7 ms·

Sounds like Multi-Head Latent Attention (MLA) from DeepSeek

• veunes 6mo ago
  Nah, those are completely different beasts. DeepSeek's MLA solves the KV cache issue via low-rank projection - they literally squeeze the matrix through a latent vector at train time. TurboQuant is just Post-Training Quantization where they mathematically compress existing weights and activations using polar coordinates
  • esafak 6mo ago
    No, it is about compressing the KV cache; see How TurboQuant works.