fix AMD CI failure of NUMA binding (#17184)

python/sglang/srt/mem_cache/hiradix_cache.py

@@ -24,7 +24,7 @@ from sglang.srt.mem_cache.radix_cache import (
     split_node_hash_value,
 )
 from sglang.srt.metrics.collector import StorageMetricsCollector
-from sglang.srt.utils import bind_to_closest_numa_node, is_numa_available
+from sglang.srt.utils import bind_to_closest_numa_node_cuda
 
 if TYPE_CHECKING:
     from sglang.srt.mem_cache.cache_init_params import CacheInitParams
@@ -44,12 +44,8 @@ class HiRadixCache(RadixCache):
                     "Page first layout is not supported with direct IO backend, switching to page first direct layout"
                 )
 
-        if (
-            not server_args.disable_hicache_numa_detect
-            and is_numa_available()
-            and torch.cuda.is_available()
-        ):
-            bind_to_closest_numa_node()
+        if not server_args.disable_hicache_numa_detect:
+            bind_to_closest_numa_node_cuda()
 
         self.page_size = params.page_size
         self.kv_cache = params.token_to_kv_pool_allocator.get_kvcache()

python/sglang/srt/utils/common.py

@@ -3956,7 +3956,7 @@ def is_numa_available() -> bool:
         return False
 
 
-def get_system_gpu_count() -> int:
+def get_system_nvgpu_count() -> int:
     """
     Get the total number of GPUs in the system (not affected by CUDA_VISIBLE_DEVICES).
 
@@ -3978,7 +3978,7 @@ def get_system_gpu_count() -> int:
 
 
 @lru_cache(maxsize=1)
-def get_current_device_numa_node() -> int:
+def get_current_device_numa_node_cuda() -> int:
     """
     Retrieve the NUMA node ID of the CPU socket closest to the currently active CUDA device.
 
@@ -4017,7 +4017,7 @@ def get_current_device_numa_node() -> int:
 
     # Fall back: distribute GPUs evenly across NUMA nodes
     numa_count = get_numa_node_count()
-    gpu_count = get_system_gpu_count()
+    gpu_count = get_system_nvgpu_count()
 
     if gpu_count >= numa_count:
         gpus_per_numa = gpu_count // numa_count  # >= 1
@@ -4031,11 +4031,20 @@ def get_current_device_numa_node() -> int:
     return numa_node
 
 
-def bind_to_closest_numa_node():
+def nvgpu_available() -> bool:
+    if not torch.cuda.is_available():
+        return False
+    if torch.version.cuda is None:
+        return False
+    return True
+
+
+def bind_to_closest_numa_node_cuda():
     """
     Bind the current process to the NUMA node closest to the active CUDA device.
 
     Uses `numa` library calls via ctypes to set the CPU affinity of the process.
     """
-    node_id = get_current_device_numa_node()
-    numa_bind_to_node(node_id)
+    if is_numa_available() and nvgpu_available():
+        node_id = get_current_device_numa_node_cuda()
+        numa_bind_to_node(node_id)