[Feature]: loading model from remote KV store such as Redis

[DellCurry]

🚀 The feature, motivation and pitch

Currently, loading models in vLLM takes such a long time, up to several minutes. 3 steps are needed:

1. Downloading model files from S3 or HuggingFace or your own repo.
2. Reading, decoding and loading Tensor from disk file into CPU memory.
3. Copy Tensor into GPU memory.

I wonder whether it is a good idea to introduce a new class named RemoteModelLoader in model_loader.py. By doing so, we just need to store model tensors and metadata into remote database only once. After that, we could load models from remote database directly, and also faster than traditional way since local disk is not involved.

Besides Redis, I noticed that some companies and organization are working on RDMA-based KV database, which is much faster theoretically. And step 3 is not necessary if using GDR. Those databases in the future may also become available by RemoteModelLoader, similar as how Redis is used.

Alternatives

No response

Additional context

No response

Before submitting a new issue...

• Make sure you already searched for relevant issues, and asked the chatbot living at the bottom right corner of the documentation page, which can answer lots of frequently asked questions.

[github-actions]

This issue has been automatically marked as stale because it has not had any activity within 90 days. It will be automatically closed if no further activity occurs within 30 days. Leave a comment if you feel this issue should remain open. Thank you!

[github-actions]

This issue has been automatically closed due to inactivity. Please feel free to reopen if you feel it is still relevant. Thank you!