diff --git a/python/sglang/srt/layers/attention/attention_registry.py b/python/sglang/srt/layers/attention/attention_registry.py
index 2bf271c29..c89fe809c 100644
--- a/python/sglang/srt/layers/attention/attention_registry.py
+++ b/python/sglang/srt/layers/attention/attention_registry.py
@@ -183,6 +183,7 @@ def attn_backend_wrapper(runner: "ModelRunner", full_attn_backend: "AttentionBac
     ), "hybrid_gdn can only be used with non-MLA models."
 
     if cfg := runner.mambaish_config:
+        from sglang.srt.layers.attention.fla.utils import check_environments
         from sglang.srt.layers.attention.hybrid_linear_attn_backend import (
             GDNAttnBackend,
             HybridLinearAttnBackend,
@@ -190,6 +191,7 @@ def attn_backend_wrapper(runner: "ModelRunner", full_attn_backend: "AttentionBac
         )
         from sglang.srt.utils import is_blackwell, is_npu
 
+        check_environments()
         if runner.hybrid_gdn_config is not None:
             if is_blackwell():
                 assert (
diff --git a/python/sglang/srt/layers/attention/fla/utils.py b/python/sglang/srt/layers/attention/fla/utils.py
index 3caf70de5..8613d611d 100644
--- a/python/sglang/srt/layers/attention/fla/utils.py
+++ b/python/sglang/srt/layers/attention/fla/utils.py
@@ -58,9 +58,6 @@ def check_environments():
     return None
 
 
-check_environments()
-
-
 def get_abs_err(x, y):
     return (x.detach() - y.detach()).flatten().abs().max().item()