diff --git a/vllm/v1/attention/backends/rocm_aiter_fa.py b/vllm/v1/attention/backends/rocm_aiter_fa.py
index c7f925817..ad454daa5 100644
--- a/vllm/v1/attention/backends/rocm_aiter_fa.py
+++ b/vllm/v1/attention/backends/rocm_aiter_fa.py
@@ -18,6 +18,7 @@ from vllm.config import VllmConfig
 from vllm.logger import init_logger
 from vllm.platforms import current_platform
 from vllm.utils.math_utils import cdiv
+from vllm.utils.platform_utils import get_cu_count
 from vllm.v1.attention.backends.utils import (
     AttentionCGSupport,
     AttentionMetadataBuilder,
@@ -38,7 +39,7 @@ if current_platform.is_rocm():
         return min(65536 // x.element_size(), triton.next_power_of_2(head_dim))
 
     def num_programs(total_tokens):
-        return min(total_tokens, current_platform.get_cu_count())
+        return min(total_tokens, get_cu_count())
 
     @triton.jit
     def cp_mha_gather_cache_kernel(