# DeepEP | DeepSeek | Knowledge Base | Instagit

DeepEP: an efficient expert-parallel communication library

GitHub Stars: 9.4k

Repository: https://github.com/deepseek-ai/DeepEP

---

## Articles

### [How DeepEP Optimizes Handle Reuse for Repeated Dispatches](/deepseek-ai/DeepEP/how-does-deepep-optimize-handle-reuse-for-repeated-dispatch)

DeepEP optimizes handle reuse for repeated dispatches by caching communication layout metadata. Improve performance by eliminating redundant calculations and GPU launches.

- Tags: performance
- Published: 2026-04-25

### [DeepEP topk_idx_t Custom Dtype Requirements: Implementation Guide](/deepseek-ai/DeepEP/what-are-the-requirements-for-deepep-topk-idx-t-custom-dtype)

Learn the requirements for DeepEP topk_idx_t custom dtype. Implement efficiently with torch.int32 or torch.int64 based on TOPK_IDX_BITS. Get the guide now.

- Tags: how-to-guide
- Published: 2026-04-25

### [Does DeepEP Support Multi-Node NVLink (MNNVL)? Configuration and Usage Guide](/deepseek-ai/DeepEP/does-deepep-support-mnnvl)

DeepEP enables Multi-Node NVLink (MNNVL) support. Learn how to configure and use MNNVL for enhanced distributed training performance via the allow_mnnvl flag or command-line option.

- Tags: how-to-guide
- Published: 2026-04-25

### [Safety Concerns with DeepEP Undefined Behavior PTX Instructions](/deepseek-ai/DeepEP/what-are-the-safety-concerns-with-deepep-undefined-behavior-ptx-instructions)

Explore the safety concerns of DeepEP undefined behavior PTX instructions. Learn how non-standard instructions can lead to incorrect results on different GPU architectures.

- Tags: deep-dive
- Published: 2026-04-25

### [How to Install and Configure the DeepEP NVSHMEM Dependency](/deepseek-ai/DeepEP/how-to-install-and-configure-deepep-nvshmem-dependency)

Learn how to install and configure DeepEP NVSHMEM dependency. Follow steps for NVSHMEM v3.3.9+, set NVSHMEM_DIR, and enable GPUDirect Async for faster internode communication in DeepEP.

- Tags: how-to-guide
- Published: 2026-04-25

### [DeepEP Adaptive Routing vs Static Routing: Performance Differences and Configuration Guide](/deepseek-ai/DeepEP/what-is-the-performance-difference-between-deepep-adaptive-routing-vs-static-routing)

Compare DeepEP adaptive routing vs static routing performance. Explore latency differences and configuration for optimal network traffic management. Discover which fits your needs.

- Tags: performance
- Published: 2026-04-25

### [How DeepEP Achieves Traffic Isolation with InfiniBand Virtual Lanes](/deepseek-ai/DeepEP/how-does-deepep-achieve-traffic-isolation-with-infini-band-virtual-lanes)

Discover how DeepEP uses InfiniBand Virtual Lanes and NVSHMEM to isolate traffic, preventing latency-sensitive MoE kernels from contending with bandwidth-heavy operations.

- Tags: internals
- Published: 2026-04-25

### [How DeepEP Utilizes TMA Instructions for Minimal SM Usage on Hopper GPUs](/deepseek-ai/DeepEP/how-does-deepep-utilize-tma-instructions-for-minimal-sm-usage)

Discover how DeepEP leverages TMA instructions for minimal SM usage on Hopper GPUs. Offloads tensor movement to TMA, achieving high bandwidth with fewer SMs.

- Tags: deep-dive
- Published: 2026-04-25

### [How DeepEP Uses Zero‑Copy Optimization to Reduce SM Usage](/deepseek-ai/DeepEP/how-does-deepep-use-zero-copy-optimization-to-reduce-sm-usage)

Discover how DeepEP s zero copy optimization slashes shared memory usage by reading directly from RDMA buffers, boosting GPU occupancy and kernel efficiency. Learn more today.

- Tags: deep-dive
- Published: 2026-04-25

### [DeepEP Shrink Mode for Dynamic Rank Masking: Runtime Fault Tolerance in Expert Parallelism](/deepseek-ai/DeepEP/what-is-deepeps-shrink-mode-for-dynamic-rank-masking)

Explore DeepEP's shrink mode for dynamic rank masking. Temporarily disable ranks during expert-parallel communication at runtime without communicator teardown. Enhance fault tolerance.

- Tags: internals
- Published: 2026-04-25

### [How to Configure DeepEP Low-Latency Mode QP (Queue Pair) for RDMA Optimization](/deepseek-ai/DeepEP/how-to-configure-deepep-low-latency-mode-qp)

Optimize RDMA performance by configuring DeepEP low-latency mode QP. Learn how to set num_qps_per_rank for improved buffer creation and faster communication.

