Qwen2.5-7B gets ~50% on GSM8K eval while Qwen2.5-3B gets ~80% using scripts/grpo_demo_llama3_qwen2.py

### Description

Hi — I’m evaluating GSM8K using:

`scripts/grpo_demo_llama3_qwen2.py`

This is **eval-only (no training), pass@1**, using the **vanilla inference engine**.

Results:

| Model | Accuracy | Note |
|------|----------|------|
| **Qwen2.5-3B-Instruct** | ~80% | consistent with Qwen’s official report |
| **Qwen2.5-7B-Instruct** | ~50% | much lower than expected |

To enable the 3B model I only added a few lines of configuration (no training logic changed).
The small patch is here for reference:

👉 **Link:** <[PASTE YOUR PR / GIST HERE](https://github.com/zyh1999/tunix/tree/only-test?tab=readme-ov-file)>

### Reproduction

python3 scripts/grpo_demo_llama3_qwen2.py --model-version=Qwen/Qwen2.5-7B-Instruct --rollout-engine=vanilla --global-batch-size=16 --num-test-batches=83 --max-tpu-to-use=4 --eval-only=True

python3 scripts/grpo_demo_llama3_qwen2.py --model-version=Qwen/Qwen2.5-3B-Instruct --rollout-engine=van
illa --global-batch-size=16 --num-test-batches=83 --max-tpu-to-use=2 --eval-only=True