I have a fairly substantial 16gb AMD GPU, and when I load in Llama 3.1 8B Instruct 128k (Q4_0), it gives me about 12 tokens per second. That’s reasonably fast enough for me, but only 50% faster than CPU (which I test by loading mlabonne’s abliterated Q4_K_M version, which runs on CPU in GPT4All, though I have no idea if that’s actually meant to be comparable in performance).
Then I load in Nous Hermes 2 Mistral 7B DPO (also Q4_0) and it blazes through at 50+ tokens per second. So I don’t really know what’s going on there. Seems like performance varies a lot from model to model, but I don’t know enough to speculate why. I can’t even try Gemma2 models, GPT4All just crashes with them. I should probably test Alpaca to see if these perform any different there…
Unfortunately I can’t even test Llama 3.1 in Alpaca because it refuses to download, showing some error message with the important bits cut off.
That said, the Alpaca download interface seems much more robust, allowing me to select a model and then select any version of it for download, not just apparently picking whatever version it thinks I should use. That’s an improvement for sure. On GPT4All I basically have to download the model manually if I want one that’s not the default, and when I do that there’s a decent chance it doesn’t run on GPU.
However, GPT4All allows me to plainly see how I can edit the system prompt and many other parameters the model is run with, and even configure multiple sets of parameters for the same model. That allows me to effectively pre-configure a model in much more creative ways, such as programming it to be a specific character with a specific background and mindset. I can get the Mistral model from earlier to act like anything from a very curt and emotionally neutral virtual intelligence named Jarvis to a grumpy fantasy monster whose behavior is transcribed by a narrator. GPT4All can even present an API endpoint to localhost for other programs to use.
Alpaca seems to have some degree of model customization, but I can’t tell how well it compares, probably because I’m not familiar with using ollama and I don’t feel like tinkering with it since it doesn’t want to use my GPU. The one thing I can see that’s better in it is the use of multiple models at the same time; right now GPT4All will unload one model before it loads another.