[Bug]: Using Ray with compiled DAG throws the "The compiled graph can't have more than 10 in-flight executions" error

[tczekajlo]

Your current environment

The output of `python collect_env.py`

Your output of `python collect_env.py` here

🐛 Describe the bug

I'm trying to run DeepSeek R1 using vLLM with Ray and aDAG. As soon as I send the first request I get the following error:

File "/home/ray/anaconda3/lib/python3.12/site-packages/vllm/engine/async_llm_engine.py", line 825, in run_engine_loop
result = task.result()
^^^^^^^^^^^^^
File "/home/ray/anaconda3/lib/python3.12/site-packages/vllm/engine/async_llm_engine.py", line 748, in engine_step
request_outputs = await self.engine.step_async(virtual_engine)
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "/home/ray/anaconda3/lib/python3.12/site-packages/vllm/engine/async_llm_engine.py", line 353, in step_async
outputs = await self.model_executor.execute_model_async(
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "/home/ray/anaconda3/lib/python3.12/site-packages/vllm/executor/ray_distributed_executor.py", line 575, in execute_model_async
dag_future = await self.forward_dag.execute_async(serialized_data)
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "/home/ray/anaconda3/lib/python3.12/site-packages/ray/dag/compiled_dag_node.py", line 2186, in execute_async
self._raise_if_too_many_inflight_executions()
File "/home/ray/anaconda3/lib/python3.12/site-packages/ray/dag/compiled_dag_node.py", line 1918, in _raise_if_too_many_inflight_executions
raise ray.exceptions.RayCgraphCapacityExceeded(
ray.exceptions.RayCgraphCapacityExceeded: System error: The compiled graph can't have more than 10 in-flight executions, and you currently have 10 in-flight executions. Retrieve an output using ray.get before submitting more requests or increase `_max_inflight_executions`. `dag.experimental_compile(_max_inflight_executions=...)`

The Ray version that I'm using is 2.41.0.

Before submitting a new issue...

• [x] Make sure you already searched for relevant issues, and asked the chatbot living at the bottom right corner of the documentation page, which can answer lots of frequently asked questions.

[darthhexx]

I'm running into the same error with 2.42.0.

[ruisearch42]

Hi, this is a known issue in ray 2.41 which we are fixing. Will fix when new ray version is released.

Please downgrade to ray 2.40 for a short-term workaround.

[ruisearch42]

This has been fixed in https://github.com/vllm-project/vllm/pull/13994

[oelhammouchi]

Hi, I've been experiencing the exact same issue with Ray 2.46, any idea why this could be?

[ruisearch42]

hmm, the issue was fixed in ray 2.43, and I have not heard about it occurring again. Could you provide a repro script and the exact stack trace? @oelhammouchi

[VinayaLee]

Hi, I've been experiencing the exact same issue with Ray 2.46, any idea why this could be?

Hi do you fix this? @oelhammouchi
I have same problem with vllm 0.10.1.1 + Ray 2.49/2.51 , running pp = 16