Model Releases
I released a softmax-free attention model at GPT-2 Medium scale (~354M params, 11.5B tokens): structural sparsity + tile-skipping kernels for long-context VRAM savings. Open weights + custom Triton kernels [R]
DGX agentA researcher released an open-source softmax-free attention model at GPT-2 Medium scale (354M parameters trained on 11.5B tokens) that uses structural sparsity and tile-skipping kernels to reduce VRAM