[CI] Fix test_inference_engines_generation after vllm 0.16.0 upgrade; Use the correct GSM8k path for test_generator_multi_turn_gsm8k_router_replay

[SumanthRH]

What does this PR do?

Fixes CI failures on main right now: https://github.com/NovaSky-AI/SkyRL/actions/runs/23218100038/job/67484003027

1. tests/backends/skyrl_train/gpu/gpu_ci/test_skyrl_gym_generator.py::test_generator_multi_turn_gsm8k_router_replay - FileNotFoundError: Unable to find '/mnt/cluster_storage/data/gsm8k/validation.parquet' -> R3 PR: Rollout Routing Replay #1273 added a router replay test but used an incorrect path for the GSM8K dataset on CI
2. tests/backends/skyrl_train/gpu/gpu_ci/inference_servers/test_weight_sync.py::TestColocatedIpcWeightUpdateFlow::test_update_weights_ipc -> Similar fix as in [CI] Skip FlashRL integration test in CI and fix failing generation test for new inference codepath #1301

=========================== short test summary info ============================
FAILED tests/backends/skyrl_train/gpu/gpu_ci/test_engine_generation.py::test_inference_engines_generation[tp2_pp1_dp2] - ray.exceptions.ActorDiedError: The actor died because of an error raised in its creation task, ray::AsyncVLLMInferenceEngine.__init__() (pid=25165, ip=10.0.25.170, actor_id=7617504c4a769d500d1bc9ef13000000, repr=<skyrl.backends.skyrl_train.inference_engines.vllm.vllm_engine.AsyncVLLMInferenceEngine object at 0x7ea2b748c800>)
  File "/home/ray/anaconda3/lib/python3.12/concurrent/futures/_base.py", line 456, in result
    return self.__get_result()
           ^^^^^^^^^^^^^^^^^^^
  File "/home/ray/anaconda3/lib/python3.12/concurrent/futures/_base.py", line 401, in __get_result
    raise self._exception
           ^^^^^^^^^^^^^^^^^^^^^
           ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
           ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "/tmp/ray/session_2026-03-17_21-54-20_786615_3475/runtime_resources/working_dir_files/s3_anyscale-production-data-cld-hxkifz7xa22mwicp21nzkds1lw_org_xc6lv84h3d7m9dljcc17esfw2i_cld_hxkifz7xa22mwicp21nzkds1lw_runtime_env_packages_pkg_95e2ed2914e100bfad4cccac453e4b5b/skyrl/backends/skyrl_train/inference_engines/vllm/vllm_engine.py", line 343, in __init__
    super().__init__(*args, **kwargs)
  File "/tmp/ray/session_2026-03-17_21-54-20_786615_3475/runtime_resources/working_dir_files/s3_anyscale-production-data-cld-hxkifz7xa22mwicp21nzkds1lw_org_xc6lv84h3d7m9dljcc17esfw2i_cld_hxkifz7xa22mwicp21nzkds1lw_runtime_env_packages_pkg_95e2ed2914e100bfad4cccac453e4b5b/skyrl/backends/skyrl_train/inference_engines/vllm/vllm_engine.py", line 112, in __init__
    self.llm = self._create_engine(*args, **kwargs)
               ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
           ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "/tmp/ray/session_2026-03-17_21-54-20_786615_3475/runtime_resources/working_dir_files/s3_anyscale-production-data-cld-hxkifz7xa22mwicp21nzkds1lw_org_xc6lv84h3d7m9dljcc17esfw2i_cld_hxkifz7xa22mwicp21nzkds1lw_runtime_env_packages_pkg_95e2ed2914e100bfad4cccac453e4b5b/skyrl/backends/skyrl_train/inference_engines/vllm/vllm_engine.py", line 364, in _create_engine
    engine = vllm.AsyncLLMEngine.from_engine_args(engine_args, stat_loggers=stat_loggers)
             ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "/home/ray/.cache/uv/builds-v0/.tmpuPz58u/lib/python3.12/site-packages/vllm/v1/engine/async_llm.py", line 251, in from_engine_args
    return cls(
           ^^^^
  File "/home/ray/.cache/uv/builds-v0/.tmpuPz58u/lib/python3.12/site-packages/vllm/v1/engine/async_llm.py", line 148, in __init__
    self.engine_core = EngineCoreClient.make_async_mp_client(
                       ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "/home/ray/.cache/uv/builds-v0/.tmpuPz58u/lib/python3.12/site-packages/vllm/v1/engine/core_client.py", line 121, in make_async_mp_client
    return DPAsyncMPClient(*client_args)
           ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "/home/ray/.cache/uv/builds-v0/.tmpuPz58u/lib/python3.12/site-packages/vllm/v1/engine/core_client.py", line 1082, in __init__
    self._ensure_stats_update_task()
  File "/home/ray/.cache/uv/builds-v0/.tmpuPz58u/lib/python3.12/site-packages/vllm/v1/engine/core_client.py", line 1091, in _ensure_stats_update_task
    assert self.stats_update_address is not None
           ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
AssertionError
FAILED tests/backends/skyrl_train/gpu/gpu_ci/test_skyrl_gym_generator.py::test_generator_multi_turn_gsm8k_router_replay - FileNotFoundError: Unable to find '/mnt/cluster_storage/data/gsm8k/validation.parquet'
ERROR tests/backends/skyrl_train/gpu/gpu_ci/inference_servers/test_weight_sync.py::TestColocatedIpcWeightUpdateFlow::test_update_weights_ipc - AssertionError

---

[gemini-code-assist]

Code Review

This pull request effectively addresses two CI failures. The first is a FileNotFoundError which is resolved by correcting a dataset path. The second failure, an AssertionError following a vllm upgrade, is fixed by parameterizing a test to include a MoE model, which likely covers the failing code path. The changes are well-targeted and correct. I have one minor suggestion to improve code clarity by removing a redundant parameter.

[gemini-code-assist]

This line explicitly sets data_path to the same value as its default in the run_generator_end_to_end function signature. To reduce redundancy and improve maintainability, you can remove this line and rely on the default value.

[devin-ai-integration]

✅ Devin Review: No Issues Found

Devin Review analyzed this PR and found no potential bugs to report.

View in Devin Review to see 2 additional findings.