- Tags: how-to-guide
- Published: 2026-04-25

### [How DeepEP Enables Hook-Based Overlapping Without SM Occupation](/deepseek-ai/DeepEP/how-does-deepep-enable-hook-based-overlapping-without-sm-occupation)

Discover how DeepEP enables hook-based overlapping without SM occupation by launching low-latency kernels and splitting operations into distinct phases. Learn more today!

- Tags: internals
- Published: 2026-04-25

### [DeepEP Intranode and Internode Kernels: Architecture and Performance Differences](/deepseek-ai/DeepEP/what-are-the-differences-between-deepep-intranode-and-internode-kernels)

Explore DeepEP intranode vs internode kernels. Understand their distinct NVLink and RDMA communication patterns and synchronization strategies for efficient distributed computing.

- Tags: architecture
- Published: 2026-04-25

### [How to Set Up DeepEP Auto-Tuning for Different Cluster Sizes](/deepseek-ai/DeepEP/how-to-set-up-deepep-auto-tuning-for-different-cluster-sizes)

Learn to set up DeepEP auto-tuning for various cluster sizes. Optimize dispatch and kernel performance by re-tuning its static configuration map for your specific hardware.

- Tags: how-to-guide
- Published: 2026-04-25

### [DeepEP CUDA Graph Compatibility: Limitations and Workarounds](/deepseek-ai/DeepEP/what-are-the-limitations-and-workarounds-for-deepep-cuda-graph-compatibility)

Explore DeepEP CUDA graph compatibility limitations. Discover workarounds for intranode NVLink operations, CPU sync, host tensor ops, and more in this technical guide.

- Tags: deep-dive
- Published: 2026-04-25

### [How to Configure IBGDA in DeepEP: Complete Setup Guide for InfiniBand GPU Direct Access](/deepseek-ai/DeepEP/how-to-configure-ibgda-in-deepep)

Configure IBGDA in DeepEP with this guide. Enable NVIDIA driver options, install NVSHMEM, and optimize your setup for low-latency GPU Direct Access. Learn how now.

- Tags: how-to-guide
- Published: 2026-04-25

### [DeepEP FP8 Dispatch and Combine Operations: Low-Precision All-to-All for MoE](/deepseek-ai/DeepEP/what-are-deepeps-fp8-dispatch-and-combine-operations)

Discover DeepEP's FP8 dispatch and combine operations for efficient all-to-all communication in MoE models. Reduce bandwidth by 4x with minimal accuracy loss.

- Tags: deep-dive
- Published: 2026-04-25

### [How DeepEP Controls Streaming Multiprocessors (SM) for Kernel Execution](/deepseek-ai/DeepEP/how-does-deepep-control-sm-for-kernel-execution)

Learn how DeepEP manages SMs for kernel execution by controlling the num_sms parameter. Discover how this impacts grid dimensions and optimizes performance.

- Tags: internals
- Published: 2026-04-25

### [How DeepEP Achieves Communication-Computation Overlapping with EventOverlap](/deepseek-ai/DeepEP/how-does-deepep-achieve-communication-computation-overlapping-with-eventoverlap)

DeepEP overlaps communication and computation using EventOverlap. Discover how it captures completion signals and defers synchronization for efficient deep learning.

- Tags: internals
- Published: 2026-04-25

### [DeepEP Buffer Class API for Expert Parallel Communication: Architecture and Usage](/deepseek-ai/DeepEP/what-is-the-deepep-buffer-class-api-for-expert-parallel-communication)

Explore the DeepEP Buffer class API for efficient expert parallel communication. This Python interface unifies CUDA NVLink and RDMA resources for low latency.

- Tags: api-reference
- Published: 2026-04-25

### [How to Configure NVLink and RDMA in DeepEP for Hybrid Clusters](/deepseek-ai/DeepEP/how-to-configure-nvlink-and-rdma-in-deepep-for-hybrid-clusters)

Configure NVLink and RDMA in DeepEP for hybrid clusters. Learn to validate GPU topology, set NVSHMEM variables, and initialize Buffers for low latency and asymmetric forwarding. Optimize deep learning performance now.

- Tags: how-to-guide
- Published: 2026-04-25

### [DeepEP Low‑Latency Dispatch and Combine Kernels for Inference Decoding](/deepseek-ai/DeepEP/what-are-deepeps-low-latency-dispatch-and-combine-kernels-for-inference-decoding)

Discover DeepEP's low latency dispatch and combine kernels for MoE inference decoding. Achieve sub millisecond token routing with RDMA and C++ runtime optimizations. Learn more.

- Tags: deep-dive
- Published: 2026-04-25

### [How DeepEP Manages MoE Expert Parallelism Communication in DeepSeek](/deepseek-ai/DeepEP/how-does-deepep-manage-moe-expert-parallelism-communication)

Discover how DeepEP manages MoE expert parallelism communication using automatic NVLink and RDMA routing for efficient, high-throughput training and low-latency inference.

- Tags: deep-dive
- Published: 2026-04-25