BLASST: Dynamic BLocked Attention Sparsity via Softmax Thresholding

Documenti analoghi

• SmoothQuant: Accurate and Efficient Post-Training Quantization for Large Language Models
  di: Xiao, Guangxuan, et al.
  Pubblicazione: (2022)
• FlashAttention-4: Algorithm and Kernel Pipelining Co-Design for Asymmetric Hardware Scaling
  di: Zadouri, Ted, et al.
  Pubblicazione: (2026)
• What Does Softmax Probability Tell Us about Classifiers Ranking Across Diverse Test Conditions?
  di: Tu, Weijie, et al.
  Pubblicazione: (2024)
• BLaST: High Performance Inference and Pretraining using BLock Sparse Transformers
  di: Okanovic, Patrik, et al.
  Pubblicazione: (2025)
• Intrinsic and Extrinsic Organized Attention: Softmax Invariance and Network Sparsity
  di: Fasina, Oluwadamilola, et al.
  Pubblicazione: (2025)