-
-
Notifications
You must be signed in to change notification settings - Fork 4.9k
New issue
Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.
By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.
Already on GitHub? Sign in to your account
[Model] Add Llama-SwiftKV model #11023
base: main
Are you sure you want to change the base?
Conversation
👋 Hi! Thank you for contributing to the vLLM project. Once the PR is approved and ready to go, your PR reviewer(s) can run CI to test the changes comprehensively before merging. To run CI, PR reviewers can do one of these:
🚀 |
Appreciate your great work! If I understand correct, swiftKV is technology which could apply to other models(but needs re-train or finetune based on new model structure), not only limited on llama model, right? |
We will likely pick this up after refactoring of the memory management layer in V1 to support the varying KV cache requirements. In V1, we also have piece wise CUDA graph (no graph capture on attention) which should allow you no longer needing to self-manage CUDA Graph. |
SwiftKV was recently announced at https://www.snowflake.com/engineering-blog/swiftkv-llm-compute-reduction/. This PR adds a SwiftKV version of Llama that can immediately be used to run models at https://huggingface.co/collections/Snowflake/swiftkv-models-674f7d7474eb789e185d31cb.
The model definition is somewhat unconventional due to the need to early-exit some tokens but not others after a specific number of layers, and we wanted to minimize the amount of changes to vLLM's core code. Specifically:
Current limitations: