add VLLM_USE_OPT_ZEROS to replace triton_ (torch.zeros)

set default_max_num_batched_tokens = 10240
update qwen3_moe of layernorm
off lightop of moe_fused_gate

vllm/attention/layer.py

@@ -297,8 +297,13 @@ class Attention(nn.Module, AttentionLayerBase):
         if self.use_output:
             output_shape = (output_shape
                             if output_shape is not None else query.shape)
+            if envs.VLLM_USE_OPT_ZEROS:
+                output = torch.empty(output_shape,
+                                    dtype=query.dtype,
+                                    device=query.device)
+            else:
                 output = torch.zeros(output_shape,
-                                 dtype=output_dtype,
+                                    dtype=query.dtype,
                                     device=query.device)
             hidden_size = output_shape[-1]
             # We skip reshaping query, key and value tensors for the MLA

vllm/config/model.py

@@ -199,7 +199,7 @@ class ModelConfig:
     """Whether to disable sliding window. If True, we will disable the sliding
     window functionality of the model, capping to sliding window size. If the
     model does not support sliding window, this argument is ignored."""
-    disable_cascade_attn: bool = True
+    disable_cascade_attn: bool = False
     """Disable cascade attention for V1. While cascade attention does not
     change the mathematical correctness, disabling it could be useful for
     preventing potential numerical issues. Note that even if this is set to

vllm/engine/arg_utils.py

@@ -1639,7 +1639,7 @@ class EngineArgs:
             # TODO(woosuk): Tune the default values for other hardware.
             default_max_num_batched_tokens = {
                 UsageContext.LLM_CLASS: 8192,
-                UsageContext.OPENAI_API_SERVER: 2048,
+                UsageContext.OPENAI_API_SERVER: 10240,
             }
             default_max_num_seqs = {
                 UsageContext.LLM_CLASS: 256,

vllm/envs.py

@@ -227,6 +227,7 @@ if TYPE_CHECKING:
     VLLM_USE_APEX_RN: bool = False
     VLLM_USE_GLOBAL_CACHE13: bool = False
     VLLM_USE_LIGHTOP: bool = False
+    VLLM_USE_OPT_ZEROS: bool = False
     VLLM_USE_OPT_CAT: bool = False
     VLLM_USE_OPT_MOE_SUM: bool = False
     VLLM_USE_LIGHTOP_MOE_SUM: bool = False
@@ -1605,6 +1606,11 @@ environment_variables: dict[str, Callable[[], Any]] = {
         lambda: (os.environ.get("VLLM_USE_LIGHTOP", "False").lower() in
                  ("true", "1")),
         
+    # vLLM will use elenmentwise not triton_
+    "VLLM_USE_OPT_ZEROS":
+        lambda: (os.environ.get("VLLM_USE_OPT_ZEROS", "False").lower() in
+                 ("true", "1")),
+        
     # vLLM will use opt cat for deepseek-v3
     "VLLM_USE_OPT_CAT":
         lambda: (os.environ.get("VLLM_USE_OPT_CAT", "True").lower() in

vllm/model_executor/layers/fused_moe/layer.py

@@ -1793,7 +1793,8 @@ class FusedMoE(CustomOp):
             assert topk_group is not None
             assert num_expert_group is not None
             if use_fused_gate:
-                if envs.VLLM_USE_LIGHTOP:
+                # if envs.VLLM_USE_LIGHTOP:
+                if False:
                     from lightop import op as op
                     topk_weights, topk_ids = op.moe_fused_gate(
                         router_logits,

vllm/model_executor/model_loader/utils.py

@@ -194,8 +194,8 @@ def _get_model_architecture(
                 os.environ['LM_NN'] = '1'
                 
             if architectures in [['DeepseekV3ForCausalLM'], ['DeepSeekMTPModel']]:
-                # if not envs.is_set("VLLM_USE_LIGHTOP"):
-                #     os.environ['VLLM_USE_LIGHTOP'] = '1'
+                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'
 

vllm/model_executor/models/qwen3_moe.py

@@ -284,7 +284,7 @@ class Qwen3MoeAttention(nn.Module):
         if envs.VLLM_USE_APEX_RN:
             q_by_head = self.q_norm.forward_apex(q_by_head)
         else:
-            q_by_head = self.q_norm(q_by_head)
+            q_by_head = self.q_norm.forward_cuda(q_by_head)
         q = q_by_head.view(q.shape)
 
         k_by_head = k.view(*k.shape[:-1], k.shape[-1] // self.head_dim,
@@ -292,7 +292,7 @@ class Qwen3MoeAttention(nn.Module):
         if envs.VLLM_USE_APEX_RN:
             k_by_head = self.k_norm.forward_apex(k_by_head)
         else:
-            k_by_head = self.k_norm(k_by_head)
+            k_by_head = self.k_norm.forward_cuda(k_by_head)
         k = k_by_head.view(k.shape)
         q, k = self.rotary_emb(positions, q, k)
         attn_output = self.attn(q, k, v)