LMOps is a foundational research and technology project by Microsoft centered on large language models (LLMs) and generative AI models, with the goal of providing general-purpose capability support for building AI products. The repository address is https://github.com/microsoft/LMOps。
Purpose and directions: Research is conducted around prompting, long context, alignment, inference acceleration, domain customization, and the principles of in-context learning.
Main capabilities (as listed in the materials):
- Better prompting: Automatic Prompt Optimization, Promptist, Extensible prompts, Universal prompt retrieval, LLM Retriever, In-Context Demonstration Selection.
- Longer context: Structured prompting, Length-Extrapolatable Transformers.
- LLM alignment: Alignment via LLM feedback.
- LLM acceleration (faster inference): Lossless Acceleration of LLMs.
- LLM customization: Adapt LLM to domains.
- Foundational research: Understanding In-Context Learning.
Related projects and papers: Associated with microsoft/unilm, microsoft/torchscale. The paper release timeline includes In-Context Demonstration Selection with Cross Entropy Difference in November 2023, Tuna, Automatic Prompt Optimization with "Gradient Descent" and Beam Search, UPRISE in October 2023, Learning to Retrieve In-Context Examples in July 2023, Inference with Reference in April 2023, and several papers and model demonstrations in December 2022, among others.
Specific technical examples: Promptist performs automatic prompt optimization through reinforcement learning; Structured Prompting consumes long-sequence prompts in an efficient way; X-Prompt provides an extensible prompt interface beyond natural language; LLMA achieves lossless acceleration by copying and verifying reference text spans, and in retrieval-augmented generation andmulti-turn dialogue and other scenarios can achieve 2~3x acceleration without requiring an additional model.
Usage and limitations: The materials do not provide a unified installation command or usage steps; specific usage requires consulting each subproject and paper. The license is available in the LICENSE file in the root directory. Contact information and recruitment information are in the original text.
