v0.0.27: Qwen2 models, Neuron SDK 2.20.2

What's Changed

• Add support for Qwen2 models (#746)
• bump Neuron SDK to 2.20.2 (#743)
• NeuronX TGI: bump router version to 3.0.0 (#748)

Bug fixes

• training: Fixes consolidation issue when TP is enabled (#739)
• inference: Fix t5 decoder compilation error since Neuron sdk 2.20 (#732)

Full Changelog: v0.0.26...v0.0.27

v0.0.26: T5 parallel support and new NeuronORPOTrainer

What's Changed

Inference

• refactoring Diffusers pipelines (#711)
• Add tensor parallel support to T5 via NxD (#697)

Training

• support resizing embeddings (#670)
• NeuronORPOTrainer (#719)

Bug fixes

• update TGI error message (#659)
• fix errors in vision/audio models docstring (#714)
• fix wrong inputs/model placement when using a single core (#725)
• fix model checkpoint saving issue when using PEFT (#727)
• fix non contiguous tensors in consolidation (#736)

Full Changelog: v0.0.25...v0.0.26

v0.0.25: SFT Trainer, Llama 3.1-3.2, ControlNet, AWS Neuron SDK 2.20

What's Changed

• Use AWS Neuron SDK 2.20 (#696) by @dacorvo
• Bump optimum to 1.22 (#686) by @JingyaHuang
• Bump transformers to 4.43.2 (#665) by @dacorvo

Inference

• Add support for multiple ControlNet (#691) by @JingyaHuang
• Add ControlNet support for SDXL (#675) by @JingyaHuang

Training

• Support SFTTrainer (#682) by @michaelbenayoun
• LoRA finetuning tutorial (#671) by @michaelbenayoun

Full Changelog: v0.0.24...v0.0.25

v0.0.24: PEFT training support, ControlNet, InstructPix2Pix, Audio models, TGI benchmarks

What's Changed

• Use AWS Neuron SDK 2.19.1 by @dacorvo in #661

Training

• Initial PEFT support by @michaelbenayoun in #612
• PEFT + TP support by @michaelbenayoun in #620
• Fix MPMD detected error during training with TP by @michaelbenayoun in #648

Inference

• Add Stable Diffusion ControlNet support by @JingyaHuang in #622
• Add InstructPix2Pix pipeline support. by @asntr in #625
• Add ViT export support and image classification by @JingyaHuang in #616
• Add wav2vec2 support - export and audio tasks modeling by @JingyaHuang in #645
• Add more audio models: ast, hubert, unispeech, unispeech-sat, wavlm by @JingyaHuang in #651

TGI

• Extending TGI benchmarking and documentation by @jimburtoft in #621
• Add support for TGI truncate parameter by @dacorvo in #647

Other changes

• enable unequal height and width by @yahavb in #592
• Skip invalid gen config by @dacorvo in #618
• Deprecate resume_download by @Wauplin in #586
• Remove a line non-intentionally merged by @JingyaHuang in #628
• Add secrets scanning workflow by @mfuntowicz in #631
• fix bad link to distributed-training how-to guide in optimum-neuron docs by @aws-amj in #627
• Do not copy local checkpoint by @dacorvo in #630
• Make neuron_cc_optlevel None by default by @michaelbenayoun in #632
• Remove print by @michaelbenayoun in #633
• Set bf16 to true when needed by @michaelbenayoun in #635
• Fix gradient checkpointing with PEFT by @michaelbenayoun in #634
• Refactor decoder tests by @dacorvo in #641
• CI cache builder by @dacorvo in #642
• Restore optimized attention score for sd15 & fix the generated images quality issue by @JingyaHuang in #646
• Add and remove some mark steps by @michaelbenayoun in #644
• Fix consolidation for TP by @michaelbenayoun in #649
• Fix spelling in error message by @jimburtoft in #656
• Update docs by @michaelbenayoun in #588
• Fixes NxDPPModel for Neuron SDK 2.19 by @michaelbenayoun in #663
• Various fixes for training by @michaelbenayoun in #654
• migrate ci by @XciD in #662
• ci: fix inference cache pipeline by @dacorvo in #667
• broken link by @pagezyhf in #669
• Bump TGI version and fix bugs by @dacorvo in #666

New Contributors

• @mfuntowicz made their first contribution in #631
• @aws-amj made their first contribution in #627
• @asntr made their first contribution in #625
• @XciD made their first contribution in #662

Full Changelog: v0.0.23...v0.0.24

v0.0.23: Bump transformers and optimum version

What's Changed

• bump required packages versions: transformers==4.41.1, accelerate==0.29.2, optimum==1.20.*

Inference

• Fix diffusion caching by @oOraph in #594
• Fix inference latency issue when weights/neff are separated by @JingyaHuang in #584
• Enable caching for inlined models by @JingyaHuang in #604
• Patch attention score far off issue for sd 1.5 by @JingyaHuang in #611

