use 1e-3 atol

Author: leonling-ll

benchmarks/third_party/sglang/extended_attention_benchmark.py

@@ -105,7 +105,7 @@ def refer_fn():
                                     b_seq_len_prefix, max_len_in_batch)
                 return o_redundant
 
-            benchmark_suit.assert_close(triton_fn, refer_fn, atol=1e-4, rtol=1e-2, err_msg='extend to refer')
+            benchmark_suit.assert_close(triton_fn, refer_fn, atol=1e-3, rtol=1e-2, err_msg='extend to refer')
 
         _, min_ms, max_ms, mean, cv = benchmark_suit.do_bench(triton_fn, n_warmup=10, n_repeat=10, quantiles=quantiles)