TLDR Tech / DevOps
Deepgemm
DeepGEMM, a new open-source CUDA library released under the MIT License, delivers high-performance tensor core operations for large language models, including FP8/FP4 GEMMs, fused mixture-of-experts, and attention mechanisms, all compiled at runtime without requiring CUDA installation. The lightweight library matches or exceeds the performance of expert-tuned alternatives while maintaining a simpler codebase designed for both production use and learning GPU kernel optimization techniques.
Read the full story