[Roadmap] Release Plan of tilescale 0.1.0

[chengyupku]

Release Plan for v0.1.0

Features

• Intra-node UVA copy
  • unrolled copy primitive on device side @chengyupku
  • create IPC handle and gather on host side @chengyupku
• Barrier and memory fence
  • barrier primitives (signal, arrive)
  • group barrier, implemented by barrier primitives
  • memory fence primitives
• Resource control
  • persistent threadblock specialization
  • multi-stream specialization
• Language
  • T.alloc operation, e.g., T.alloc(scope=”system”, level=”L3”)
  • T.view operation, e.g., T.view(scope=”device”, layout=T.FullRow)

---

Kernels

• DeepEP
  • Intra-node
  • Inter-node
• AFD
• Other patterns
  • all-to-all
  • all-reduce
  • ag-gemm
  • gemm-rs
  • Cannon
  • SUMMA