Support for Qwen 3.5

[jonnor]

Hi, and thank you for this integration of vLLM into RunPod. The fast autoscaling of your platform is super impressive.

I wanted to run Qwen 3.5 with this project, as people say it is better than GPT-OSS for agentic coding.

I tried creating and endpoint using worker-vllm and model unsloth/Qwen3.5-35B-A3B-GGUF.
This failed with an error on vLLM startup like:

{"requestId": null, "message": "Worker startup failed: 1 validation error for ModelConfig\n  Value error, Model architectures ['Qwen3_5MoeForConditionalGeneration'] are not supported for now.

This seems to be because v 2.14.0 of this project looks to be shipping vLLM 0.16.0,
Whereas full Qwen 3.5 support is available in vLMM release v0.17.0.

[jonnor]

Upgrading to vLLM 0.17.1 seems to work. Ref https://github.com/jonnor/worker-vllm/tree/update-vllm
Also needed to restrict transformers to 4.x - ref #258

Have tested successfully on RunPod now with Qwen/Qwen3.5-35B-A3B on RTX PRO 6000, using MAX_MODEL_LEN=120000
If anyone else wants to try, you can use the Docker image jonnor/worker-vllm:vllm-0.17.1-2.
I copied the default environment variables from the standard vLLM worker template.
NB: You also need to bump Docker disk size to 150 GB

I have also pushed jonnor/worker-vllm:vllm-0.18.0-2, but that has not been tested.

[jonnor]

Note that to use with a coding agent one should use the following, as documented in the Qwen 3.5 usage guide in vLLM.

REASONING_PARSER=qwen3
TOOL_CALL_PARSER=qwen3_coder
ENABLE_AUTO_TOOL_CHOICE=true