diff --git a/python/sglang/srt/model_executor/runner/eager_runner.py b/python/sglang/srt/model_executor/runner/eager_runner.py
index 8a61fe40f..c4b24e8dc 100644
--- a/python/sglang/srt/model_executor/runner/eager_runner.py
+++ b/python/sglang/srt/model_executor/runner/eager_runner.py
@@ -100,11 +100,7 @@ class EagerRunner(BaseRunner):
 
             max_bs = ceil_align(max_bs, self.attn_tp_size)
             max_bs = ceil_align(max_bs, get_cp_padding_align_size())
-        prefill_ceiling = (
-            sa.max_prefill_buffer_tokens()
-            if sa.chunked_prefill_size and sa.chunked_prefill_size > 0
-            else mr.max_total_num_tokens
-        )
+        prefill_ceiling = max(mr.max_total_num_tokens, sa.max_prefill_buffer_tokens())
         max_num_token = max(prefill_ceiling, max_bs * num_tokens_per_bs)
         if require_mlp_sync(sa):
             max_num_token = ceil_align(max_num_token, self.attn_tp_size)