Overview¶

MUSA AI Tensor Engine (MATE) accelerates generative AI workloads on Moore Threads GPUs by providing high-performance operator implementations, especially Attention and GEMM, along with compatibility wrappers for selected CUDA-oriented Python operator interfaces.

For a deeper explanation of how MATE is structured, see Design and Architecture.

Key Principles¶

Wrapper-first: when a wrapper matches your existing package surface, keep the upstream import path and high-level API shape as stable as possible.
Direct API fallback: use MATE Python APIs when no wrapper matches your workload or wrapper coverage is insufficient.
Diagnostics early: verify the runtime with mate check, mate show-config, and mate env before debugging deeper failures.

Key Goals¶

Run high-performance generative AI workloads on Moore Threads GPUs with optimized Attention and GEMM operators.
Reduce migration work for CUDA-oriented integrations by preserving familiar package surfaces when wrapper coverage exists.
Provide a clear path from installation to wrapper selection, runtime verification, and failure diagnosis.
Surface actionable debug artifacts, including logs, configuration, dumps, and replay data, when an integration fails.

Typical Workflow¶

Prepare a supported runtime.

Start with a MUSA-enabled torch / torch_musa stack.
Choose the integration surface.

Start with FlashAttention-3, SageAttention, FlashMLA, FlashKDA, or DeepGEMM when one matches your framework surface.
Install the delivered package.

Install the matching wrapper from the wheel source when possible. The wrapper pulls in the matching mate package automatically. Install mate directly when no wrapper matches.
Verify the runtime.

Run mate check, mate show-config, and mate env.
Debug or fall back to APIs.

If wrapper coverage does not meet your needs, continue with MATE Python APIs.

Next steps: Installing MATE -> Wrappers -> CLI & Diagnostics -> Python APIs

For version-specific changes, see GitHub Releases.