Add a shared MegaMoE forward path and expert weight builder that can be used by GLM MoE without DeepSeek-only type assumptions.
Build ModelOpt NVFP4 MegaMoE sidecar weights before FlashInfer TRTLLM alignment so the original runner layout remains available as fallback.
Keep the runtime fast path gated to DeepGEMM's FP4 activation pre-dispatch because this fork does not carry the upstream DSV4 JIT FP8 pre-dispatch kernels yet.
Constraint: preserve none+flashinfer_trtllm runner layout as fallback.
Feature-flag: --moe-a2a-backend=megamoe and SGLANG_OPT_DEEPGEMM_MEGA_MOE_USE_FP4_ACTS.
Conflict-hotspots: python/sglang/srt/layers/moe/mega_moe.py, python/sglang/srt/layers/quantization/modelopt_quant.py.
Scope-risk: DeepGEMM runtime APIs and GPU e2e are not available on this local machine.
Tested: PYTHONPYCACHEPREFIX=/private/tmp/sglang_pycache python3 -m py_compile python/sglang/srt/layers/moe/mega_moe.py python/sglang/srt/layers/quantization/modelopt_quant.py.
Tested: git diff --check.
Not-tested: GLM 5.2 MegaMoE GPU runtime; local environment lacks deep_gemm and target GPU.
c13d4556ff