diff --git a/python/sglang/srt/layers/moe/mega_moe.py b/python/sglang/srt/layers/moe/mega_moe.py
index db7ce6871..5fb0f3879 100644
--- a/python/sglang/srt/layers/moe/mega_moe.py
+++ b/python/sglang/srt/layers/moe/mega_moe.py
@@ -91,6 +91,17 @@ def _get_mega_moe_symm_buffer(
     return buf
 
 
+def _deep_gemm_supports_fp8_mega_moe() -> bool:
+    if not deep_gemm_wrapper.ENABLE_JIT_DEEPGEMM:
+        return False
+    if not torch.cuda.is_available():
+        return False
+    # Current sgl-deep-gemm exposes fp8_mega_moe, but its Blackwell MegaMoE
+    # symmetric buffer only accepts fp8xfp4. SM90 uses a separate buffer path
+    # that can back fp8_mega_moe.
+    return torch.cuda.get_device_capability()[0] == 9
+
+
 def should_use_mega_moe(moe, hidden_states: torch.Tensor) -> bool:
     if not get_moe_a2a_backend().is_megamoe():
         return False
@@ -98,7 +109,7 @@ def should_use_mega_moe(moe, hidden_states: torch.Tensor) -> bool:
         return False
     weight_format = getattr(moe.experts, "_mega_moe_weight_format", None)
     if weight_format == "fp8":
-        return deep_gemm_wrapper.ENABLE_JIT_DEEPGEMM
+        return _deep_gemm_supports_fp8_mega_moe()
     if weight_format != "fp4":
         return False
     if not envs.SGLANG_OPT_DEEPGEMM_MEGA_MOE_USE_FP4_ACTS.get():
diff --git a/test/registered/unit/moe/test_glm_megamoe.py b/test/registered/unit/moe/test_glm_megamoe.py
index 73a7a738d..e3e43c2a6 100644
--- a/test/registered/unit/moe/test_glm_megamoe.py
+++ b/test/registered/unit/moe/test_glm_megamoe.py
@@ -82,7 +82,7 @@ class TestGLMMegaMoE(unittest.TestCase):
         with patch.object(
             mega_moe, "get_moe_a2a_backend", return_value=_MegaBackend()
         ), patch.object(
-            mega_moe.deep_gemm_wrapper, "ENABLE_JIT_DEEPGEMM", True
+            mega_moe, "_deep_gemm_supports_fp8_mega_moe", return_value=True
         ), envs.SGLANG_OPT_DEEPGEMM_MEGA_MOE_USE_FP4_ACTS.override(False):
             self.assertTrue(mega_moe.should_use_mega_moe(moe, hidden_states))