diff --git a/vllm/config.py b/vllm/config.py
index 421c036003c13..6c003bdf2cfc0 100644
--- a/vllm/config.py
+++ b/vllm/config.py
@@ -22,7 +22,7 @@
     get_hf_text_config, get_pooling_config,
     get_sentence_transformer_tokenizer_config, is_encoder_decoder, uses_mrope)
 from vllm.utils import (GiB_bytes, cuda_device_count_stateless, get_cpu_memory,
-                        identity, is_mi250, print_warning_once,
+                        identity, is_mi250, is_navi, print_warning_once,
                         resolve_obj_by_qualname)
 
 if TYPE_CHECKING:
@@ -1036,6 +1036,12 @@ def __init__(
                 "Disabled the custom all-reduce kernel because it is not "
                 "working correctly on multi AMD MI250.")
 
+        if is_navi() and self.tensor_parallel_size <= 2:
+            self.disable_custom_all_reduce = True
+            logger.info(
+                "Disabled the custom all-reduce kernel because it is not "
+                "working correctly when using two AMD Navi GPUs.")
+
     @property
     def use_ray(self) -> bool:
         return self.distributed_executor_backend == "ray" or (