diff --git a/python/sglang/srt/layers/attention/fla/fused_recurrent.py b/python/sglang/srt/layers/attention/fla/fused_recurrent.py
index 058662280..8f4130c10 100644
--- a/python/sglang/srt/layers/attention/fla/fused_recurrent.py
+++ b/python/sglang/srt/layers/attention/fla/fused_recurrent.py
@@ -136,7 +136,7 @@ def fused_recurrent_gated_delta_rule_fwd(
     B, T, H, K, V = *k.shape, v.shape[-1]
     HV = v.shape[2]
     N = B if cu_seqlens is None else len(cu_seqlens) - 1
-    BK, BV = triton.next_power_of_2(K), min(triton.next_power_of_2(V), 8)
+    BK, BV = triton.next_power_of_2(K), min(triton.next_power_of_2(V), 32)
     NK, NV = triton.cdiv(K, BK), triton.cdiv(V, BV)
     assert NK == 1, "NK > 1 is not supported yet"
     num_stages = 3