Support for VLMs on Roadmap

[Jeremalloch]

Hi there,

I wanted to inquire about the roadmap for VLM training.

I see that VLM support is still under the experimental section of torchtitan. I did not see any mention of VLMs on the PyTorch distributed H1 2026 roadmap. What does the roadmap look like for VLMs?

I'm looking to adopt a new training framework for multimodal pytorch training and I want to ensure that if I select torchtitan VLM support will see continued investment.

[tianyu-l]

We are working on supporting Qwen3-VL.

cc @shuhuayu

[shuhuayu]

@Jeremalloch Thanks for the ask. We are working on the enablement of Qwen3-VL right now, and will release it in recent weeks.

[Jeremalloch]

Awesome, thank you both!

[xiao-hu]

Hi @shuhuayu @tianyu-l ,
Just notice this thread. I want to share that I have implemented a prototype for Qwen3-VL in my fork. It currently lives in experiments/qwen3_vl to avoid touching core torchtitan code.

I extended Qwen3 text implementation with DeepStack visual feature injection. The vision encoder is currently just a thin wrapper over the HF Qwen3VLVisionModel. Next I plan to reimplement it to enable more better control on vision model compilation.

I’d love to contribute this as a PR to the experiments/ folder. Are there specific design patterns you'd prefer I align with before submitting?

[shuhuayu]

@xiao-hu Thanks for the reference! We currently have one implementation under testing and will push it soon, and would appreciate your input given your expertise on it.