[BUG] `AsyncEngineArgs` unknown arg `disable_log_requests` with freshest vllm

[vadimkantorov]

 File "/usr/lib/python3.12/concurrent/futures/_base.py", line 449, in result
 return self.__get_result()
 ^^^^^^^^^^^^^^^^^^^
 File "/usr/lib/python3.12/concurrent/futures/_base.py", line 401, in __get_result
 raise self._exception
 ^^^^^^^^^^^^^^^^^^^^^
 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
 File "/tmp/ray/session_2025-09-04_19-02-52_192985_111429/runtime_resources/working_dir_files/_ray_pkg_efc8b708aebb3d24/trinity/common/models/vllm_model.py", line 64, in __init__
 engine_args = vllm.AsyncEngineArgs(
 ^^^^^^^^^^^^^^^^^^^^^
TypeError: AsyncEngineArgs.__init__() got an unexpected keyword argument 'disable_log_requests'`

Fixing this is needed for upgrading to vllm nightly which is built against pytorch 2.8.0 which fixes inductor cache problem

[hiyuchang]

For now, Trinity-RFT supports vllm>=0.9.1,<=0.10.0.

We will upgrade the versions of VLLM recently and pay attention to this issue. Thank you!

[vadimkantorov]

We had to enable CUDA graphs / torch.compile in rollouts. And pytorch 2.7.0 has the bug of races in torch.compile cache - which forces us to use VLLM_COMPILE_CACHE_DISABLE=1 causing slow-downs:

• [Bug]: Strange error AssertionError: failed to get the hash of the compiled graph when running Qwen/Qwen3-8B via LLM class vllm-project/vllm#18851

This seems to have been fixed in PyTorch 2.8.0, to benefit from this we need to try to use vllm nightly (and hope for a new vllm service release built against 2.8.0)

[vadimkantorov]

It appears that instead of , disable_log_requests=arg , enable_log_requests=(not arg) need be used (maybe simply upgrade all the config classes / config yamls)

Here is an excerpt from vllm:

    @property
    @deprecated(
        "`disable_log_requests` is deprecated and has been replaced with "
        "`enable_log_requests`. This will be removed in v0.12.0. Please use "
        "`enable_log_requests` instead.")
    def disable_log_requests(self) -> bool:
        return not self.enable_log_requests

[vadimkantorov]

Vllm 0.10.2 (depending on pytorch 2.8.0) was released yesterday. Might be nice to migrate to this release to benefit from 2.8.0...

[pan-x-c]

Fixed in #278