Proto commits in ServerlessLLM/ServerlessLLM

These 5 commits are when the Protocol Buffers files have changed:

2024-11-29

[FIX] Fix loading loop in `SllmLocalStore` (#165) ## Description This PR enhances the `SllmLocalStore` by addressing issues in the loading loop and adapting to the latest automatic hardware configuration detection feature. ## Motivation - Fixes the first issue mentioned in #131. - Prepares the codebase for the upcoming live-migration feature release. ## Type of Change - [x] Bug fix - [x] Enhancement ## Checklist - [x] I have read the [CONTRIBUTING](https://github.com/ServerlessLLM/ServerlessLLM/blob/main/CONTRIBUTING.md) guide. - [x] The changes are tested and verified to maintain compatibility with existing features.

The documentation is generated from this commit.

2024-10-21

Prepare for 0.5.0 release (#112) ## Description This PR includes necessary modifications for the 0.5.0 release: 1. Refactored the repository to shorten file paths (e.g., `serverless_llm/` to `sllm` and `serverless_llm/store/serverless_llm_store` to `sllm_store/sllm_store`). 2. Updated the README to clarify that `sllm` is pronounced "slim." 3. Added `pip install` instructions in the documentation. ## Motivation Previously, using the full name `serverless_llm` made it cumbersome to reference the codebase during discussions. We have now abbreviated `serverless_llm` to `sllm` (pronounced "slim"), as well as `serverless_llm_store` to `sllm_store` (pronounced "slim store"). However, "sllm" is unavailable on PyPI due to its similarity to an existing project `slim`. As a result, we will retain the full name `serverless_llm` for the `pip install` command but use `sllm` for all other references within the codebase and documentation. ## Type of Change - [x] Breaking change - [x] Documentation update ## Checklist - [x] I have read the [CONTRIBUTION](https://github.com/ServerlessLLM/ServerlessLLM/blob/main/CONTRIBUTING.md) guide. - [x] I have updated the tests (if applicable). - [x] I have updated the documentation (if applicable).

2024-09-19

Fy/unified model path (#82) ## Description This PR cleans up ServerlessLLM Store design by provide dedicated interface for PyTorch and Transformers. ServerlessLLM Store now uses `model_path` instead of `model_name` as the global ID for a model checkpoint. By providing a clean API, ServerlessLLM now supports storage-aware model loading scheduling for vLLM backend. ## Changes - Move `save_model` and `load_model` to `serverless_llm_store.transformers` - Move 'save_dict` and `load_dict` to `serverless_llm_store.torch` - Use `serverless_llm_store.torch.save_dict/load_dict` in vLLM - Support and require model registration for vLLM backend

2024-07-30

feat: storage aware scheduler (#22) This PR implements a storage manager and a storage-aware model loading scheduler.

2024-07-10

Initial commit: Launching ServerlessLLM This release includes basic modules of ServerlessLLM. This project builds upon our research published in OSDI 2024. As we transitioned from an internal development environment to this open-source platform, the detailed commit history and authorship information could not be preserved. Hence, we are taking this opportunity to formally acknowledge all contributors. - [Yao Fu](https://github.com/future-xy) - [Leyang Xue](https://github.com/drunkcoding) - [Yeqi Huang](https://github.com/Chivier) - [Andrei-Octavian Brabete](https://github.com/andrei3131) - [Matej Sandor](https://github.com/MatejSandor) - [Ruiqi Lai](https://github.com/lrq619) - [Siyang Shao](https://github.com/SiyangShao) - [Xinyuan Tong](https://github.com/JustinTong0323) Thank you to everyone involved for your dedication and expertise.