deepseek-ai/EPLB

Health v2 · maintenance only

Flags: no_releases

Adoption not part of the score

1424 stars · 209 forks observed · 2026-08-28

What it is AI-extracted, prompt v1, taxonomy v1, 2026-08-30, confidence not recorded

EPLB is DeepSeek's open-source Expert Parallelism Load Balancer for Mixture-of-Experts models. It computes balanced expert replication and placement plans across GPUs and nodes using redundant-expert strategies, with hierarchical and global load balancing policies.

Use cases

• balance expert load across GPUs in MoE training
• place duplicated experts for expert parallelism in DeepSeek-V3 style models
• compute expert replication plans from estimated expert loads
• reduce inter-node traffic with group-limited expert routing
• balance MoE experts separately for prefilling and decoding stages

When to choose

• you are running MoE models with expert parallelism and uneven expert loads
• you need a heuristic replication and placement plan for experts across nodes and GPUs
• you use group-limited expert routing and want groups co-located on nodes

When to avoid

• you need automatic expert load prediction - this repo expects load estimates as input
• you are not using mixture-of-experts or expert parallelism
• you need a full distributed training framework rather than a placement algorithm

Facets

library · maturity active

machine-learning llm-inference llm-training gpu-computing scheduling machine-learning deep-learning large-language-models gpu-computing microservices python cross-platform mixture-of-experts expert-parallelism load-balancing moe deepseek gpu linux

Member repositories

For agents

markdown · JSON · MCP: product_card(name="deepseek-ai/EPLB")