API Reference

Greyhound exposes fused kernels through functional operators and small nn.Module wrappers. The functional layer is the most direct way to call kernels from model code, while the module layer is useful when replacing standard transformer blocks.

Detailed API reference pages are generated from Python docstrings with scripts/generate_api_docs.py.

Functional API

Function	Purpose
greyhound.nn.functional.cross_entropy	Cross-entropy using the standalone loss-and-logits-gradient kernel.
greyhound.nn.functional.autograd_loss_and_logits_grad	Adapter from scalar PyTorch losses to chunked loss-and-logits-gradient callbacks.
greyhound.nn.functional.chunked_linear_loss	Chunked linear projection with a user-provided per-chunk loss-and-logits-gradient function.
greyhound.nn.functional.chunked_linear_cross_entropy	Chunked linear cross-entropy using a fused cross-entropy-and-logits-gradient kernel per logits tile.
greyhound.nn.functional.causal_conv1d	Depthwise causal 1D convolution with optional bias and activation.
greyhound.nn.functional.selective_log_softmax	Selected log-softmax values without materializing full log-probabilities.

Generated functional reference

nn.Module Wrappers

Module	Purpose
greyhound.nn.causal_conv1d.GreyhoundCausalConv1d	Depthwise causal Conv1d wrapper backed by Greyhound’s fused convolution kernel.

Generated module wrapper reference

Bonus Integrations

Utility	Purpose
greyhound.bonus.newton_schultz.orthogonalize_via_newton_schulz	Muon-style Newton-Schulz orthogonalization composed with Quack symmetric GEMM when available.

Bonus integrations are experimental utilities that compose Greyhound code with kernels from other providers. They are documented separately from the core functional and module API because they may require optional provider dependencies.

Source Modules

The public API is implemented in the Python package under src/greyhound/:

Surface	Source
Functional namespace	src/greyhound/nn/functional.py
Module wrappers	src/greyhound/nn/
Autograd ops	src/greyhound/ops/
CuTe DSL kernels	src/greyhound/kernels/
Bonus integrations	src/greyhound/bonus/