Reduced and now seemingly removed support for Mali?

[federicoparra]

❓ General Questions

I've noticed that your documentation is not thorough with regards to the Mali GPU, and now even the blog has been removed.

This is in my opinion such a bad idea :(

Perhaps the very best argument to move from llama.cpp to MLC was the support for this very cheap computer (the Orange Pi).

[tqchen]

Thank you for input. We love to support mali, the new link can be found here https://blog.mlc.ai/2024/04/20/GPU-Accelerated-LLM-on-Orange-Pi

I think it was due to we updated the post for llama3 and changed a date, so the blogpost link get updated. If you find there are more places for improvements, we are all ears

[tqchen]

I also added a redirection so the old link also works. Thanks @federicoparra for spotting it

[federicoparra]

Could you build pip versions for Orange Pi 5?
If there you need someone to do this I could !

[federicoparra]

Also, since I got your attention :) @tqchen could answer me this #2244 namely is there a way with MLC compilation to run an iterative optimization for one's specific harware akin to what AutoTVM does?
Also, is there a way to use GGUF weights that have already been quantized #2273 with MLC ? In the case of Phi-3 something rather miracoulous happened: Microsoft themselves posted int4 quantized weights, and I tested them and it works just like the 32 bit unquantized version, I suspect this is due to quantization-aware training, something that can't be achieved with quantization after training the likes MLC does, so it would be cool to be able to use their quantized weights ?

[tqchen]

On pip packages, MLC pip packages are automated nightly via https://github.com/mlc-ai/package/, we use github actions. I think one main barrier for orange pi(ARM64 build) is to have GH action(which runs on x86) to be able to cross build for that. We do already have Mac arm64 runner. So if docker can work in that env, maybe that is another route.