2023 H2 Roadmap

In 2023 H2, ByteIR will focus on PyTorch 2.0 supports, highlighting transparent optimizations for GPUs, and shift to more community-driven development.

Highlight features

Frontend/Torch

• PyTorch 2.0 compilation with ByteIR backend (using single GPU)
• PyTorch 2.0 inference demos and tutorials, highlighting LLMs
• PyTorch 2.0 training demos and tutorials, highlighting LLMs

Compiler

• Enhance Linalg optimizations, including elementwise fusion, and tile-and-fuse, highlighting supporting multiple attention variants
• Enhance CAT backend (AI Template) experience, such compilation/autotuning speed and configuration
• Enhance GPU backend optimizations, including elementwise, broadcast, reduction, and their combination

Cross-repo

• ByteIR x ByteMLPerf initiate