Articles by tag: triton

ROCm is catching up to CUDA in AI: OneROCm and Triton

AMD unified ROCm into OneROCm for all accelerators. Triton writes kernels for AMD/Nvidia. vLLM inference on par with CUDA. Six-week releases. Test the stack on Strix Halo.

ruGPT3XL with 8k context: sparse attention

Expansion of ruGPT3XL to 8k tokens via sparse attention and fine-tuning. PPL 13.00 on gazeta, Triton/SDPA optimization ×1.85. Instructions for developers.

GPU acceleration of the Game of Life: CUDA Triton benchmarks

Comparison of PyTorch, CUDA, Triton for the cellular automaton "Game of Life" on A40. Results 22.5 ms, kernel code. Study optimizations for memory-bound tasks. Read the analysis.