mirror of
https://github.com/hanzoai/ml.git
synced 2026-11-20 22:54:34 +00:00
Three-way per file through the rename: base 39355c6c91, theirs the upstream
merge of main with 35d7ae7c (via 7ef6a694d), ours origin/main.
Lands the barrier thread pool, runtime CPU feature dispatch, aarch64 and x86
repacked quantized kernels (indexed_gemv, gemv_fused_shared_lhs), quantized
embedding on CPU/Metal/CUDA, paged flash-attn kernels, cutile fused MoE, the
in-place op aliasing API, parking_lot storage locks, Metal buffer builders and
debug labels, safetensors 0.8, hf-hub 1.0, tokenizers 0.23.
Ours keeps ROCm, Vulkan, wgpu, the IQ/MXFP4/NVFP4/TQ dtypes, and the i-quant
weight/activation block split in matmul. Our backends ride InplaceOpN and the
quantized embedding; Metal get_rows gains the IQ kernels. ug leaves, as upstream.
Co-authored-by: Hanzo Dev <[email protected]>
hanzo-ml-quantized-phi
ML implementation of various quantized Phi models.
Running an example
$ cargo run --example quantized-phi --release -- --prompt "The best thing about coding in rust is "
> - it's memory safe (without you having to worry too much)
> - the borrow checker is really smart and will catch your mistakes for free, making them show up as compile errors instead of segfaulting in runtime.
>
> This alone make me prefer using rust over c++ or go, python/Cython etc.
>
> The major downside I can see now:
> - it's slower than other languages (viz: C++) and most importantly lack of libraries to leverage existing work done by community in that language. There are so many useful machine learning libraries available for c++, go, python etc but none for Rust as far as I am aware of on the first glance.
> - there aren't a lot of production ready projects which also makes it very hard to start new one (given my background)
>
> Another downside: