docs.rs failed to build teeny-kernels-0.2.0
Please check the build logs for more information.
See Builds for ideas on how to fix a failed build, or Metadata for how to configure docs.rs builds.
If you believe this is docs.rs' fault, open an issue.

teeny-kernels

CPU/GPU kernel implementations of teenygrad's nn layers (attention — including a Flash Attention 2 forward/backward implementation — math ops, and graph lowering), written against the teeny-triton DSL and compiled via teeny-compiler.

Prerequisites

Rust: any stable or nightly toolchain to cargo build/cargo doc this crate — no system-library dependency at build time (the cuda feature pulls in teeny-cuda, which does have a CUDA toolkit build requirement — see its README).
Compiling/running kernels at runtime needs the custom teenyc compiler — see teeny-compiler's README for the cargo-teeny/TEENYC_PATH setup.
On Blackwell (sm_120) GPUs, teenyc's default PTX version for sm_120a may be rejected by the installed driver's JIT compiler (PTX .version 8.6 does not support .target sm_120a); if so, set TEENYC_PTX_VERSION=87 (see teeny-compiler's TEENYC_PTX_VERSION env var).

Benchmarks

benches/conv2d_bn_silu.rs compares the three fused Conv2d+BatchNorm+SiLU kernel variants (scalar, channel-tiled, GEMM/tensor-core) across shapes that straddle the shape-based dispatch thresholds in graph/mod.rs:

cargo bench -p teeny-kernels --features cuda,training --bench conv2d_bn_silu

Needs a real CUDA device (same runtime teenyc requirement as above); results are written to target/criterion/report/index.html.

Features

Feature	Default	Description
`cuda`	✅	Enables the `teeny-cuda` backend (requires the CUDA toolkit to build — see `teeny-cuda`'s README).
`training`	✅	Enables `teeny-core`'s `training` feature passthrough.

Getting started

[dependencies]
teeny-kernels = "0"

License