Glossary¶

This glossary defines key MATE terms for new users.

This page is part of the MATE 0.2.7 documentation set. See Documentation History for the release history.

Core platform¶

Term	Definition
MATE	MUSA AI Tensor Engine, an operator library for generative AI and large language model workloads on Moore Threads GPUs.
MUSA	Moore Threads’ GPU software and hardware platform.
MUSA SDK	The compiler and runtime toolkit for MUSA development.
TorchMUSA	The PyTorch integration layer for MUSA devices.
GPU	Moore Threads GPU hardware.

Compatibility wrappers¶

Term	Definition
wrapper	A compatibility package that preserves an upstream package surface while routing supported calls to MATE.
native MATE API	The direct `mate` Python APIs. Use these when no wrapper matches your workload or when you need lower-level control.
FlashAttention-3 wrapper (`flash_attn_3`)	Compatibility wrapper for FlashAttention-3 style attention APIs.
FlashMLA wrapper (`flash_mla`)	Compatibility wrapper for FlashMLA-style MLA APIs.
FlashKDA wrapper (`flash_kda`)	Compatibility wrapper for Kimi Delta Attention APIs.
FlashInfer wrapper (`flashinfer-python`)	Compatibility wrapper for selected FlashInfer GEMM, Norm, RoPE, and sparse decode APIs.
DeepGEMM wrapper (`deep-gemm`)	Compatibility wrapper for DeepGEMM-style GEMM and logits APIs.
MSA wrapper (`fmha_sm100`)	Compatibility wrapper for MiniMax Sparse Attention APIs.

Feature terms¶

Term	Definition
FlashAttention	The MATE FMHA attention path, including dense, varlen, and KV-cache workflows.
FlashMLA	The MATE MLA path for metadata, decode, and sparse prefill workflows.
KDA	Kimi Delta Attention. The MATE KDA APIs cover chunked KDA and decode.
FlashInfer Norm	The FlashInfer-compatible RMSNorm, LayerNorm, fused add Norm, and quantization APIs.
DeepGEMM	The DeepGEMM-compatible GEMM and logits API surface.
MSA (MiniMax Sparse Attention)	The MATE attention path for dense, paged, and sparse MSA workflows.
GDN (Gated Delta Network)	The MATE APIs for GDN decode and prefill workflows.

Installation and runtime¶

Term	Definition
wheel	A Python distribution artifact that can be installed with `pip`.
package source	The `pip` index or repository that provides a package.
editable install	An install that points Python at the local checkout while you edit it.
source build	Building a package directly from the repository checkout.
local wheel	A wheel built from the local checkout.
build baseline	The minimum version needed to build or run a feature.
runtime	The stage where an installed package or binary is executed.
validation	The checks used to confirm that installation or behavior is correct.
environment variable	A shell or process setting that changes MATE behavior.
AOT	Ahead-of-time compilation.
JIT	Just-in-time compilation.
MUBIN	Pre-generated kernel metadata and binary artifacts used by selected operator paths.
`+musa`	The PEP 440 local version suffix used by MUSA wrapper packages.