ACCELERATING MOES WITH TRITON GROUPED GEMM

Summary from the TLDR AI newsletter

PyTorch has introduced a cache-aware Triton BF16 Grouped GEMM kernel optimized for Mixture-of-Experts models like DeepSeekv3. The implementation offers up to 2.6× speedup over baseline PyTorch loops by batching independent GEMMs into a single kernel call.