Dan Filimon

Hi there! Currently we're not using concurrent data structures to keep the overhead as low as possible. There's a thread about this on the Mahout mailing list [1] about this...

Yeah, no worries, I patched it up and ran it. Could you please look at the thread on the mailing list? :)

Would love to get an update for this as well. :)

I'm encountering it too, and the best I can find is that it's likely a recently introduced bug as part of a refactor - I see references to a refactor...

I can confirm that checking out the [parent of that commit](https://github.com/mlc-ai/relax/commit/ec548eb6145171b9cdeb654d96b9e39db1bf771e) works for TVM - ```shell (mlc-src) ➜ tvm git:(ec548eb61) ✗ git checkout ec548eb ``` But you then also need...

Yes, you need to uninstall and rebuild both from source, but that will happen in the pip step automatically at the end. For MLC: - https://llm.mlc.ai/docs/install/mlc_llm.html#option-2-build-from-source For TVM: - https://llm.mlc.ai/docs/install/mlc_llm.html#option-2-build-from-source...

I am not sure it's a RAM issue as much as an issue with the model compilation for older iOS devices. The error is from the Metal compiler as it...

I will add that the crash persists even with a 0.5B parameter model - The Metal compilation error makes me think that somehow the TVM-generated Metal code can't be correctly...

Going on a hunch, but I'm wondering whether one of the authors of https://github.com/mlc-ai/relax/blame/79a69ae4a92c9d4f23e62f93ce5b0d90ed29e5ed/src/runtime/metal/metal_module.mm#L100 could help us out? Please @echuraev @tqchen? 😄

I thought about it more and it's interesting that the function is a matrix multiplication operation specifically and others seem to work. This got me thinking that maybe the problem...