试过了，不支持。感觉lmdeploy尤其是turbomind的模型适配已经停滞了，可惜了这么一个好推理引擎。

Thank you for your feedback. We understand your concerns, but we’d like to respectfully clarify a few points:

Turbomind’s model adaptation is not stagnant – we recently added support for models like GPT-OSS and InternVL 3.5, showing continued progress.
The pace is slower, but not due to lack of priority. Adapting new model architectures and operators at the C++ level—while meeting Turbomind’s strict performance standards—requires substantial engineering effort, which naturally takes time.
Qwen3 VL presents additional complexity due to its DeepStack architecture, making the integration of the video encoder with the backbone model particularly challenging.
We are currently focused on refactoring Turbomind to better support future models, and merge [POC] Encoder Disaggregation #4047 first to enable Turbomind to use PyTorch-based encoders. This will simplify integration, but it requires additional development time.

We genuinely welcome constructive community feedback and appreciate your understanding of the technical depth and cycle involved in this work.

[Feature] turbomind引擎是否支持qwen3 vl #4193

Description