How to quantize both vision encoder and llm together?

[XULU42]

I am trying to quantize dots_ocr model, whose submodules are ['model','vision_tower','lm_head'], as shown below:

By tracing the code down, I found that this code is for parsing modules for quantization, but it only return one module. By experiment, I can quantize either vision_tower or model separately and loading by vllm， but I can't quantize both due to above mechanism. Please help me figure out how to solve this, thanks^_^

[LiMa-cas]

I have quantized thinker or talker seperately，but I can not put them together，same questions

[XULU42]

I was able to quantize dots_ocr by changing code here and below to a for loop. It works, but looks ugly, looking forward to official mechanism to support both model quantization.

[Qubitium]

I was able to quantize dots_ocr by changing code here and below to a for loop. It works, but looks ugly, looking forward to official mechanism to support both model quantization.

@XULU42 Can you post your code diff? I would be awesome if you can create a PR to add this support. Don't worry about breaking things,once you create a PR and an associated unit test to tests/models (just copy another model's unit teset and modify) I can help with the PR testing and review.

I am already overworked so don't have the time to implement everything that everyone wants without external help.

[XULU42]

I would issue an PR later, but it seems to be a root change in module loop mechanism, may cause problem for other supported models. Thanks for your dedication to community. Your health is precious, please take care.

[Qubitium]

but it seems to be a root change in module loop mechanism, may cause problem for other supported models.

Don't worry about breakage. Module looper needs needs to eventually handle parallel models so I expect structual changes for world models where text, vision, audio module exists concurrently in the modules tree. Right now, the module tree auto extraction is expecting a linear single branch from top to bottom but current VL models are now structured where essentially there are two+ parallel branches in the module tree.

[XULU42]

but it seems to be a root change in module loop mechanism, may cause problem for other supported models.

Don't worry about breakage. Module looper needs needs to eventually handle parallel models so I expect structual changes for world models where text, vision, audio module exists concurrently in the modules tree. Right now, the module tree auto extraction is expecting a linear single branch from top to bottom but current VL models are now structured where essentially there are two+ parallel branches in the module tree.

The situation is clear. Let's try to add mechanism to parse and quantize parallel branches.

[LWQuestc]

I was able to quantize dots_ocr by changing code here and below to a for loop. It works, but looks ugly, looking forward to official mechanism to support both model quantization.
@XULU42 I've been working on something similar lately. If you don't mind, could I take a look at your code for reference? Thanks!

[namgyu-youn]

Quantizing the vision encoder is uncommon because: (1) calibration requires multimodal (image-text) samples, whereas most pipelines are built around text-only inputs, (2) vision encoders represent a small fraction of total parameters (e.g., ~1.5% / 0.42B out of 27.44B in Gemma3-27B).