tvm VS tch-rs

Yes. Web-llm is a wrapper of tvmjs: https://github.com/apache/tvm

Just wrappers all the way down

Yes, this is coming! Myself and others at OctoML and in the TVM community are actively working on multi-gpu support in the compiler and runtime. Here are some of the merged and active PRs on the multi-GPU (multi-device) roadmap:

Support in TVM’s graph IR (Relax) - https://github.com/apache/tvm/pull/15447

Would it make sense to have a backend support for OpenXLA, Apache TVM, Jittor or other similar to get free GPU, TPU and other accelerators for free ?

I have tried the iPhone app. It's fast. They're using Apache TVM which should allow better use of native accelerators on different devices. Like using metal on Apple and Vulcan or CUDA or whatever instead of just running the thing on the CPU like llama.cpp.

I was going to answer the same, I find the approach of machine learning compilers that directly compile models to host and device code better than having to bring a huge runtime. There are exciting projects in this area like TVM Unity, IREE [2], or torch.export [3]

[1] https://github.com/apache/tvm/tree/unity

[2] https://pytorch.org/get-started/pytorch-2.0/#inference-and-e...

[3] https://pytorch.org/get-started/pytorch-2.0/#inference-and-e...

Apache TVM home page: https://tvm.apache.org/

It's pretty clear its referring to the output of Apache TVM and Meta's Glow

As mentioned in sibling comments, Torch is indeed the glue in this implementation. Other glues are TVM[0] and ONNX[1]

These just cover the neural net though, and there is lots of surrounding code and pre-/post-processing that isn't covered by these systems.

For models on Replicate, we use Docker, packaged with Cog for this stuff.[2] Unfortunately Docker doesn't run natively on Mac, so if we want to use the Mac's GPU, we can't use Docker.

I wish there was a good container system for Mac. Even better if it were something that spanned both Mac and Linux. (Not as far-fetched as it seems... I used to work at Docker and spent a bit of time looking into this...)

[0] https://tvm.apache.org/

Or use TVM, the idea is to compile your model into code that you can load at runtime. Similar to onnxruntime, it only does DNN inference; so you need domain-specific code.

I wanted to do something like this but then I would miss on proper CUDA acceleration and lose performance compared to using torchlib.

I wrote a forgettable llama implementation for https://github.com/LaurentMazare/tch-rs (pytorch's torchlib rust binding).

I first encountered OCaml's PyTorch bindings because apparently they generate a C wrapper around PyTorch's C++ API, and Rust's PyTorch bindings use OCaml's C wrapper. See: https://github.com/LaurentMazare/tch-rs

You could try looking at the min-GPT example of tch-rs. I'd also strongly suggest watching Karpathy's video to understand what's going on.

If you pefer to see it in code there's a succint gpt implementation here https://github.com/LaurentMazare/tch-rs/blob/main/examples/m...

Rust is fully compatible w/ C bindings, so even Python libraries written in C can be easily set up to work in Rust (and have been). For example, see PyTorch Rust bindings, which actually works faster than in Python because all of the glue code around the C++ API is in Rust instead of Python.

Pytorch

I have looked into this a bit and found some crates which help in loading torchscript models called tch-rs