[torch.compile] add deepseek v2 compile (vllm-project#9775)

Signed-off-by: youkaichao <youkaichao@gmail.com>
Signed-off-by: Randall Smith <Randall.Smith@amd.com>

Author: youkaichao

vllm/model_executor/models/deepseek_v2.py

@@ -28,6 +28,7 @@
 from transformers import PretrainedConfig
 
 from vllm.attention import Attention, AttentionMetadata
+from vllm.compilation.decorators import support_torch_compile
 from vllm.config import CacheConfig
 from vllm.distributed import (get_pp_group,
                               get_tensor_model_parallel_world_size,
@@ -403,6 +404,7 @@ def forward(
         return hidden_states, residual
 
 
+@support_torch_compile
 class DeepseekV2Model(nn.Module):
 
     fall_back_to_pt_during_load = False