HN Explorer

9 comments

[ 3.6 ms ] story [ 35.3 ms ] thread
johntash
Can this be run locally without beefy GPUs by any chance?
ops
I haven't used this yet, but I am currently running GPT-J on my Mac Studio, so I suspect so.
jerpint
There have been CPU implementations of LLAMA (7b parameters, comparable in size) with very impressive performance
quesomaster9000
ggml (https://github.com/ggerganov/ggml) has a GPT-J example, the 6B parameter model runs happily on the CPU 16gb of ram and 8 cores at a couple of words per second, no GPUs necessary.

    gptj_model_load: ggml ctx size = 13334.86 MB
    gptj_model_load: memory_size =  1792.00 MB, n_mem = 57344
    gptj_model_load: model size = 11542.79 MB / num tensors = 285
    main: number of tokens in prompt = 12

    An example of GPT-J running on the CPU is shown in Fig. [4](#Fig4

    main: mem per token = 16179460 bytes
    main:     load time =  7463.20 ms
    main:   sample time =     3.24 ms
    main:  predict time =  4887.26 ms / 232.73 ms per token
    main:    total time = 13203.91 ms
jarrell_mark
It should work with about 12gb GPU RAM.

I got it to load on a GTX 1070 with 8GB GPU RAM, but then it crashed before it could generate a response.

It needs less RAM than regular GPT-J because the weights are converted to 8-bit

two_handfuls
"Privacy-First", but also working in a colab notebook - meaning running on someone else's machine? That doesn't seem very private.
was_a_dev
Download the notebook and run locally?
jarrell_mark
Yes, the GitHub has the Jupyter .ipynb notebook that can be run locally: https://github.com/jarrellmark/chatgpt-j

And even in Colab, it's privacy first in the sense that user input or model output isn't being sent anywhere. The data is local to your Colab session.