tinygrad
GPTQ-for-LLaMa
tinygrad | GPTQ-for-LLaMa | |
---|---|---|
58 | 75 | |
17,800 | 2,916 | |
- | - | |
9.7 | 8.6 | |
10 months ago | 9 months ago | |
Python | Python | |
MIT License | 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.
tinygrad
- tinygrad: extreme simplicity, easiest framework to add new accelerators to
-
GGML – AI at the Edge
Might be a silly question but is GGML a similar/competing library to George Hotz's tinygrad [0]?
[0] https://github.com/geohot/tinygrad
-
Render neural network into CUDA/HIP code
at first glance i thought may its like tinygrad. but looks has many ops than that tiny grad but most maps to underlying hardware provided ops?
i wonder how well tinygrad's apporach will work out, ops fusion sounds easy, just a walk a graph, pattern match it and lower to hardware provided ops?
Anyway if anyone wants to understand the philosophy behind tinygrad, this file is great start https://github.com/geohot/tinygrad/blob/master/docs/abstract...
-
llama.cpp now officially supports GPU acceleration.
There are currently at least 3 ways to run llama on m1 with GPU acceleration. - mlc-llm (pre-built, only 1 model has been ported) - tinygrad (very memory efficient, not that easy to integrate into other projects) - llama-mps (original llama codebase + llama adapter support)
- George Hotz building an AMD competitor to Nvidia.
-
George Hotz ROCm adventures
Hopefully we will see now full support with AMD hardware on https://github.com/geohot/tinygrad. You can read more about it on https://tinygrad.org/
-
The Coming of Local LLMs
tinygrad
https://github.com/geohot/tinygrad/tree/master/accel/ane
But I have not tested it on Linux since Asahi has not yet added support.
llama.cpp runs at 18ms per token (7B) and 200ms per token (65B) without quantization.
- Everything we know about Apple's Neural Engine
- Everything we know about the Apple Neural Engine (ANE)
- How 'Open' Is OpenAI, Really?
GPTQ-for-LLaMa
-
[P] Early in 2023 I put in a lot of work on a new machine learning project. Now I'm not sure what to do with it.
First I want to make it clear this is not a self promotion post. I hope many machine learning people come at me with questions or comments about this project. A little background about myself. I did work on the 4 bits quantization of LLaMA using GPTQ. (https://github.com/qwopqwop200/GPTQ-for-LLaMa). I've been studying AI in-depth for many years now.
-
GPT-4 Details Leaked
Deploying the 60B version is a challenge though and you might need to apply 4-bit quantization with something like https://github.com/PanQiWei/AutoGPTQ or https://github.com/qwopqwop200/GPTQ-for-LLaMa . Then you can improve the inference speed by using https://github.com/turboderp/exllama .
If you prefer to use an "instruct" model à la ChatGPT (i.e. that does not need few-shot learning to output good results) you can use something like this: https://huggingface.co/TheBloke/Wizard-Vicuna-30B-Uncensored...
-
Rambling
I use gptq-for-llama - from this https://github.com/qwopqwop200/GPTQ-for-LLaMa and Pygmalion 7B.
-
Now that ExLlama is out with reduced VRAM usage, are there any GPTQ models bigger than 7b which can fit onto an 8GB card?
exllama is an optimized implementation of GPTQ-for-LLaMa, allowing you to run 4-bit quantized language models with GPU at great speeds.
-
GGML – AI at the Edge
With a single NVIDIA 3090 and the fastest inference branch of GPTQ-for-LLAMA https://github.com/qwopqwop200/GPTQ-for-LLaMa/tree/fastest-i..., I get a healthy 10-15 tokens per second on the 30B models. IMO GGML is great (And I totally use it) but it's still not as fast as running the models on GPU for now.
-
New quantization method AWQ outperforms GPTQ in 4-bit and 3-bit with 1.45x speedup and works with multimodal LLMs
And exactly what Triton version are they comparing against? I just tried the latest version of this, and on my 4090/12900K I get 77 tokens per second for Llama 7B-128g. My own GPTQ CUDA implementation gets 151 tokens/second on the same model, same hardware. That makes it 96% faster, whereas AWQ is only 79% faster. For 30B-128g I'm currently only getting a 110% speedup over Triton compared to their 178%, but it still seems a little disingenuous to compare against their own CUDA implementation only, when they're trying to present the quantization method as being faster for inference.
-
Introducing Basaran: self-hosted open-source alternative to the OpenAI text completion API
Thanks for the explanation. I think some repos, like text generation webui used gptq for llama (I don't know if it's this repo or another one), anyway most repo that I saw use external things (like gptq for llama)
-
How to use AMD GPU?
cd ../.. git clone https://github.com/qwopqwop200/GPTQ-for-LLaMa.git -b triton cd GPTQ-for-LLaMa pip install -r requirements.txt mkdir -p ../text-generation-webui/repositories ln -s ../../GPTQ-for-LLaMa ../text-generation-webui/repositories/GPTQ-for-LLaMa
-
Help needed with installing quant_cuda for the WebUI
cd repositories git clone https://github.com/qwopqwop200/GPTQ-for-LLaMa pip install -r requirements.txt
-
The installed version of bitsandbytes was compiled without GPU support
# To use the GPTQ models I need to Install GPTQ-for-LLaMa and the monkey patch mkdir repositories cd repositories git clone https://github.com/qwopqwop200/GPTQ-for-LLaMa.git -b triton cd GPTQ-for-LLaMa pip install ninja pip install -r requirements.txt cd cd text-generation-webui # download random model python download-model.py xxx/yyy # try to start the gui python server.py # It returns this warning but it runs bin /home/gm/miniconda3/envs/chat/lib/python3.10/site-packages/bitsandbytes/libbitsandbytes_cpu.so /home/gm/miniconda3/envs/chat/lib/python3.10/site-packages/bitsandbytes/cextension.py:34: UserWarning: The installed version of bitsandbytes was compiled without GPU support. 8-bit optimizers, 8-bit multiplication, and GPU quantization are unavailable. warn("The installed version of bitsandbytes was compiled without GPU support. " /home/gm/miniconda3/envs/chat/lib/python3.10/site-packages/bitsandbytes/libbitsandbytes_cpu.so: undefined symbol: cadam32bit_grad_fp32
What are some alternatives?
Pytorch - Tensors and Dynamic neural networks in Python with strong GPU acceleration
llama.cpp - LLM inference in C/C++
bitsandbytes - Accessible large language models via k-bit quantization for PyTorch.
openpilot - openpilot is an open source driver assistance system. openpilot performs the functions of Automated Lane Centering and Adaptive Cruise Control for 250+ supported car makes and models.
text-generation-webui - A Gradio web UI for Large Language Models. Supports transformers, GPTQ, AWQ, EXL2, llama.cpp (GGUF), Llama models.
llama - Inference code for Llama models
qlora - QLoRA: Efficient Finetuning of Quantized LLMs
tensorflow_macos - TensorFlow for macOS 11.0+ accelerated using Apple's ML Compute framework.
private-gpt - Interact with your documents using the power of GPT, 100% privately, no data leaks
stable-diffusion-webui-docker - Easy Docker setup for Stable Diffusion with user-friendly UI