Skip to content

Adds support for Qwen#408

Merged
tazlin merged 15 commits intomainfrom
qwen
Nov 24, 2025
Merged

Adds support for Qwen#408
tazlin merged 15 commits intomainfrom
qwen

Conversation

@db0
Copy link
Member

@db0 db0 commented Sep 6, 2025

  • Supports models with multi-filetype downloads. For example Qwen requires a unet, a vae and a clip. The new code allows hordelib to download all 3. It will only keep the unet in VRAM however as the code was easy to port into our custom node loader. The vae and clip loader are much more advanced and was not possible to import into our own node loader to keep them in VRAM as well.
  • Support Qwen model. However it requires min 24G of VRAM, so only the very high-end workers will be able to run it consistently.
  • Switches to 71ed4a399ec76a75aa2870b772d2022e4b9a69a3 commit of comfyUI to support Qwen, which should hopefully also gives a lot of bugfixes and more features. This required some rework to support its new async definitions.

@db0 db0 marked this pull request as ready for review November 20, 2025 20:56
@db0
Copy link
Member Author

db0 commented Nov 20, 2025

All tests are now passing. Ofc the qwen one is very very heavy. @tazlin how do we make this optional?

@tazlin tazlin merged commit 205997d into main Nov 24, 2025
2 of 3 checks passed
@tazlin tazlin deleted the qwen branch November 24, 2025 21:17
tazlin added a commit that referenced this pull request Nov 24, 2025
* can now correctly download models with multiple file typoes

* more asyncio

* added await on the IsChangedCache

* fix: basic inference capability; comment out qwen while WIP

* fix: show correct frame info in `InterceptHandler`

Previous to this change, messages would show as originating from `logging:handle` due to not crawling past `comfy_horde` or `loguru` in the call stack.

* fix: use gc.collect/torch.empty cache post cleanup

This ensures memory/vram is completely released as intended

* lora tests

* wip: working with native loaders

* wip: working through our own loader

* fix: Workaround to work with latest version of comfyui

* ci: all tests passing with qwen changes

* fix: handle qwen from baseline

* feat: make determining file_type a method

---------

Co-authored-by: tazlin <tazlin.on.github@gmail.com>
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

2 participants