Release v0.12.1

Highlights

• New / Updated PTQ Algorithms:
  • Qronos support #1311
  • Fixes / improvements to rotation equalization #1310, #1312
  • DDP-like bias correction for SDXL (experimental) #1342
• Improved layer support:
  • Quantization of SDPA without FX #1299
• New export flows:
  • Initial SHARK Export support: #1300
  • Initial GGUF Export: #1291
• Improved examples:
  • Qronos examples (paper) #1326
  • "Benchmark" experiments for stable diffusion, imagenet #1281
  • Post-training model expansion examples (paper) #1355
• Allow signed scales #1308
• QONNX export with dynamo=True #1234

What's Changed

• Fix (setup): solve incompatibility between isort and yapf by @Giuseppe5 in #1293
• Feat (graph/hadamard): 152 had support by @Giuseppe5 in #1295
• feat (ex/llm): fully parametrise attention quantization by @nickfraser in #1287
• Feat (ex/llm): "auto" dtype by @pablomlago in #1301
• Setup: update transformers version by @Giuseppe5 in #1304
• Feat (brevitas_examples/llm): quant SDPA without FX by @Giuseppe5 in #1299
• Fix (brevitas_examples/llm): update README and yaml by @Giuseppe5 in #1305
• Setup: temporary pin pytest by @Giuseppe5 in #1307
• Feat (brevitas_examples/llm): configurable expansion step by @Giuseppe5 in #1280
• Versioning support in documentation by @Giuseppe5 in #1298
• Feat (graph/rotate): improve R2 region in SDPA by @Giuseppe5 in #1310
• Docs: improve docs build by @Giuseppe5 in #1314
• Fix (graph/rotation): rotation on subset of channels for SDPA by @Giuseppe5 in #1312
• Feat (brevitas_examples/llm): GGUF export by @Giuseppe5 in #1291
• Setup: bump torch version by @Giuseppe5 in #1205
• Fix (ex/imagenet): add forward pass in imagenet ptq example by @Giuseppe5 in #1316
• Feat (qronos): initial implementation of Qronos by @i-colbert in #1311
• Fix (graph/equalize): correct class check during rotation merging by @Giuseppe5 in #1317
• docs (core): Typo fix in docstring by @nickfraser in #1318
• Fix (llm): removing duplicate set_seed function by @i-colbert in #1319
• Fix (ex/common): save scales during optimization by @pablomlago in #1313
• Fix (llm): compatibility with non-uniform RMSNorm shapes by @i-colbert in #1324
• Fix (graph/gpxq): fix memory leak with weight_orig by @Giuseppe5 in #1325
• Shark LLM export by @Giuseppe5 in #1300
• Feat (scaling): rescaled min-max scaling and zero point by @i-colbert in #1320
• Fix (gguf): derived modify_tensors() returns generator by @i-colbert in #1329
• Fix (gpxq): device management of weight_orig for GPxQ by @i-colbert in #1330
• Fix (gguf): resolving zero point permutation issue with LlamaModel by @i-colbert in #1332
• Feat (examples): refactor imagenet and stable_diffusion entrypoints by @pablomlago in #1281
• Feat: skipping rotation optimization with load_checkpoint by @i-colbert in #1331
• Feat (core): Remove assumptions on positiveness of scales by @pablomlago in #1308
• Fix (export/qonnx): Add export support with dynamo=True by @nickfraser in #1234
• Fix (core/ops_ste): preserve dtype during clamp by @Giuseppe5 in #1340
• Fix (eq/rotation): find value if passed as kwarg by @nickfraser in #1338
• Feat (ex): tests Stable Diffusion and ImageNet by @pablomlago in #1339
• Feat (ex/sdxl): DDP-like bias correction for SDXL by @pablomlago in #1342
• Feat (graph): Minor refactoring layerwise_layer_handler by @pablomlago in #1335
• Fix (torch_utils): remove deprecated functions by @Giuseppe5 in #1344
• Fix (setup): test against latest 2.1 torch by @Giuseppe5 in #1348
• Rotation fix by @Giuseppe5 in #1334
• Docs (qronos): adding docs and configs by @i-colbert in #1326
• Feat: Added ONNX export to BNN-PYNQ example by @nickfraser in #916
• Fix (deps): set accelerate<1.10 by @nickfraser in #1352
• Fix (copyright): Fix some missing copyright headers by @nickfraser in #1353
• Fix (ex/llm/benchmark): import error by @nickfraser in #1356
• Setup: temporarily pin diffusers version by @Giuseppe5 in #1360
• Fix (core/scaling): handle edge cases with signed scale by @Giuseppe5 in #1347
• Fix(core/stochastic_round): adjust stochastic round device by @Giuseppe5 in #1359
• Feat (papers): expansion paper configs by @Giuseppe5 in #1355
• Fix (docs): correct link to docs in initial README by @Giuseppe5 in #1361
• Fix (setup.py): Update author and contact in setup.py by @nickfraser in #1365
• Docs (README): Update maximum PyTorch version by @nickfraser in #1366
• Docs (getting started): Typo fix in example by @nickfraser in #1367
• requirements: Updated PyTorch, python versions by @nickfraser in #1370
• Feat (core/scaling): Add option to restrict the output of (scale / threshold) by @nickfraser in #1369
• deps (ex) update accelerate version by @nickfraser in #1371

Full Changelog: v0.12.0...v0.12.1