What's Changed

• Fix Llava Next Reset Parameters by @alex-jw-brooks in #504
• Multimodal Mistral by @alex-jw-brooks in #502
• 🐛 Fix vision features attention calculation by @gkumbhat in #507

New Contributors

• @gkumbhat made their first contribution in #507

Full Changelog: v1.6.0...v1.7.0

What's Changed

• expansion adaper to allow language model for llava-next by @rzbhatti in #486
• ModelConfig.updated(**kwargs) method to support mulimodel config by @rzbhatti in #484
• Fixed the _hf_to_fms_rope by @rzbhatti in #489
• Granite-vision / llava_next fix for AIU: explicit call embedding layer by @sahilsuneja1 in #490
• [doc] MultiHeadAttention: add kvheads to docstring by @yannicks1 in #487
• Fixed paged implementation to include cuda by @JRosenkranz in #481
• Update training scripts to new FMS conventions by @ani300 in #473
• Fix Reloaded Model Config Types for Composite Models by @alex-jw-brooks in #492
• Add Tests for FMS Model Kwarg Correctness by @alex-jw-brooks in #493
• Refactor HF Config -> FMS Model Kwarg Building by @alex-jw-brooks in #494
• Added granite_v4.py to support 8b model by @rzbhatti in #497
• Fix Parallel Config for Nested Models in Llava Next by @alex-jw-brooks in #500

New Contributors

• @yannicks1 made their first contribution in #487
• @alex-jw-brooks made their first contribution in #492

Full Changelog: v1.5.0...v1.6.0

What's Changed

• Add Roberta for classification to FMS by @ani300 in #466
• Add BERT support to FMS by @ani300 in #467
• updated transformers to 4.55.4 and fixed bug where fms expected cache to be None by @JRosenkranz in #474
• Granite 2b & 3b expand K Q V and Dense to head dim 128 to compile for AIU by @rzbhatti in #472
• Chunked prefill support for paged attention by @ani300 in #463
• head_dim expansion managed at get_model() level by @rzbhatti in #476

Full Changelog: v1.4.0...v1.5.0

What's Changed

• Add the ability to get the last_n_tokens from forward by @JRosenkranz in #470

Full Changelog: v1.3.0...v1.4.0

What's Changed

• updated version to 1.2.1 by @JRosenkranz in #458
• mpnet first version by @ppnaik1890 in #432
• fix: n_groups expansion by @garrett361 in #462
• automate gha build and publish workflow by @tedhtchang in #460
• Fix license errors with latest setuptools by @ani300 in #469
• Refactor Llama model to use same structure as all other models by @ani300 in #465

New Contributors

• @ppnaik1890 made their first contribution in #432
• @garrett361 made their first contribution in #462
• @tedhtchang made their first contribution in #460

Full Changelog: v1.2.1...v1.3.0

What's Changed

• Change how quantized layers are sharded during Tensor Parallel to support FP8 and other cases by @ani300 in #457

Full Changelog: v1.2.0...v1.2.1

What's Changed

• [utils] add option to pad right on pad_input_ids by @kcirred in #426
• Verification for unused layers to remove warnings by @flaviabeo in #414
• Adding siglip vision model by @sahilsuneja1 in #405
• Initial FP8 linear support for FMS by @ani300 in #415
• Fix granite adapter by @ani300 in #440
• Update init.py by @prashantgupta24 in #424
• Add Tie Embeddings in GPTBigCode by @fabianlim in #431
• Tekken tokenizer class added to support the Mistral models by @rzbhatti in #434
• Fix gptq init after FP8 updates by @ani300 in #442
• [tokenizers] deprecation warning by @kcirred in #429
• [typing_extensions] for backwards compatibility by @kcirred in #441
• Devstral small 2505 to fms by @rzbhatti in #427
• Adding llava_next model by @sahilsuneja1 in #420

New Contributors

• @flaviabeo made their first contribution in #414
• @prashantgupta24 made their first contribution in #424
• @fabianlim made their first contribution in #431

Full Changelog: v1.1.0...v1.2.0

What's Changed

• [tokenizer] enable new methods/attributes for tokenizer that are commonly used by @kcirred in #409
• [hf_adapter] added GenerationMixin for transformers version compatibility by @kcirred in #418
• fix get_signature so that it no longer ignores some optional params by @JRosenkranz in #413
• Enable Granite HF Adapter by @kcirred in #402
• Don't set cache_dir in snapshot_download by @tjohnson31415 in #417
• Delay the eos-based loop break to at least get one timing in by @ani300 in #422
• fixed errors in llama and gpt_bigcode docstring by @Zephyr271828 in #390
• paged attention implementation using attn_kwargs by @JRosenkranz in #411

New Contributors

• @kcirred made their first contribution in #409
• @tjohnson31415 made their first contribution in #417
• @Zephyr271828 made their first contribution in #390

Full Changelog: v1.0.0...v1.1.0

What's Changed

Bug Fixes

• fixed bug where when hf_configured set as architecture in get_model, weights get downloaded by @JRosenkranz in #373
• Fix issue where CICD fails occasionally on a test case by @JRosenkranz in #378
• Fixed mypy failure in CI by @JRosenkranz in #381
• fix some bugs in local/test training by @nairbv in #393
• Fix performance when use_contiguous_cache=True by @ani300 in #389
• Fix extra graph generation during generate with contiguous cache by @ani300 in #397
• fixed bug where expectation tests were being skipped by @JRosenkranz in #410

Changes

• Change HF default checkpoint from bin to safetensors by @ani300 in #377
• Support for Granite GPTQ model weight adaptation from HF by @JRosenkranz in #375
• Roberta Question-Answering by @andrea-fasoli in #379
• Add code of conduct by @spzala in #374
• Select linear layer based on module_name by @andrea-fasoli in #369
• Update dependencies for FMS to a more modern stack by @ani300 in #380
• Bamba Model Support by @JRosenkranz in #372
• Added option to specify model name in the model consistency test suite expectation file path by @JRosenkranz in #383
• Fixed issues inferring Roberta QA as well as added features for encoder-only testing by @JRosenkranz in #386
• Update TP and MoE kernels with modern pytorch constructs by @ani300 in #382
• Add fixes for full dynamic with masks by @ani300 in #387
• Add support for fms_mo-based INT8 Granite by @andrea-fasoli in #391
• Add a carve-out for Bamba SSM layers in contiguous check by @ani300 in #394
• Flexible model inputs in generate by @JRosenkranz in #388
• added optional_params/input_ids option to consistency expectation testing by @JRosenkranz in #398
• mistralai/Mistral-7B-Instruct-v0.3 model ported to fms by @rzbhatti in #395
• Add handling of token_type_ids to RoBERTa by @andrea-fasoli in #399
• Raise exception in case of shape mismatch during ckpt loading by @andrea-fasoli in #400
• Add Rope implementations and corrections for llama 3 and llama 3.1 by @ani300 in #385
• Inject custom attention op into MultiHeadAttention by @JRosenkranz in #408

New Contributors

• @rzbhatti made their first contribution in #395

Full Changelog: v0.0.8...v1.0.0

What's Changed

• Enable linear layer selection and weight unfusion in RoBERTa by @andrea-fasoli in #358
• Fix Granite 3.0 TP by @ani300 in #360
• GPTQ CPU implementation by @JRosenkranz in #359
• Support for delayed initialization of modules by @ani300 in #365

New Contributors

• @andrea-fasoli made their first contribution in #313

Full Changelog: v0.0.7...v0.0.8