Open dnhkng opened 5 months ago
I second this! It is a very impressive SLM that passed all my preliminary tests. Curious as to why it failed? @dnhkng did you use the most up-to-date driver?
I used the docker instructions from https://github.com/Pelochus/ezrknn-llm
I think this probably needs official support from the rockchip team, as the model is different to Phi-2.
Hmm... quickly reviewed it and it doesn't seem to use the newest driver. The docker file should be updated to use this OS:
https://github.com/Pelochus/armbian-build-rknpu-0.9.6
As that removes the 4GB limit which is what I believe is causing your error but all an assumption.
The error was in converting the model to RKLLM format, not running the model. The error I posted was in loading the model in the conversation script.
Rockchip should opensource the SDK. There are people ready, willing and able to make it compatible and functional. It seems that this is not a high priority for Rockchip. Let the community help make this a viable product. If Rockchip had LLM functionality that worked as well as the vision models, more people would be willing to purchase boards with their chips.
v1.0.1 has been released, it supports Phi-3.
I will soon post a project link that uses this in a very interesting way!
I will soon post a project link that uses this in a very interesting way!
Oh I had no idea you wanted to run that on an RK3588 SBC neat. I actually have been looking for an SBC/mini pc to try and mirror what you have done.
I will soon post a project link that uses this in a very interesting way!
Will it run GlaDOS? Would love to see that. I just started digging into it last night (saw you post it on Reddit some weeks ago).
That's the goal 😉
I tried to do a model conversion for the Microsoft Phi-3, and got the following error:
This model should have a very high priority. Its probably the only model that can run at a good tokens/second on the 3588 and also be useful. It would be a very good showcase for the 3588 and similar systems.