DeepEP V2 Released

DeepSeek released DeepEP V2, completely restructuring Expert Parallelism.

• It cleans up V1's legacy, operating with fewer SMs while supporting larger scale-up / scale-out domains.
• The backend was switched from NVSHMEM to the lighter NCCL Gin.
• It provides JIT compilation, a high-throughput/low-latency path unified into a single API, EP2048 support, and analytical calculation of SM/QP counts.
• As experimental features, it adds 0 SM Engram, 0 SM Pipeline Parallelism, and 0 SM Context Parallelism all-gather.

In internal benchmarks, under conditions of 8K tokens/batch, 7168 hidden dim, top-8 experts, FP8 dispatch, and BF16 combine, it reduced SM count from 24 to 4-6 while maintaining equal or better performance, reporting up to 1.3x peak performance and up to 4x SM savings.

Alongside the same announcement, the TileKernels repository was also released. It is a TileLang-based GPU kernel library for LLMs, released under the MIT License, and includes MoE routing, quantization (FP8/FP4/E5M6), transpose, Engram, Manifold HyperConnection, and modeling layers.