diff --git a/vllm/model_executor/warmup/kernel_warmup.py b/vllm/model_executor/warmup/kernel_warmup.py
index e4f575d412..c3725064a6 100644
--- a/vllm/model_executor/warmup/kernel_warmup.py
+++ b/vllm/model_executor/warmup/kernel_warmup.py
@@ -106,6 +106,12 @@ def kernel_warmup(worker: "Worker"):
         )
 
 
+# TODO: remove once FlashInfer upstream fixes the persistent file cache
+# to resolve collisions like `use_8x4_sf_layout=True/False`, which causes
+# invalid tactics to be chosen
+_FLASHINFER_USE_PERSISTENT_CACHE = False
+
+
 def flashinfer_autotune(runner: "GPUModelRunner") -> None:
     """
     Autotune FlashInfer operations.
@@ -122,6 +128,16 @@ def flashinfer_autotune(runner: "GPUModelRunner") -> None:
     import vllm.utils.flashinfer as fi_utils
     from vllm.distributed.parallel_state import get_world_group
 
+    if not _FLASHINFER_USE_PERSISTENT_CACHE:
+        with torch.inference_mode(), fi_utils.autotune():
+            runner._dummy_run(
+                num_tokens=runner.scheduler_config.max_num_batched_tokens,
+                skip_eplb=True,
+                is_profile=True,
+            )
+        get_world_group().barrier()
+        return
+
     world = get_world_group()
     is_leader = world.rank_in_group == 0