• Zhen Xin's avatar
    BoltOps: Backend Optimized Low-latency Triton/TileLang Operators · 4c19e8ba
    Zhen Xin authored
    
    
    Introduce BoltOps (Backend Optimized Low-latency Triton/TileLang Operators) —
    also named for the lightning bolt (low latency, lightning-fast by design) —
    the AI Compiler team's high-performance operator library for Hygon DCU,
    shipping production-oriented kernels in Triton, TileLang, and other DSLs
    behind clean Python APIs for inference frameworks and model-side callers.
    
    Included operators:
    - fused_experts: fused MoE experts with quantized paths and autotune tooling
    - recompute_w_u_fwd: Gated Delta Net chunked WY recompute
    - fused_mamba_state_scatter_with_mask: MTP Mamba state writeback
    
    Also establish the per-operator package layout (triton/tilelang),
    docs, correctness tests, and wheel packaging.
    Co-authored-by: default avatarCursor <cursoragent@cursor.com>
    4c19e8ba
This project is licensed under the MIT License. Learn more
LICENSE 1.04 KB