Investigate: Qwen3.5-9B GRPO — how does Unsloth actually support it?

[akashgit]

Background

Unsloth advertises Qwen3.5 support, but when we tried `unsloth/Qwen3.5-9B`:

• It loads as `Qwen3_5ForConditionalGeneration` (a vision-language model)
• Crashes during GRPO generation with:
  ```
  RuntimeError: The size of tensor a (16) must match the size of tensor b (0)
  at non-singleton dimension 1
  ```
  in `compute_3d_position_ids` — a multimodal position encoding function

We switched to Qwen3-8B (pure text CausalLM) which works fine. But Qwen3.5's hybrid Gated DeltaNet architecture (Mamba+Transformer) is interesting and may have better efficiency.

Questions

1. Does Unsloth's Qwen3.5 support require specific model variants (e.g., text-only vs multimodal)?
2. Is there a `Qwen3.5-9B-Base` or similar text-only variant?
3. Does Qwen3.5 GRPO need special generation config (disable 3D position IDs for text-only)?
4. Check Unsloth Discord / GitHub issues for known workarounds

References

• Unsloth Qwen3.5 docs: https://unsloth.ai/docs/models/qwen3.5/fine-tune
• Qwen3.5 blog (Gated DeltaNet): https://qwen.ai/blog?id=qwen3.5
• Related issues: [Bug] Error when I run the Qwen3_(4B)-GRPO.ipynd, I do not know why the batch size will change during training, and I get an error. unslothai/unsloth#3003, #3864, #3149