diff --git a/python/sglang/multimodal_gen/runtime/platforms/cuda.py b/python/sglang/multimodal_gen/runtime/platforms/cuda.py
index cebd7dd53..d59326de1 100644
--- a/python/sglang/multimodal_gen/runtime/platforms/cuda.py
+++ b/python/sglang/multimodal_gen/runtime/platforms/cuda.py
@@ -246,6 +246,13 @@ class CudaPlatformBase(Platform):
         elif selected_backend == AttentionBackendEnum.SAGE_SLA_ATTN:
             logger.info("Using Sage Sparse Linear Attention backend")
             return "sglang.multimodal_gen.runtime.layers.attention.backends.sparse_linear_attn.SageSparseLinearAttentionBackend"
+        elif selected_backend == AttentionBackendEnum.FA2:
+            from sglang.multimodal_gen.runtime.layers.attention.backends.flash_attn_2 import (  # noqa: F401
+                FlashAttention2Backend,
+            )
+
+            logger.info("Using FlashAttention2 backend")
+            return "sglang.multimodal_gen.runtime.layers.attention.backends.flash_attn_2.FlashAttention2Backend"
         elif selected_backend in [
             AttentionBackendEnum.FA,
         ]: