[Test] Skip DeepEP MoE layer tests without P2P access (#47946)

Signed-off-by: Tyler Michael Smith <[email protected]>
Signed-off-by: Tyler Michael Smith <[email protected]>
Co-authored-by: Claude <[email protected]>

tests/kernels/moe/test_moe_layer.py

@@ -36,6 +36,9 @@ from vllm.distributed import (
     get_eplb_group,
     tensor_model_parallel_all_gather,
 )
+from vllm.distributed.device_communicators.all_reduce_utils import (
+    gpu_p2p_access_check,
+)
 from vllm.distributed.eplb.eplb_communicator import create_eplb_communicator
 from vllm.distributed.eplb.rebalance_execute import rearrange_expert_weights_inplace
 from vllm.forward_context import set_forward_context
@@ -106,6 +109,8 @@ if has_deep_ep():
 if has_nixl_ep():
     BACKENDS += ["nixl_ep"]
 
+DEEPEP_BACKENDS = {"deepep_high_throughput", "deepep_low_latency"}
+
 QUANT_METHODS = [
     None,
     "fp8",
@@ -416,6 +421,22 @@ def generate_valid_test_configs(
     return configs
 
 
+@functools.cache
+def visible_devices_have_peer_access(world_size: int) -> bool:
+    if not current_platform.is_cuda():
+        return True
+
+    try:
+        return all(
+            gpu_p2p_access_check(src, dst)
+            for src in range(world_size)
+            for dst in range(world_size)
+            if src != dst
+        )
+    except RuntimeError:
+        return False
+
+
 # TODO: break this up into sections
 def is_valid_config(config: MoETestConfig) -> tuple[bool, str | None]:
     # routed_input_transform only makes sense with shared_experts (latent MoE)
@@ -1802,6 +1823,9 @@ def test_moe_layer(
     if enable_eplb and not use_ep:
         pytest.skip("EPLB requires EP.")
 
+    if backend in DEEPEP_BACKENDS and not visible_devices_have_peer_access(world_size):
+        pytest.skip("DeepEP backends require peer access between visible GPUs.")
+
     verbosity = pytestconfig.getoption("verbose")
 
     if os.environ.get("VLLM_LOGGING_LEVEL") is None: