skills.rest
yangwhale
yangwhaleCOMMUNITY
Explore all skills by this author →
GitHub

deepep-installer

Install, configure, and troubleshoot DeepEP on NVIDIA GPU systems.

9|1|Updated Oct 20, 2025
Search Tags:#CUDA#GPU-NIC mapping#GPU installation#deepep#NVSHMEM#DeepEP#DOCA-OFED
One-click install
npx skills add https://github.com/yangwhale/gpu-tpu-pedia --skill deepep-installer
Or copy as Structured Prompt for Agent▼
Please help me install this Agent Skill.
Skill: deepep-installer
Source: https://github.com/yangwhale/gpu-tpu-pedia/tree/main/VibeCoding/claude-code/skills/deepep-installer
Command: npx skills add https://github.com/yangwhale/gpu-tpu-pedia --skill deepep-installer

SYSTEM DOCUMENTATION & REQUIREMENTS

💡 This Skill requires cuda-toolkit-12-9, doca-ofed, cmake, ninja-build, python3-venv, python3-pip, python3.12-dev, build-essential, git, curl, and includes scripts (resource) components.

What problem does it solve?

This Skill helps users install, configure, and troubleshoot DeepEP (DeepSeek Expert Parallelism) on NVIDIA GPU systems, covering end-to-end setup of CUDA, DOCA-OFED, NVSHMEM with IBGDA support, and the DeepEP library itself, including debugging workflows for common failures.

Core Features & Use Cases

  • End-to-end installation: Guides users through CUDA Toolkit, DOCA-OFED, NVSHMEM IBGDA, and DeepEP installation.
  • GPU compatibility and networking: Optimized for B200/H100/A100 GPUs with RoCE/InfiniBand networking, enabling high-performance MoE communication.
  • Troubleshooting: Provides structured steps to diagnose and fix common installation failures, including NIC mapping issues.
  • Use Case: A data center engineer can deploy a production DeepEP environment on a cluster with 8 GPUs per node and 2 nodes, ensuring proper NIC-to-GPU mapping.

Quick Start

Run the installer script to begin Phase 1 (CUDA + DOCA + PeerMappingOverride) and reboot as prompted. After reboot, run Phase 2 to install NVSHMEM (without GDRCopy), PyTorch, and DeepEP with PR #466 GPU-to-NIC mapping. Source the generated environment script and verify that importing DeepEP succeeds.

Frequently Asked Questions about deepep-installer

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

FAQPage Schema
How do I install DeepEP on NVIDIA H100 GPUs with InfiniBand networking?▼

To install DeepEP on NVIDIA H100 GPUs with InfiniBand, you need to set up CUDA Toolkit 12.9, DOCA-OFED 3.2.1, NVSHMEM with IBGDA support, and build the DeepEP library from source with a specific GPU-to-NIC mapping patch.

What is the correct GPU-to-NIC mapping patch for DeepEP deployment?▼

The correct GPU-to-NIC mapping for DeepEP deployment is applied via PR #466 during the source build, ensuring proper communication routing between multi-node GPUs and network interfaces.

Why does my NVSHMEM IBGDA installation fail when building DeepEP from source?▼

NVSHMEM IBGDA installation failures during DeepEP builds often stem from missing dependencies or incorrect configuration, requiring structured troubleshooting to verify CUDA, DOCA-OFED, and NVSHMEM v3.5.19-1 setups.

Does DeepEP installation work without GDRCopy for NVSHMEM?▼

Yes, DeepEP installation configures NVSHMEM explicitly without GDRCopy, building the library from source while relying on IBGDA support for high-performance MoE communication across B200, H100, or A100 GPUs.

Can I use DeepEP with PyTorch 2.9.1 and CUDA Toolkit 12.9?▼

Yes, DeepEP is fully compatible with PyTorch 2.9.1+cu129 and requires CUDA Toolkit 12.9, installing both during the second phase of the setup process after configuring DOCA-OFED and rebooting.

What are the limitations when troubleshooting DeepEP network mapping on RoCE?▼

Troubleshooting DeepEP network mapping on RoCE requires ensuring proper PeerMappingOverride configurations and exact NVSHMEM versions, as incorrect GPU-to-NIC mapping can severely degrade multi-node MoE communication performance.

Related Skills

Explore More Software Engineering Skills→
igamenovoerigamenovoer
7

pixi-install-nvidia

Install NVIDIA CUDA tooling in a user-space Pixi environment without sudo.

Community
Intermediate
yo-stevenyo-steven

perf-moe-comm-overlap

Overlap expert-parallel dispatch and combine communication with expert compute in Megatron-Bridge.

Community
Advanced
graphsignalgraphsignal
241

graphsignal-sdk

Instrument Python AI inference workloads with Graphsignal SDK for tracing and profiling.

Official
Advanced
BioKTBioKT
1

gmx-install

Automate GROMACS source installation with PLUMED patching and CUDA/MPI configuration.

Official
Advanced