[fix]解决moe_fused_gate编译错误，去掉mla中mtp部分的修改

restore the default settings of disable_cascade_attn
add VLLM_USE_OPT_ZEROS to replace triton_ (torch.zeros)
set default_max_num_batched_tokens = 10240
update qwen3_moe of layernorm

vllm/attention/layer.py

@@ -219,9 +219,14 @@ class Attention(nn.Module):
         if self.use_output:
             output_shape = (output_shape
                             if output_shape is not None else query.shape)
-            output = torch.zeros(output_shape,
-                                 dtype=query.dtype,
-                                 device=query.device)
+            if envs.VLLM_USE_OPT_ZEROS:
+                output = torch.empty(output_shape,
+                                    dtype=query.dtype,
+                                    device=query.device)
+            else:
+                output = torch.zeros(output_shape,
+                                    dtype=query.dtype,
+                                    device=query.device)
             hidden_size = output_shape[-1]
             # We skip reshaping query, key and value tensors for the MLA
             # backend since these tensors have different semantics and are

vllm/config.py

@@ -326,7 +326,7 @@ 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

@@ -1664,7 +1664,7 @@ 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

@@ -1107,6 +1107,10 @@ environment_variables: dict[str, Callable[[], Any]] = {
     "VLLM_USE_LIGHTOP":
         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", "False").lower() in

vllm/model_executor/models/deepseek_mtp.py

@@ -224,7 +224,7 @@ class DeepSeekMTP(nn.Module, SupportsPP):
         params_dict = dict(self.named_parameters())
         loaded_params: set[str] = set()
         for name, loaded_weight in weights:
-            if "rotary_emb.inv_freq" in name:
+            if "rotary_emb.inv_freq" in name or "indexer" in name:
                 continue
             spec_layer = get_spec_layer_idx_from_weight_name(self.config, name)
             if spec_layer is None:

vllm/model_executor/models/qwen3_moe.py

@@ -234,7 +234,7 @@ 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,
@@ -242,7 +242,7 @@ 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)