add VLLM_USE_OPT_ZEROS to replace triton_ (torch.zeros)

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/envs.py

@@ -165,6 +165,7 @@ 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_MUL_ADD: bool = False
@@ -1105,6 +1106,10 @@ 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