diff --git a/python/sglang/srt/configs/moe_model_registry.py b/python/sglang/srt/configs/moe_model_registry.py index 0fbae099cd..104db14eef 100644 --- a/python/sglang/srt/configs/moe_model_registry.py +++ b/python/sglang/srt/configs/moe_model_registry.py @@ -14,6 +14,7 @@ _DEEPEP_V2_MODELS: dict[str, bool] = { "DeepseekV3ForCausalLM": False, "DeepseekV4ForCausalLM": False, "Qwen3MoeForCausalLM": False, + "Glm5NextForConditionalGeneration": False, } diff --git a/python/sglang/srt/models/glm5_next.py b/python/sglang/srt/models/glm5_next.py index dad0c2fdc1..1c901e997f 100644 --- a/python/sglang/srt/models/glm5_next.py +++ b/python/sglang/srt/models/glm5_next.py @@ -1094,7 +1094,11 @@ class Glm5NextModel(nn.Module): ) self.layers_to_capture = [] self.dflash_capture = False - if get_moe_a2a_backend().is_deepep() or get_moe_a2a_backend().is_mooncake(): + if ( + get_moe_a2a_backend().is_deepep() + or get_moe_a2a_backend().is_mooncake() + or get_moe_a2a_backend().is_deepep_v2() + ): self.enable_a2a_moe = True else: self.enable_a2a_moe = False