diff --git a/vllm/env_override.py b/vllm/env_override.py
index 78270c2bee37..aed826a87e42 100644
--- a/vllm/env_override.py
+++ b/vllm/env_override.py
@@ -758,3 +758,51 @@ def _exec_then_patch(module):
 
 
 _patch_cpp_indirect_assert_if_needed()
+
+# ===================================================
+# Triton Autotuner disable
+# ===================================================
+# Replace Autotuner.run so it always picks configs[0] and skips benchmarking.
+# Used to eliminate autotuning variability when measuring kernel perf.
+# Gated on VLLM_TRITON_FORCE_FIRST_CONFIG=1 so it is opt-in.
+from vllm.triton_utils import HAS_TRITON  # noqa: E402
+
+
+def _disable_triton_autotuner():
+    if not HAS_TRITON:
+        return
+    if os.environ.get("VLLM_TRITON_FORCE_FIRST_CONFIG", "0").strip().lower() not in (
+        "1",
+        "true",
+    ):
+        return
+    import importlib
+
+    Autotuner = importlib.import_module("triton.runtime.autotuner").Autotuner
+    seen_kernels: set[str] = set()
+
+    def _run_first_config(self, *args, **kwargs):
+        config = self.configs[0]
+        self.best_config = config
+        kernel_name = getattr(self.fn, "__name__", repr(self.fn))
+        if kernel_name not in seen_kernels:
+            seen_kernels.add(kernel_name)
+            logger.info(
+                "[triton-autotune-disabled] kernel=%s configs=%d picked=%s",
+                kernel_name,
+                len(self.configs),
+                config,
+            )
+        if config.pre_hook is not None:
+            full_nargs = {
+                **dict(zip(self.arg_names, args)),
+                **kwargs,
+                **config.all_kwargs(),
+            }
+            config.pre_hook(full_nargs)
+        return self.fn.run(*args, **kwargs, **config.all_kwargs())
+
+    Autotuner.run = _run_first_config
+
+
+_disable_triton_autotuner()