diff --git a/benchmarks/benchmark_throughput_autotune.py b/benchmarks/benchmark_throughput_autotune.py
index 407393f..4bbd843 100644
--- a/benchmarks/benchmark_throughput_autotune.py
+++ b/benchmarks/benchmark_throughput_autotune.py
@@ -1067,34 +1067,75 @@ def main() -> None:
                 oot_enabled = False
                 oot_whitelist = None
 
-            val = run_benchmark_multi(
-                label=label,
-                throughput_args=throughput_args,
-                num_runs=args.num_runs,
-                run_dir=run_dir,
-                log_counter=log_counter,
-                use_flaggems=use_flaggems,
-                gems_whitelist=gems_whitelist,
-                oot_whitelist=oot_whitelist,
-                oot_enabled=oot_enabled,
-                config_dir=round_config_dir,
-                config_payload=op_config_payload,
-            )
+            if backend == "reference":
+                best_ref_type = "FLASH_ATTN"
+
+                for attn_mode in ["FLASH_ATTN", "FLASHINFER"]:
+                    logger.info(f"  [Autotune Reference] Trying {attn_mode}...")
+
+                    test_env = {"VLLM_ATTENTION_BACKEND": attn_mode}
+                    os.environ["VLLM_ATTENTION_BACKEND"] = attn_mode
+
+                    val = run_benchmark_multi(
+                        label=f"op_{op}_reference_{attn_mode.lower()}",
+                        throughput_args=throughput_args,
+                        num_runs=args.num_runs,
+                        run_dir=run_dir,
+                        log_counter=log_counter,
+                        use_flaggems=use_flaggems,
+                        gems_whitelist=gems_whitelist,
+                        oot_whitelist=oot_whitelist,
+                        oot_enabled=oot_enabled,
+                        config_dir=round_config_dir,
+                        config_payload=op_config_payload,
+                    )
 
-            per_op_backend_results.setdefault(op, {})[backend] = val
-            write_results_csv(
-                csv_path,
-                baseline_result,
-                per_op_results,
-                baseline_fake_result=baseline_fake_result,
-                baseline_enable_result=baseline_enable_result,
-                op_backends=tuned_op_backends,
-                per_op_backend_results=per_op_backend_results,
-            )
-            if val is not None and baseline_total is not None:
-                if best_result is None or val[0] > best_result[0]:
-                    best_result = val
-                    best_backend = backend
+                    per_op_backend_results.setdefault(op, {})[backend] = val
+                    write_results_csv(
+                        csv_path,
+                        baseline_result,
+                        per_op_results,
+                        baseline_fake_result=baseline_fake_result,
+                        baseline_enable_result=baseline_enable_result,
+                        op_backends=tuned_op_backends,
+                        per_op_backend_results=per_op_backend_results,
+                    )
+                    if val is not None and baseline_total is not None:
+                        if best_result is None or val[0] > best_result[0]:
+                            best_result = val
+                            best_backend = backend
+                            best_ref_type = attn_mode
+
+                    os.environ["VLLM_ATTENTION_BACKEND"] = best_ref_type
+            else:
+                val = run_benchmark_multi(
+                    label=label,
+                    throughput_args=throughput_args,
+                    num_runs=args.num_runs,
+                    run_dir=run_dir,
+                    log_counter=log_counter,
+                    use_flaggems=use_flaggems,
+                    gems_whitelist=gems_whitelist,
+                    oot_whitelist=oot_whitelist,
+                    oot_enabled=oot_enabled,
+                    config_dir=round_config_dir,
+                    config_payload=op_config_payload,
+                )
+
+                per_op_backend_results.setdefault(op, {})[backend] = val
+                write_results_csv(
+                    csv_path,
+                    baseline_result,
+                    per_op_results,
+                    baseline_fake_result=baseline_fake_result,
+                    baseline_enable_result=baseline_enable_result,
+                    op_backends=tuned_op_backends,
+                    per_op_backend_results=per_op_backend_results,
+                )
+                if val is not None and baseline_total is not None:
+                    if best_result is None or val[0] > best_result[0]:
+                        best_result = val
+                        best_backend = backend
 
         per_op_results[op] = best_result
         if (
diff --git a/vllm_fl/dispatch/backends/reference/reference.py b/vllm_fl/dispatch/backends/reference/reference.py
index 78514d0..63520f5 100644
--- a/vllm_fl/dispatch/backends/reference/reference.py
+++ b/vllm_fl/dispatch/backends/reference/reference.py
@@ -134,7 +134,14 @@ def attention_backend(self, use_mla: bool = False) -> str:
         # Return vLLM's native flash attention backend as reference
         from vllm.attention.backends.registry import AttentionBackendEnum
 
+        import os
+
+        env_backend = os.environ.get("VLLM_ATTENTION_BACKEND", "FLASH_ATTN").upper()
+
         if use_mla:
             # vLLM native MLA backend
             return AttentionBackendEnum.MLA.get_path()
-        return AttentionBackendEnum.FLASH_ATTN.get_path()
+        if env_backend == "FLASHINFER":
+            return AttentionBackendEnum.FLASHINFER.get_path()
+        else:
+            return AttentionBackendEnum.FLASH_ATTN.get_path()