[Bug] Geo3k multi turn reward abnormal

[Zhou-jiecheng]

Bug Description

I'm trying to reproduce the reward curve of examples/geo3k_multi_turn, but I failed; the rewards do not increase. Even the first step raw reward is around 0.1, which is expected to be around 0.3.....

Steps to Reproduce

1. use the official Docker container;
2. run examples/geo3k_multi_turn training script.

Expected Behavior

The first step raw reward is around 0.3

Actual Behavior

The first step raw reward is around 0.1

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

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.

[ZHZisZZ]

#1727

This issue is probably caused by a bug in SGLang 0.5.9 that leaves the Qwen3-VL vision encoder unloaded.

[Zhou-jiecheng]

#1727

This issue is probably caused by a bug in SGLang 0.5.9 that leaves the Qwen3-VL vision encoder unloaded.

Let me try.

[Zhou-jiecheng]

It does not work for me, I switched to docker slime v0.2.2 and worked

[Alzheimer-hz]

Same problem, I pull the latest docker on Qwen3-VL-2B-Instrcuct. SGLang seems to have much lower performance on geo3k test
┌─────────────┬──────────┬──────┐
│ Server │ Accuracy │ Time │
├─────────────┼──────────┼──────┤
│ vLLM TP=8 │ 55.0% │ 686s │
├─────────────┼──────────┼──────┤
│ SGLang DP=8 │ 10.0% │ 158s │
└─────────────┴──────────┴──────┘

[ZHZisZZ]

@Alzheimer-hz

Have you had a chance to try this PR?
#1727

I suspect it might be related to the issue. I’d appreciate it if you could help verify.