Cuda error in RULER notebook

I am using the RULER notebook to train a model, but I get this error:

---------------------------------------------------------------------------
RuntimeError                              Traceback (most recent call last)
/usr/local/lib/python3.12/dist-packages/unsloth_zoo/vllm_utils.py in load_vllm(model_name, config, gpu_memory_utilization, max_seq_length, dtype, training, float8_kv_cache, random_state, enable_lora, max_lora_rank, max_loras, use_async, use_engine, disable_log_stats, enforce_eager, enable_prefix_caching, compilation_config, conservativeness, max_logprobs, use_bitsandbytes, unsloth_vllm_standby, return_args)
   1499             if use_async:
-> 1500                 llm = AsyncLLMEngine.from_engine_args(AsyncEngineArgs(**engine_args))
   1501             elif use_engine:

31 frames
RuntimeError: torch.cuda.MemPool doesn't currently support expandable_segments.

During handling of the above exception, another exception occurred:

RuntimeError                              Traceback (most recent call last)
/usr/local/lib/python3.12/dist-packages/unsloth_zoo/vllm_utils.py in load_vllm(model_name, config, gpu_memory_utilization, max_seq_length, dtype, training, float8_kv_cache, random_state, enable_lora, max_lora_rank, max_loras, use_async, use_engine, disable_log_stats, enforce_eager, enable_prefix_caching, compilation_config, conservativeness, max_logprobs, use_bitsandbytes, unsloth_vllm_standby, return_args)
   1525                 )
   1526             else:
-> 1527                 raise RuntimeError(error)
   1528         pass
   1529     pass

RuntimeError: torch.cuda.MemPool doesn't currently support expandable_segments.

I have tried upgrading transformers, vllm, and ART, and I have also tried multiple models, including GPT-OSS 20b and Qwen/Qwen2.5-7B-Instruct, but nothing resolved this issue. Here is my notebook's code: https://colab.research.google.com/drive/13Ax7eQ313WxTHXzosUciHdBYXlnG9047?usp=sharing

Open the linked Google Colab notebook and first run the RULER training flow that reaches vLLM, recording the environment and the reported torch.cuda.MemPool error. Done means the notebook can load a selected model and start training without this runtime error, with the result documented in the issue.