_address}): NVML memory affinity omits the "
+                f"sysfs numa_node",
             )