Skip to content

Torq Framework version 2.2.0

Choose a tag to compare

View all tags
@ayounes-synaptics ayounes-synaptics released this 19 Sep 14:33
· 4 commits to v2.2 since this release
v2.2.0
e508cd5
This commit was created on GitHub.com and signed with GitHub’s verified signature.
GPG key ID: B5690EEEBB952194
Verified
Learn about vigilant mode.

v2.2.0 is a major release of the Torq-Compiler with changes all other the places; more performance, more coverage, more stability, more testing and more tools. This version is still based on IREE 3.10.

Highlights

  • New torq-lab / torq lab unified model workflow CLI, shipped as a wheel with a pytest adapter and CI coverage.
  • New torq-decode-llm runtime driver: model-agnostic autoregressive decode for LLMs.
  • torq-gen-config is now a standalone CLI shipped in the compiler wheel, plus a new torq-mlir-query tool.
  • Broader ONNX quantized coverage: DynamicQuantizeLinear (per-tensor, dynamic, uint8) and integer MatMul via MatMulInteger + dequantization.
  • LayerNorm now runs on the NSS through bf16 stash retargeting.
  • Large-input tiling, size-aware matmul tiling, program deduplication and NSS program packing in XRAM reduce memory pressure and code size.
  • QEMU simulation replaced by MPACT; kernel-module dependencies upgraded to Astra 2.5.
  • NPU devfreq is boosted during active jobs on SL2610.

Extented release notes

Full Changelog

Assets 10
Loading