llm-foundry
sparsegpt
llm-foundry | sparsegpt | |
---|---|---|
37 | 16 | |
3,730 | 634 | |
4.0% | 5.0% | |
9.7 | 2.4 | |
5 days ago | about 1 month ago | |
Python | Python | |
Apache License 2.0 | Apache License 2.0 |
Stars - the number of stars that a project has on GitHub. Growth - month over month growth in stars.
Activity is a relative number indicating how actively a project is being developed. Recent commits have higher weight than older ones.
For example, an activity of 9.0 indicates that a project is amongst the top 10% of the most actively developed projects that we are tracking.
llm-foundry
-
Fine Tuning Mistral 7B on Magic the Gathering Draft
Related comment from gwern: https://news.ycombinator.com/item?id=38438859
Also - why qlora rather than a full finetune? Using LambdaLabs, It'd cost roughly the same as your quote. Cheaper I think if you're willing to gamble with fp8: https://github.com/mosaicml/llm-foundry/tree/main/scripts/tr.... And fewer hyperparameters to tune as well
-
Consortium launched to build the largest open LLM
Traditionally, training runs can "explode" and fail, but there are methods to incrementally back them up and resume when that happens, see https://www.mosaicml.com/blog/mpt-7b
-
Applying All Recent Innovations To Train a Code Model
MosaicML released the MPT-7B model, which has a context of 60k tokens, thanks to the ALiBi position encoding.
-
Fine Tuning Language Models
Most AI runners just ignore licensing and run LLaMA finetunes.
But if you want to avoid the non commercial LLaMA license, you have 3 good options for a base model.
- OpenLlama 13B
- MPT 30B
- Falcon 40B
Of these, Falcon 40B is very difficult to run (slow in 4 bit, basically requires a professional GPU, no good cpu offloading yet).
OpenLLaMA 13B only supports a context size of 2048 as of today... But that could change soon.
So you probably want MPT instruct 30B, specifically this one:
https://huggingface.co/TheBloke/mpt-30B-instruct-GGML
As the page says, you can try it out on a decent PC of your own with the OpenCL build of KoboldCPP. Change it to "instruct" mode, use the template on the page, offload as many layers as you can to your PC's dGPU, and run it in instruct mode. It may already work for your summarization needs.
If not, you can finetune it with MPT's code and summarization d
https://github.com/mosaicml/llm-foundry
Or train OpenLLaMA 13B with SuperHOT + summarization data using QLORA.
-
Finetune MPT-30B using QLORA
BTW. they finally merged a MPT patch to work with lora: https://github.com/mosaicml/llm-foundry/issues/304
- [N] Meet MPT-30B: A Fully OpenSouce LLM that Outperforms GPT-3 - Dr. Mandar Karhade, MD. PhD.
-
MPT-30B QLoRA on 24 GB VRAM
Did you run into this error while using qlora on MPT30b?: https://github.com/mosaicml/llm-foundry/issues/413
-
MosaicML Agrees to Join Databricks to Power Generative AI for All
Yes? Their github is under Apache, their base model is under apache, the training data is not theirs, and they provide scripts how to convert it for the pretrain step. They have scripts for pretraining and finetuning as well. Basically for everything.
-
Best model for commercial use?
mosaicml/llm-foundry: LLM training code for MosaicML foundation models (github.com)
-
MosaicML launches MPT-30B: A new open-source model that outperforms GPT-3
MosaicML, a company that provides a platform for training and deploying large language models (LLMs), has recently released its second open-source foundation model called MPT-30B. The model is part of the MosaicML Foundation Series and comes after the smaller MPT-7B model that was launched in May 2023.
sparsegpt
-
(1/2) May 2023
SparseGPT: Massive Language Models Can Be Accurately Pruned in One-Shot (https://arxiv.org/abs/2301.00774)
- Why Falcon going Apache 2.0 is a BIG deal for all of us.
-
New Open-source LLMs! 🤯 The Falcon has landed! 7B and 40B
There is this : https://github.com/IST-DASLab/sparsegpt
-
Webinar: Running LLMs performantly on CPUs Utilizing Pruning and Quantization
Check the paper here, it's intersting: https://arxiv.org/abs/2301.00774
-
OpenAI chief goes before US Congress to propose licenses for building AI
There's no chance that we've peeked from a bang for buck sense - we still haven't adequately investigated sparse networks.
Relevantish: https://arxiv.org/abs/2301.00774
The fact that we can reach those levels of sparseness with pruning also indicates that we're not doing a very good job of generating the initial network conditions.
Being able to come up with trainable initial settings for sparse networks across different topologies is hard, but given that we've had a degree of success with pre-trained networks, pre-training and pre-pruning might also allow for sparse networks with minimally compromised learning capabilities.
If it's possible to pre-train composable network modules, it might also be feasible to define trainable sparse networks with significantly relaxed topological constraints.
-
How to run Llama 13B with a 6GB graphics card
Training uses gradient descent, so you want to have good precision during that process. But once you have the overall structure of the network, https://arxiv.org/abs/2210.17323 (GPTQ) showed that you can cut down the precision quite a bit without losing a lot of accuracy. It seems you can cut down further for larger models. For the 13B Llama-based ones, going below 5 bit per parameter is noticeably worse, but for 30B models you can do 4 bits.
The same group did another paper https://arxiv.org/abs/2301.00774 which shows that in addition to reducing the precision of each parameter, you can also prune out a bunch of parameters entirely. It's harder to apply this optimization because models are usually loaded into RAM densely, but I hope someone figures out how to do it for popular models.
- SparseGPT: Language Models Can Be Accurately Pruned in One-Shot
What are some alternatives?
qlora - QLoRA: Efficient Finetuning of Quantized LLMs
StableLM - StableLM: Stability AI Language Models
basaran - Basaran is an open-source alternative to the OpenAI text completion API. It provides a compatible streaming API for your Hugging Face Transformers-based text generation models.
github-copilot-product-specific-terms
RasaGPT - 💬 RasaGPT is the first headless LLM chatbot platform built on top of Rasa and Langchain. Built w/ Rasa, FastAPI, Langchain, LlamaIndex, SQLModel, pgvector, ngrok, telegram
promptfoo - Test your prompts, models, and RAGs. Catch regressions and improve prompt quality. LLM evals for OpenAI, Azure, Anthropic, Gemini, Mistral, Llama, Bedrock, Ollama, and other local & private models with CI/CD integration.
LMFlow - An Extensible Toolkit for Finetuning and Inference of Large Foundation Models. Large Models for All.
chat-ui - Open source codebase powering the HuggingChat app
prompt-engineering - ChatGPT Prompt Engineering for Developers - deeplearning.ai
intel-extension-for-pytorch - A Python package for extending the official PyTorch that can easily obtain performance on Intel platform
llm-numbers - Numbers every LLM developer should know
geov - The GeoV model is a large langauge model designed by Georges Harik and uses Rotary Positional Embeddings with Relative distances (RoPER). We have shared a pre-trained 9B parameter model.