Issues

• RFC: io_uring DMA-BUF copier, NVMe straight into GPU memory

  foundation-model-stack/fastsafetensors#139 · Filter by author mcgrof opened on Nov 4, 2026
  2

• [Discussion] Loader building blocks for row-gathered tables (Engram, PLE), following #127

  foundation-model-stack/fastsafetensors#132 · Filter by author gitbisector opened on Oct 11, 2026

• Expert-parallel shuffle OOMs on large fused MoE tensors (full-tensor on-device clone in shuffle())

  foundation-model-stack/fastsafetensors#127 · Filter by author igor-susic1 opened on Oct 6, 2026
  4

• feat: checkpoint writing and Spyre support

  foundation-model-stack/fastsafetensors#118 · Filter by author takeshi-yoshimura opened on Sep 30, 2026
  1

• SafeTensorsFileLoader retains ~2x file size in memory permanently after close(), even with nogds=True + explicit gc/empty_cache

  foundation-model-stack/fastsafetensors#94 · Filter by author ZackSample opened on Jul 23, 2026
  3

  

• [Discussion] Sub-file batching — Decouple GPU memory usage from model file size

  foundation-model-stack/fastsafetensors#71 · Filter by author ABNER-1 opened on Apr 27, 2026
  21