diff --git a/src/prime_rl/trainer/models/qwen3_8_flash_next/modeling_qwen3_8_flash_next.py b/src/prime_rl/trainer/models/qwen3_8_flash_next/modeling_qwen3_8_flash_next.py
index 8b8ccf8196f..e8f4f20560d 100644
--- a/src/prime_rl/trainer/models/qwen3_8_flash_next/modeling_qwen3_8_flash_next.py
+++ b/src/prime_rl/trainer/models/qwen3_8_flash_next/modeling_qwen3_8_flash_next.py
@@ -194,7 +194,7 @@ def forward(
             seq_lens.to(inputs_embeds.device),
             total_tokens=None if seq_lens_are_pre_shard else inputs_embeds.shape[1],
         )
-        torch._dynamo.mark_dynamic(cu_seqlens, 0)
+        torch._dynamo.maybe_mark_dynamic(cu_seqlens, 0)
         position_embeddings = self.rotary_emb(inputs_embeds, position_ids)
 
         hidden_states = inputs_embeds.repeat(1, 1, self.config.hc_count)