diff --git a/vllm/engine/arg_utils.py b/vllm/engine/arg_utils.py
index 921f314..d350eb5 100644
--- a/vllm/engine/arg_utils.py
+++ b/vllm/engine/arg_utils.py
@@ -2277,6 +2277,15 @@ class EngineArgs:
             jit_monitor_verbose=self.jit_monitor_verbose,
         )
 
+        # Blockwise-FP8 decode on Hopper picks the CUTLASS kernel, which is
+        # tuned for throughput shapes. A decode-bound server runs this GEMM at
+        # M equal to the running batch (single digits), where the Triton
+        # blockwise kernel has per-shape tuned configs. Numerics are unchanged:
+        # same fp8 weights, same 128x128 block scales.
+        if self.kernel_config.linear_backend == "auto":
+            self.kernel_config.linear_backend = "triton"
+            logger.info("Using the Triton blockwise-FP8 linear kernel.")
+
         # Compilation config overrides
         compilation_config = copy.deepcopy(self.compilation_config)
         if self.cudagraph_capture_sizes is not None: