Effective LLMS devices: Service and Light Communication Strategies

Links table

Abstract and 1. Introduction

2 relevant business

3 methodology and 3.1 causal language model as a classification model

3.2 functional code

3.3 Data set

3.4 Form Development and Training

4 experiments and 4.1 Android job calls

4.2 Extension of vehicle, screams, and Doordash

4.3 Full and partial training data sets and 4.4 full training and training Lora

4.5 Call the parallel and overlapping job and 4.6 likely to lose special symbols

5 discussion, business and future references

Excessive

A.1 Examples of Android job

A.2 Examples of car function

Publishing language forms on the device Due to the lower memory restrictions and speeds of reasoning, spreading larger models on edge devices such as computers or smartphones is a challenge. However, the efforts made to publish large LLMS language models on the edge devices are ongoing. Open source models for controlled sizes, such as GEMMA-2B, GEMMA-7B and Stablecode-3B [31]And llama-7B [47]It was presented. To enhance this reasoning speed on devices, search initiatives such as Llama CPP [24] It was developed. MLC LLM framework [46] It allows the operation of 7B language models on mobile phones and other edge devices, which indicates compatibility via different devices, including AMD, NVIDIA, Apple and Intel GPU.

The function of the invitation in the language forms Fast developments have been observed in the potential to summon jobs for smaller models. Projects like Nexusraven [42]Toolformer [37]TOLLALPACA [44]Gorilla [30]Toollama [32] And taskmatrix [20] It has proven that the 7B and 13B models can call the external applications programming facades similar to the GPT-4. The leading 2B Octopus V1 project enabled the GPT-4 to perform with GPT-4. This working group uses a method based on recalling jobs, as the relevant functional form recovers from a large gathering based on the user’s inquiry, then it creates a response using these jobs as a context.

Refinement and transformers of language models Microscopic language models have become a prevailing practice, with various efforts allocated to this endeavor. Laura [14] The method of selecting training forms is often under GPU restrictions. We use both full model training and Lora training in our work, and compare its performance. The remarkable benefit of Lora is to facilitate the extended functions in the models, indicating the possibility of adapting our current framework to a wide range of applications.