ascend-npu

A toolkit for AI agents used for development on Triton-Ascend for Ascend NPU

RAG for visually rich documents — layout, formula and table structure survive the pipeline. Ported from CUDA to Huawei Ascend NPUs and tested on Ascend 910C.

Bilingual SGLang tutorials and source-code walkthroughs for LLM inference, AI infrastructure, scheduling, KV cache, attention, Ascend NPU, and kernel optimization. | SGLang 中英双语教程与源码解析

Visual workflow studio for orchestrating AI assistant pipelines: intent routing, retrieval, tool calls, and model orchestration, prototyped for AI PC / HarmonyOS-style assistant scenarios.

This is the NPU version of the repository https://github.com/NJUNLP/trans0 . The README provides users with a simple guide for setting up the NPU environment. For more detailed descriptions of the methods proposed in the paper, please refer to the original repository.

Mini-SGLang port for Ascend NPUs with FIA, paged KV cache, ragged continuous batching, and request lifecycle safety.

High-performance FlashAttention forward kernels for Ascend NPUs, built with TritonAscend.

Lightweight realtime heterogeneous accelerator monitoring with workload history.

MindX DL 组件部署/ 静态虚拟化、动态虚拟化 6.0.RC2

16卡昇腾910B2C部署DeepSeek-V4-Flash：512K上下文、DSpark、vLLM-Ascend性能调优与故障排查实战

Cholesky decomposition reference implementation on Ascend NPU

C++ daemon for multi-camera streaming (GStreamer + WebRTC via MediaMTX) and hardware-accelerated YOLO12 object detection on Orange Pi AIpro 20T (Ascend NPU). Features robust USB camera hot-plug recovery.

Native AscendC Mamba2 selective scan / SSD forward-backward custom operator for Huawei Ascend 910B3 and 950PR, with CANN, torch_npu, A100 benchmarks and msprof profiling.

GRPO reinforcement learning for LLMs on math problems with verifiable reward — trained on Ascend NPU with verl

两张昇腾910B2C部署Qwen3.8-27B W8A8：128K上下文、多模态、工具调用与vLLM-Ascend实战

Inference acceleration framework for breast cancer classification based on UNI and Huawei Ascend NPU, integrating structured pruning, Decoupled Knowledge Distillation (DKD), SVD, and INT8 quantization for edge-cloud deployment. 基于 UNI 病理大模型与华为昇腾 NPU 的乳腺癌分类推理加速方案，集成结构化剪枝、解耦知识蒸馏 (DKD)、SVD 低秩分解及 INT8 量化，实现端云协同部署。