Add ColQwen3.5 and BiQwen3.5 model support#400
Conversation
…tialization and evaluation parameters
|
@ManuelFay feedback addressed! |
|
I'm super sorry, I had reviewed but didn't press submit :/ main thing is the 320 dim comment which I feel should probably not be the hqrdcoded standard. rest LGTM |
|
great ! Are you sure your model load correctly with this new way of injecting the config ? If so, I 'll merge |
We're gonna find out soon enough: https://buildkite.com/vllm/ci/builds/56432/steps/canvas If you want, let's wait until it's merged on vLLM first. EDIT: Also updated to point to the latest checkpoint. |
|
That's awesome ! All right, let's wait and when we have full confidence, just indicate it here and I'll merge it in. Thanks again for the contribution ! |
@ManuelFay It's ready to go! |
|
Great, merging now! |
|
Congrats on PR #400 @athrael-soju , nice milestone ! |
Haha nice one! |
2 participants
This pull request adds support for the new ColQwen3.5 and BiQwen3.5 models, including their model and processor implementations, integration into the codebase, a training script, and comprehensive tests. These changes enable document retrieval using the latest Qwen3.5 backbone and provide both single-vector (BiQwen3.5) and multi-vector (ColQwen3.5) retrieval capabilities.
Model and Processor Additions:
ColQwen3_5andBiQwen3_5model classes, along with their respective processor classes (ColQwen3_5Processor,BiQwen3_5Processor), implementing document and vision-language retrieval based on the Qwen3.5 backbone. [1] [2] [3] [4] [5] [6] [7]__init__.pyfile for easy import and usage.Training and Configuration:
train_colqwen3_5_model.pyunderscripts/configs/qwen3_5/, allowing users to train ColQwen3.5 models with configurable arguments for optimizer, loss, PEFT, and datasets.Testing:
ColQwen3_5model, including model loading, forward pass with images and queries, and integration with retrieval tasks.Documentation:
CHANGELOG.mdto document the addition of ColQwen3.5 and BiQwen3.5 support, including a reference to the pretrained checkpoint.A pretrained checkpoint is available at athrael-soju/colqwen3.5-v1 (4.5B params, Apache 2.0)