17 articles found
ALTK-Evolve enables your AI agents to retain long-term, on-the-job learning, solving the 'eternal in...
TorchInductor now supports NVIDIA's CuteDSL backend, offering you new avenues for state-of-the-art G...
Generalized Dot-Product Attention delivers up to 2x speedup in GPU training forward pass, hitting 1,...
TorchSpec introduces fully disaggregated inference and training for speculative decoding, enabling y...
Optimize your Mamba-2 SSD modules with a fused Triton kernel for 1.50x-2.51x speedups on NVIDIA A100...