Train LLM on AMD APU. In this scenario, we’ll use an APU because most laptops with a Ryzen CPU include an iGPU; specifically, this post should work with iGPUs based on the “GCN 5.0” architecture, or “Vega” for friends. We’ll use an AMD Ryzen 2200G in this post, an entry-level processor equipped with 4C/4T and an integrated GPU.

UMA buffer size is the size of memory used by APU. It is set on the motherboard, often limited to 2GB. But LLM AI could use 16GB or more.

My Account for Motherboerd Asus PRIME X570-P, registered here.

GGML quantized models. They would let you leverage CPU and system RAM, instead of having to rely on a GPU’s. This could save you a fortune, especially if go for some used AMD Epyc platforms. This could be more viable for the larger models, especially the 30B/65B parameters models which would still press or exceed the VRAM on the P40.

Replit is a site where I can run any REPL online. Can be used for AI

High level how to train a model

The OSCAR project (Open Super-large Crawled Aggregated coRpus) is an Open Source project aiming to provide web-based multilingual resources and datasets for Machine Learning (ML) and Artificial Intelligence (AI) applications.

Dataset is just a zip of files

Research community developing various code models, small and big. Models may not be instruct

Another open source small (1B) model.

They have the 1.3B version!!! This may be the best to start with Newspeak. Should work train even on huggingcface

Chat uses model geared for coding.

Another possible model. For coding capabilities, Deepseek Coder achieves state-of-the-art performance among open-source code models on multiple programming languages and various benchmarks.

With the optimizers of bitsandbytes (like 8 bit AdamW), you would need 2 bytes per parameter, or 14 GB of GPU memory.

Another potential model to use for Newspeak, but it is NOT open source. Adventage: 2.5B params, so should be usable in small GPUs

training a model like Llama with 2.7 billion parameters outperformed a larger model like Vicuna with 13 billion parameters. Especially when considering resource consumption, this might be a good alternative to using a 7B Foundation model instead of a full-blown ChatGPT. The best price-to-performance base model for our use case turned out to be Mistral 7b. The model is compact enough to fit into an affordable GPU with 24GB VRAM and outperforms the other models with 7B parameters.

Comparison of LLM models for coding

Online UI to Openchat. This seems really good, open source etc. It uses the LLama2 and Mistral models, according to https://github.com/imoneoi/openchat

Open source with lots of information. Uses Multiple undrelying models. Not sure how I would train for it

The Mixtral model is new, and seems to be good. Click on “Demo“ to test it

Article has comparison with other code-LLM models

Includes code generation models

Review of LLM specialized for code generation

List of LLM models on Wikipedia

Chat, models. Not open source, but instruct and relatively small (3B). The 3B instruct may be the best to try on Newspeak.

Main page of ChatGPT

Chat on Mistral. Does well on Python and Smalltalk

Their models are open source.

Model which generates code for Python, Javascript, Go, Shell, Perl, Swifg, Ruby, PHP

Chat with Huggingface's chat engine

Gemini chat from Google. Can generate Python and other other code.

Advanced coding Our first version of Gemini can understand, explain and generate high-quality code in the world’s most popular programming languages, like Python, Java, C++, and Go. Using a specialized version of Gemini, we created a more advanced code generation system, AlphaCode 2,

AI Code tools : Good summary. Does not talk about which pre-trained model they use. One is gemini (bard) -> alphacode2

Jupyter notebook to test Bert

Natural Languge processing - full course.