[feat]支持glm4_moe_mtp使用torch compile，实现mtp cudagraph模式

vllm/model_executor/models/glm4_moe_mtp.py

@@ -49,6 +49,7 @@ from .glm4_moe import (
     get_spec_layer_idx_from_weight_name,
 )
 from .utils import maybe_prefix
+from vllm.compilation.decorators import support_torch_compile
 
 
 class SharedHead(nn.Module):
@@ -184,6 +185,7 @@ class Glm4MoeMultiTokenPredictor(nn.Module):
         return logits
 
 
+@support_torch_compile
 class Glm4MoeMTP(nn.Module, Glm4MixtureOfExperts):
     def __init__(self, *, vllm_config: VllmConfig, prefix: str = ""):
         super().__init__()