skills.rest
Ascend
AscendOFFICIAL
Explore all skills by this author →
WebsiteGitHub

env-install

Install MindIE-SD and third-party inference frameworks on Ascend NPU environments.

14|5|Updated Mar 30, 2026
Search Tags:#environment setup#vllm-omni#modelscope#source build#ascend npu#docker deployment#mindiesd
One-click install
npx skills add https://github.com/Ascend/MindIE-SD --skill env-install-ascend
Or copy as Structured Prompt for Agent▼
Please help me install this Agent Skill.
Skill: env-install
Source: https://github.com/Ascend/MindIE-SD/tree/main/.agents/skills/env-install
Command: npx skills add https://github.com/Ascend/MindIE-SD --skill env-install-ascend

SYSTEM DOCUMENTATION & REQUIREMENTS

💡 This Skill requires paramiko, and includes scripts (resource) and references (resource) components.

What problem does it solve? Setting up a working Ascend NPU inference environment involves compiling MindIE-SD from source, building the full vLLM-Omni stack (torch, torch_npu, vllm, vllm-ascend, vllm-omni), deploying LightX2V and DiffSynth-Engine, and downloading large model weights to remote containers. This Skill consolidates those installation paths, version compatibility matrices, and known pitfalls into one guided workflow. ## Core Features & Use Cases - MindIE-SD Compile & Install: Local Ascend direct install, SSH push to a remote Docker container via an incremental deploy script, or direct use of official prebuilt images, with compatibility prechecks (PyTorch, TorchNPU, CANN, Python) and build workarounds such as the build_tik_ops.sh fix. - Third-Party Framework Full-Stack Builds: Source builds of the vLLM-Omni stack when official images do not cover the target architecture, plus DiffSynth-Engine and LightX2V editable deployments with the required PLATFORM=ascend_npu setup. - Model Weight Preparation: Confirm whether weights already exist remotely before downloading, then fetch partitions from ModelScope (default) or HuggingFace with integrity verification. - Use Case: You need to push modified MindIE-SD code from a local dev machine into a remote Ascend container, compile it, and prepare MiniMax-H3 weights for vLLM-Omni serving. The Skill walks you through credential confirmation, incremental transfer, in-container build, and weight validation. ## Quick Start Ask the assistant to install MindIE-SD and the vLLM-Omni stack into your remote Ascend container and prepare the required model weights.

Frequently Asked Questions about env-install

High-intent search queries and answers about installing and using this skill.

FAQPage Schema
How do I install MindIE-SD on a remote Ascend NPU container?▼

Use the deploy_to_remote.py script with host, user, workspace, container, and local-root arguments. It incrementally transfers source files over SFTP, converts CRLF line endings, then runs python setup.py build_py and pip install -e . inside the container after sourcing the CANN environment.

How do I build the vLLM-Omni full stack from source for Ascend?▼

When official images do not cover your architecture, install torch 2.11.0+cpu and matching torch_npu, then build vllm with VLLM_TARGET_DEVICE=empty, vllm-ascend with --no-deps --no-build-isolation, and vllm-omni with VLLM_OMNI_TARGET_DEVICE=npu. Finish by building mindiesd with triton-ascend 3.2.1.

Why does importing lightx2v fail with ERR99999 on Ascend?▼

The error occurs because PLATFORM=ascend_npu was not exported before importing lightx2v, so device initialization follows the default platform path. Export PLATFORM=ascend_npu first; it is an environment configuration issue, not a code problem.

Should I download model weights from ModelScope or HuggingFace?▼

ModelScope is the default because it is reachable domestically without a proxy and its mirrors of HF gated repos usually need no authentication. Use HuggingFace with a token only when ModelScope lacks the repo or a specific HF revision is required.

Why does vllm multi-card startup fail with hcclCommInitRootInfoConfig error code 4?▼

The container is missing the HCCL ranktable directory. Mount or copy /usr/local/Ascend/driver/topo into the container at the same path, either at docker run time or afterward with docker cp.

What are the limits of this environment installation workflow?▼

It stops at installation completion: successful mindiesd import, matched versions, and weights in place. Service startup, feature enablement, performance tier selection, accuracy judgment, and SSH transport troubleshooting are delegated to separate skills.

Related Skills

Explore More Software Engineering Skills→
djcyphersdjcyphers

remote-management

Execute remote commands on Windows servers via WinRM, PowerShell remoting, and SSH.

Community
Advanced
TorchedHatTorchedHat
6

vllm-compile

Diagnoses and explains JetBrains Odyssey Odyssey Odyssey Odyssey Odyssey Odyssey Odyssey Odyssey Odyssey Odyssey Odyssey issues.

Official
Advanced
sqjiansqjian

generate-profile

Launch an SGLang server, validate accuracy, and generate a Chrome/Perfetto trace.

Community
Advanced
NVIDIANVIDIA
2.8k

exec-local-compile

Build and install TensorRT-LLM from source in a Docker container with visible GPUs.

Official
Intermediate