Can it support qwen3-vl-235b-a22b-instruct?
finally saw the EAGLE3 on the qwen3-vl series!
i would like to ask if it is possible to replicate the work of '30b-a3b' on 'qwen3-vl-235b-a22b', what changes need to be made to the specForge source code?
For qwen3-vl-235b-a22b, you'll need tp size > 1 support. I added the qwen3_vl_moe.py file from transformers and you'll need to modify it (see https://docs.sglang.ai/SpecForge/advanced_features/customization.html#customize-target-model) and then wire up the plumbing in auto.py::AutoDistributedTargetModel::_model_mapping. We're currently in the process of verifying correctness of the draft model training (with these v0's) and adding support into sglang/vllm, afterwards we may add tp size > 1 support!
thanks, i will try it now~
by the way, are the dense models like 'qwen3-vl-8b' and the '30b-a3b' models of the MOE architecture adapted to eagle3 consistent?
yes both qwen3-vl-8b and qwen3-vl-30b-a3b use the same code since they both can fit on a single gpu, the configs were adapted from the target model configs.
hi guys, i run it work! and next problem is how to use it to deploy?
i use sglang 0.5.4 post1 can now work, are there any recommend deployment?