fix(moe): gate fp8 megamoe on deepgemm capability

Avoid entering the FP8 MegaMoE fused path when the installed DeepGEMM exposes fp8_mega_moe but cannot allocate an fp8xfp8 MegaMoE symmetric buffer. Current B300 sgl-deep-gemm asserts that MegaMoE buffer mma_type is fp8xfp4, so FP8 weights must fall back until dependency support lands.

Tested: PYTHONPYCACHEPREFIX=/private/tmp/sglang_pycache python3 -m py_compile python/sglang/srt/layers/moe/mega_moe.py test/registered/unit/moe/test_glm_megamoe.py

Tested: git diff --check

Remote-tested: root@95.133.252.48 B300 DeepGEMM capability probe showed fp8xfp8 MegaMoE buffer fails with mma_type_str == fp8xfp4.

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():