Ascend
Official@ascend
昇腾万里，让智能无所不及
Agent Skills by Ascend
Showing 101 vetted skills indexed across 2 GitHub repositories.
Ascenddummy-run
Validates model architecture compatibility on Ascend NPU using random weights without downloading real weights.
Ascenddit-parallel-opt
Selects and validates multi-card parallelism strategies for DiT inference on Ascend NPUs.
Ascendhost-opt
Optimizes host-side fixed overhead in inference services including encoding, transfer, loading, and warmup.
Ascendenv-install
Install MindIE-SD and third-party inference frameworks on Ascend NPU environments.
Ascenddev-workflow
Orchestrates MindIE-SD repository development with Test-First loops, parallel execution, and retrospective archiving.
Ascendmindie-sd-community-governance
Enforces MindIE-SD documentation, commit, PR, and governance conventions with fixed review steps.
Ascendaclgraph-dev
Develops and tunes NPUGraph static capture and replay for batched NPU inference dispatch.
Ascendprofiling-analyze
Analyzes Ascend NPU profiling data to diagnose bottlenecks and compare kernel-level performance across runs.
Ascendpattern-dev
Develops and debugs PyTorch Inductor pattern matcher fusions for NPU compile graphs.
Ascendvae-opt
Validates and implements bit-exact multi-card sharding for VAE/TAE video decoders.
Ascendmodel-auto-optimization
Orchestrates staged NPU optimization of diffusion models across fusion, parallelism, and quantization.
Ascendremote-access
Execute commands, select idle NPU cards, and transfer files on remote Ascend servers over SSH.
Ascendquantization-dev
Reverse-engineer quantization contracts and verify bit-exact alignment of custom quantizers on Ascend NPU.
Ascendaccuracy-gate
Validates whether code optimizations preserve numerical results using layered equivalence criteria and staged acceptance tests.
Ascendcode-standards
Enforces MindIE-SD Python formatting and lint rules via ruff, pylint, bandit, and pre-commit.
Ascendbenchmark-dev
Benchmarks and extends the mindie_bench performance toolchain for MindIE-SD core operators.
Ascendperf-gate
Validates performance optimization claims on shared Ascend NPU machines using same-window A/B measurement protocols.
Ascendoperator-dev
Route operator development and optimization tasks to cannbot-skills with MindIE-SD-specific integration guidance.
Ascenddit-perf-opt
Selects and validates DiT compute optimization tiers for quantization, sparsity, caching, and compilation on NPU.
Ascendmarkdown-lint
Validates and fixes Markdown files against markdownlint rules in pre-commit and CI workflows.
Ascendframework-integration
Enable and verify mindiesd inference features in third-party frameworks like vLLM-Omni and LightX2V.
Ascendperformance-optimization
Routes confirmed performance bottlenecks to the correct optimization module for MindIE-SD diffusion models.
Ascendfusion-scope-analyze
Determines operator fusion boundaries and estimates fusion benefit from profiling data.
Ascendprofiling-collect
Collect NPU profiling data from remote Ascend devices via SSH and patch-based instrumentation.
Frequently Asked Questions About Ascend
FAQPage SchemaWhat specific tasks can engineers perform using these Ascend skills?▼
Engineers can develop, profile, and optimize custom operators using AscendC or Triton, migrate models from GPU to NPU architectures, and manage distributed training configurations. The suite includes capabilities for precision debugging, memory error analysis, and automated documentation generation for operator interfaces.
Which technical personas are the primary users of these capabilities?▼
These skills are designed for AI infrastructure engineers, kernel developers, and machine learning researchers working on Huawei Ascend hardware. They are specifically tailored for professionals responsible for hardware-level performance tuning, model porting, and maintaining large-scale distributed training clusters.
What are the prerequisites for deploying these operator development skills?▼
Deployment requires a configured Linux environment with the CANN toolkit installed and accessible Ascend NPU hardware. Users must have the necessary driver and firmware versions verified via npu-smi, along with containerized environments for isolated operator compilation and testing.