diff --git a/requirements.txt b/requirements.txt
index 209ad00..77598f2 100644
--- a/requirements.txt
+++ b/requirements.txt
@@ -1,7 +1,7 @@
 accelerate>=1.9.0
 torch==2.6.0
 torchaudio==2.6.0
-transformers==4.49.0
+transformers>=4.50.0,<5.0.0
 fastapi>=0.116.1
 librosa>=0.11.0
 pydantic>=2.11.7
diff --git a/src/mimo_audio/modeling_mimo_audio.py b/src/mimo_audio/modeling_mimo_audio.py
index f9681ad..95246a0 100644
--- a/src/mimo_audio/modeling_mimo_audio.py
+++ b/src/mimo_audio/modeling_mimo_audio.py
@@ -13,6 +13,7 @@
 from transformers.generation.utils import (
     GenerateOutput,
     GenerationConfig,
+    GenerationMixin,
     StoppingCriteriaList,
     is_deepspeed_zero3_enabled,
 )
@@ -221,7 +222,12 @@ def to_dict(self):
         }
 
 
-class MiMoAudioForCausalLM(Qwen2PreTrainedModel):
+class MiMoAudioForCausalLM(Qwen2PreTrainedModel, GenerationMixin):
+    # transformers>=4.50 移除了 PreTrainedModel._supports_cache_class，
+    # 但 prepare_inputs_for_generation 中仍引用它来决定是否传递 cache_position。
+    # MiMo 使用 DynamicCache，显式声明为 True。
+    _supports_cache_class = True
+
     def __init__(
         self,
         config: MiMoAudioConfig | Qwen2Config,
@@ -746,9 +752,10 @@ def slm_sample(
             this_peer_finished,
             synced_gpus,
             device=input_ids.device,
-            cur_len=cur_len,
-            max_length=max_length,
         ):
+            # 长度检查（transformers>=4.50 不再由 _has_unfinished_sequences 负责）
+            if cur_len >= max_length:
+                break
             # prepare model inputs
             model_inputs = self.prepare_inputs_for_generation(input_ids, **model_kwargs)