评估 CANN 算子仓的「进阶教程 / 开发指南」类文档质量——通读文档 + 对照算子代码**静态**查证(默认不跑),按五轴(找得到/信得过/学得会/可操作/读得懂)找出 漏讲/讲不清/过时/对不上代码/概念讲错,产出带证据与改进建议的体检报告(MD + HTML)。涉及「评进阶教程 / 开发指南文档质量 / 文档对不对得上代码 / 教程审稿 / tutorial 体检 / 文档信不信得过」等意图时使用。只评不改不跑,只对着文档与代码出诊断。
Run NPU inference/training tests on a remote SSH server with vllm-ascend Docker container. Use when the user asks to test models on NPU, run inference on Ascend devices, or deploy models to an SSH server.
Track daily PRs and Issues from vllm-project/vllm and vllm-project/vllm-ascend, filter by model (DeepSeek/Qwen/GLM/MiniMax/Kimi) and tech topics (PD disaggregation, MTP, quantization, graph mode, performance), analyze with LLM, and generate a Markdown report. Use when user wants vllm daily tracker, PR/Issue digest, or Ascend inference ecosystem monitoring.
Automatically fetch InferenceX benchmark data and generate daily performance reports for LLM inference on various hardware (NVIDIA, AMD, etc.). Supports email delivery, data change detection, and 8k1k sequence length performance analysis. Use when needing to track LLM inference performance trends, compare hardware configurations, or monitor benchmark updates.
通用 PyTorch 项目 Ascend NPU 迁移可行性分析。系统化扫描代码库中的 CUDA/GPU 依赖，按 7 大域分类评估（设备层、注意力机制、自定义算子、分布式通信、精度策略、第三方依赖、编译加速），并基于 Wan2.2 实际迁移经验提供逐项替代方案。适用于评估任何 DiT/Transformer 类模型（视频生成、LLM、多模态等）在昇腾 NPU 上的运行可行性与迁移工作量估算。
End-to-end AscendC custom operator development for Ascend NPU in an ascend-kernel (csrc/ops + build.sh + torch_npu PyTorch custom op) project. Use to design, generate, build, test, document, and tune a new AscendC operator from a name and a math/functional spec. Covers project init, two-level tiling design, op_host/op_kernel code generation, framework registration, compile/install/debug, PyTorch-style API docs, precision evaluation and root-cause debugging, torch_npu.profiler performance benchmarking, performance optimization, and security code review.
模型层 Tensor 打点与精度对比工具。用于在模型 forward 过程中捕获模型各层中间 tensor，实现 GPU/NPU 精度对比调试。支持 vLLM、SGLang 推理框架。When to use： When you need to debug precision issues between GPU and NPU,or validate layer-wise tensor outputs during inference.
Migrate any HuggingFace model to Ascend NPU torchair graph mode (torch.compile) and benchmark it for accuracy and performance against NPU eager and CPU eager. Use when running, compiling, or benchmarking HF models (vision, text, image-text encoders such as SigLIP2, DINOv3, ViT, CLIP, Qwen-VL, SAM) on Ascend 910B/CANN with torch_npu and torchair; when a torch.compile graph-mode run on NPU fails (Dynamo TorchRuntimeError, unsupported op, interpolate/contiguous errors); or when comparing torchair vs npu_eager vs cpu with cosine similarity, max abs diff, and p50/p95/p99 latency.