update VLLM_USE_OPT_CAT

vllm/envs.py

@@ -1094,15 +1094,15 @@ environment_variables: dict[str, Callable[[], Any]] = {
     "VLLM_USE_GLOBAL_CACHE13":
         lambda: (os.environ.get("VLLM_USE_GLOBAL_CACHE13", "False").lower() in
                  ("true", "1")),
-    # vLLM will use global cache for moe
+    # vLLM will use lightop for deepseek-v3
     "VLLM_USE_LIGHTOP":
         lambda: (os.environ.get("VLLM_USE_LIGHTOP", "False").lower() in
                  ("true", "1")),
-    # vLLM will use global cache for moe
+    # vLLM will use opt cat for deepseek-v3
     "VLLM_USE_OPT_CAT":
-        lambda: (os.environ.get("VLLM_USE_OPT_CAT", "True").lower() in
+        lambda: (os.environ.get("VLLM_USE_OPT_CAT", "False").lower() in
                  ("true", "1")),  
-    # vLLM will use opt merge_aatn_states,not triton
+    # vLLM will use opt merge_aatn_states, not triton
     "VLLM_USE_MERGE_ATTN_STATES_OPT":
         lambda: (os.environ.get("VLLM_USE_MERGE_ATTN_STATES_OPT", "True").lower() in
                  ("true", "1")),  

vllm/model_executor/layers/fused_moe/layer.py

@@ -1442,8 +1442,8 @@ class FusedMoE(torch.nn.Module):
         if current_platform.is_tpu():
             return self.forward_impl(hidden_states, router_logits)
         else:
-            return torch.ops.vllm.moe_forward(hidden_states, router_logits, shared_output,
-                                              self.layer_name)
+            return torch.ops.vllm.moe_forward(hidden_states, router_logits,
+                                              self.layer_name, shared_output)
 
     def forward_impl_chunked(self, full_hidden_states: torch.Tensor,
                              full_router_logits: torch.Tensor):

vllm/model_executor/model_loader/utils.py

@@ -247,6 +247,8 @@ def get_model_architecture(
             if (architectures == ['DeepseekV3ForCausalLM'] or architectures == ['DeepSeekMTPModel']):
                 if not envs.is_set("VLLM_USE_LIGHTOP"):
                     os.environ['VLLM_USE_LIGHTOP'] = '1'
+                if not envs.is_set("VLLM_USE_OPT_CAT"):
+                    os.environ['VLLM_USE_OPT_CAT'] = '1'
                 
             if os.getenv('GEMM_PAD') != '1': 
                 os.environ['GEMM_PAD'] = '0'

vllm/v1/attention/backends/mla/test_concat.py

@@ -7,7 +7,7 @@ import torch
 import math
 import vllm.envs as envs
 
-if envs.VLLM_USE_LIGHTOP:
+if envs.VLLM_USE_OPT_CAT:
     from lightop import ds_cat
 
 def test_concat_Acc_prefill(shape_pair, dim):