[Bug] ValueError: rollout_routed_experts is required in rollout_data when use_rollout_routing_replay is set.

[fxxkai]

Bug Description

I want to use R3, and set the following cmds:

--use-rollout-routing-replay
--sglang-enable-return-routed-experts

but it doesn't work, the error is:

File "/root/work/externalstorage/gpfsprd/sunjiahui/slime/slime/backends/megatron_utils/actor.py", line 413, in train_actor
self.fill_routing_replay(data_iterator, num_microbatches, rollout_data)
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "/root/work/externalstorage/gpfsprd/sunjiahui/slime/slime/backends/megatron_utils/actor.py", line 269, in fill_routing_replay
raise ValueError(
ValueError: rollout_routed_experts is required in rollout_data when use_rollout_routing_replay is set.

Steps to Reproduce

bash run-qwen3-30B-A3B.sh

Expected Behavior

the generated output has rollout_routed_experts

Actual Behavior

the generated output doesn't has rollout_routed_experts.

Environment

• slime version:
• Python version:
• PyTorch version:
• CUDA/ROCm version:
• GPU type and count:
• OS:
• SGLang version (if relevant):
• Megatron-LM version (if relevant):

Logs

I print the keys in output, meta_info. They doesn't return rollout_routed_experts.

(RolloutManager pid=55507) [2026-05-14 07:23:33] sglang_rollout.py:224 - payload keys: dict_keys(['sampling_params', 'return_logprob', 'return_routed_experts', 'input_ids'])
(RolloutManager pid=55507) [2026-05-14 07:23:33] sglang_rollout.py:225 - output keys: dict_keys(['text', 'output_ids', 'meta_info'])
(RolloutManager pid=55507) [2026-05-14 07:23:33] sglang_rollout.py:226 - meta_info keys: dict_keys(['id', 'finish_reason', 'prompt_tokens', 'weight_version', 'total_retractions', 'queue_time', 'prefill_launch_delay', 'prefill_launch_latency', 'prefill_finished_ts', 'input_token_logprobs', 'output_token_logprobs', 'completion_tokens', 'cached_tokens', 'cached_tokens_details', 'e2e_latency', 'request_received_ts', 'request_sent_to_scheduler_ts', 'decode_finished_ts', 'inference_time', 'decode_throughput', 'response_sent_to_client_ts'])
(RolloutManager pid=55507) [2026-05-14 07:23:33] sglang_rollout.py:227 - has routed_experts: False

Additional Context

No response

Pre-submission Checklist

• I have read the CONTRIBUTING.md and understand the collaboration scope.
• I have read the documentation and my issue is not addressed there.
• I have searched for existing issues and this is not a duplicate.
• I have provided a minimal, reproducible example.

[zhuzilin]

Could you check your sglang router version and if the "return_routed_experts" has passed to sglang within the payload of the generate request (you may need to manually add print to the sglang for this)? This issue is often caused by sglang_router ignore the payload.

[fxxkai]

Could you check your sglang router version and if the "return_routed_experts" has passed to sglang within the payload of the generate request (you may need to manually add print to the sglang for this)? This issue is often caused by sglang_router ignore the payload.

I checked sglang-router, and it was indeed caused by an incorrect version. After switching to the image v0.2.4, it worked fine.
Thanks for your relpy!