v0.12.1 Patch Release

[nickfraser]

Issue to track what we plan to include in the next patch release. The "wishlist" contains items that we would like to include if we have the time. Keep in mind, plans and priorities can always change, as such, the list below is subject to change.

Planned Additions:

• Benchmark YAML updates for SDXL, ImageNet Feat (examples): refactor imagenet and stable_diffusion entrypoints #1281
• Basic end-to-end tests for SDXL, ImageNet entry-point Feat (ex): tests Stable Diffusion and ImageNet #1339
  • SDXL attention quantization test
• Initial Qronos support Feat (qronos): initial implementation of Qronos  #1311
  • Blog post + documentation Docs (qronos): adding docs and configs #1326
• R2 rotation equalization fix to replace the current workaround Feat (graph/rotate): improve R2 region in SDPA #1310
  • R2 follow up Fix (graph/rotation): rotation on subset of channels for SDPA #1312
• Expansion paper yamls Expansion paper configs #1327
• Expansion flag Feat (brevitas_examples/llm): configurable expansion step #1280
• Programmatic quantization for FINN new example (finn): programmatic quantization & PTQ of MNv2 for FINN #1283
• Further attention quantization options in LLM example feat (ex/llm): fully parametrise attention quantization #1287
• Fix stochastic rounding device Stochastic rounding breaks on GPU #1294
• Scaled min-max quantization Feat (scaling): rescaled min-max scaling and zero point #1320
• QONNX: add flag to generate IntQuant, not Quant (pending confirmation from QONNX team) Not needed
• Feat (core): Remove assumptions on positiveness of scales #1308
• Getting started link broken #1336

Wishlist:

• QONNX export with dynamo Fix (export/qonnx): Add export support with dynamo=True #1234
• ONNX + MX export
• SHARK export Sharkitas #1233
• Initial GGUF export Feat (brevitas_examples/llm): GGUF export #1291
• Feat (graph): Minor refactoring layerwise_layer_handler #1335
• Fix cross-dependency on LLM from stable diffusion
• Create utility which fixes seeds for all tests
  • Switch from per-file to per-test setting of seeds for better reproducibility