Lukasz Stafiniak

Work out a JS or WASM compilation target with a WebGPU-based backend

2

This is in addition to the regular, native compilation with the multiple backends. The JS target would support the interpreter and WebGPU backends, and the native target would support all...

See what we can learn for efficiency from the `ggml` project

1

[ggml](https://github.com/ggerganov/ggml)

[Extremely sparse computational graphs produced by training and pruning RNNs, in Tinygrad.](https://cprimozic.net/blog/growing-sparse-computational-graphs-with-rnns/)

Incorporate examples from François Fleuret's lectures

1

[DEEP LEARNING COURSE](https://fleuret.org/dlc/) It has handouts, slides, video recordings. 13 lectures with 1.5-2h of recordings per lecture. PyTorch based.

Reproduce Mamba

1

https://github.com/state-spaces/mamba _Mamba: Linear-Time Sequence Modeling with Selective State Spaces_ Albert Gu, Tri Dao [Paper](https://arxiv.org/abs/2312.00752)

https://arxiv.org/abs/2401.17992

Do a deep dive into TVM

2

https://github.com/apache/tvm GeoHot mentioned it inspired some TinyGrad solutions.

https://github.com/jafioti/luminal From-scratch framework in Rust. Mentioned on the TinyGrad discord.

Currently it always reports 0.

Consider allowing multiple virtual tensors to share the same for loop

1

In the initial implementation, we in effect assume without checking that for-loops only compute a single node. This is currently true in particular because we do not have for-loops in...