feat(glm): route GLM MoE through megamoe backend

Add the GLM MoE MegaMoE fast path before the existing DeepEP/normal branch selection.

Keep the existing none+flashinfer_trtllm path as the miss path when MegaMoE weights, envs, or token caps do not satisfy should_use_mega_moe().

Match upstream MegaMoE shared-expert TP1 semantics for the separate shared expert path.

Constraint: do not change the baseline none+flashinfer_trtllm dispatch path.

Feature-flag: --moe-a2a-backend=megamoe.

Conflict-hotspots: python/sglang/srt/models/glm4_moe.py.

Scope-risk: shared-expert TP1 behavior matters when shared expert fusion is disabled.

Tested: PYTHONPYCACHEPREFIX=/private/tmp/sglang_pycache python3 -m py_compile python/sglang/srt/models/glm4_moe.py.

Tested: git diff --check.

Not-tested: GLM 5.2 MegaMoE GPU e2e; local environment lacks target runtime and hardware.

python/sglang/srt/models/glm4_moe.py

@@ -415,6 +415,7 @@ class Glm4MoeSparseMoeBlock(nn.Module):
                     if get_moe_a2a_backend().is_deepep()
                     or get_moe_a2a_backend().is_mooncake()
                     or get_moe_a2a_backend().is_flashinfer()
+                    or get_moe_a2a_backend().is_megamoe()
                     or should_use_flashinfer_cutlass_moe_fp4_allgather()
                     else {}
                 ),
@@ -463,6 +464,13 @@ class Glm4MoeSparseMoeBlock(nn.Module):
         should_allreduce_fusion: bool = False,
         use_reduce_scatter: bool = False,
     ) -> torch.Tensor:
+        from sglang.srt.layers.moe.mega_moe import (
+            forward_mega_moe,
+            should_use_mega_moe,
+        )
+
+        if should_use_mega_moe(self, hidden_states):
+            return forward_mega_moe(self, hidden_states, forward_batch)
 
         if not get_moe_a2a_backend().is_deepep():
             if (