TGI

• Fix excessive CPU memory consumption on TGI startup by @dacorvo in #595
• Avoid clearing all pending requests on early user cancellations by @dacorvo in #609
• Include tokenizer during export and simplify deployment by @dacorvo in #610

Training

• Performance improvements and neuron_parallel_compile and gradient checkpointing fixes by @michaelbenayoun in #602

New Contributors

• @pagezyhf made their first contribution in #601

Full Changelog: v0.0.22...v0.0.23

v0.0.22: Mixtral support, pipeline for sentence transformers, compatibility with Compel

What's Changed

Training

• Integrate new API for saving and loading with neuronx_distributed by @michaelbenayoun in #560

Inference

• Add support for Mixtral by @dacorvo in #569
• Improve Llama models performance by @dacorvo in #587
• Make Stable Diffusion pipelines compatible with compel by @JingyaHuang and @neo in #581 (with tests inspired by the snippets sent from @Suprhimp)
• Add SentenceTransformers support to pipeline for feature-extration by @philschmid in #583
• Allow download subfolder for caching models with subfolder by @JingyaHuang in #566
• Do not split decoder checkpoint files by @dacorvo in #567

TGI

• Set up TGI environment values with the ones used to build the model by @oOraph in #529
• TGI benchmark with llmperf by @dacorvo in #564
• Improve tgi env wrapper for neuron by @oOraph in #589

Caveat

Currently traced models with inline_weights_to_neff=False have higher than expected latency during the inference. This is due to the weights are not automatically moved to Neuron devices. The issue will be fixed in #584, please avoid setting inline_weights_to_neff=False in this release.

Other changes

• Improve installation guide by @JingyaHuang in #559
• upgrade optimum and then install optimum-neuron by @shub-kris in #533
• Cleanup obsolete code by @michaelbenayoun in #555
• Extend TGI integration tests by @dacorvo in #561
• Modify benchmarks by @dacorvo in #563
• Bump PyTorch to 2.1 by @JingyaHuang in #502
• fix(decoder): specify libraryname to suppress warning by @dacorvo in #570
• missing \ in quickstart inference guide by @yahavb in #574
• Use AWS 2.18.0 AMI as base by @dacorvo in #572
• Update TGI router version to 2.0.1 by @dacorvo in #577
• Add guide for LoRA adapters by @JingyaHuang in #582
• eos_token_id can be a list in configs by @dacorvo in #580
• Ease the tests when there is no hf token by @JingyaHuang in #585
• Change inline weights to Neff default value to True by @JingyaHuang in #590

New Contributors

• @yahavb made their first contribution in #574

Full Changelog: v0.0.21...v0.0.22

v0.0.21: Expand caching support for inference, GQA training support, TGI improved performance

What's Changed

Training

• Add GQA optimization for Tensor Parallel training to support the case tp_size > num_key_value_heads by @michaelbenayoun in #498
• Mixed-precision training with both torch_xla or torch.autocast by @michaelbenayoun in #523

Inference

• Add caching support for traced TorchScript models (eg. encoders, stable diffusion models) by @JingyaHuang in #510
• Support phi model on feature-extraction, text-classification, token-classification tasks by @JingyaHuang in #509

TGI

• TGI improvements by @dacorvo in #522

Caveat

AWS Neuron SDK 2.18 doesn't support the compilation of SDXL's unet with weights / neff separation, inline_weights_to_neff=True is forced through:

• Disable weights / neff separation of SDXL's UNET for neuron sdk 2.18 by @JingyaHuang in #554

Other changes

• Fix/ami authorized keys by @shub-kris in #517
• Skip weight load during parallel compile by @michaelbenayoun in #524
• fixing format in getting-started.ipynb by @jimburtoft in #526
• Removing colab links in notebooks.mdx by @jimburtoft in #525
• ADD stale bot by @philschmid in #530
• Bump optimum version by @JingyaHuang in #534
• Fix style by @JingyaHuang in #538
• Fix GQA permutation computation and sequential weight initialization / loading when doing TP by @michaelbenayoun in #531
• Add setup runtime step for K8S by @glegendre01 in #541
• Disable logging during precompilation by @michaelbenayoun in #539
• Do not use deprecated list_files_info by @Wauplin in #536
• Adding link to existing Fine-tuning example in Notebooks by @jimburtoft in #527
• Add missing notebooks to doc by @JingyaHuang in #543
• fix: bug in get_available_cores within container by @oOraph in #546
• Init on the xla device by @michaelbenayoun in #521
• Adding CodeLlama-7B inference and compilation example notebook by @jimburtoft in #549
• Add tools for auto filling traced models cache by @JingyaHuang in #537
• Remove print that should not be there by @michaelbenayoun in #552
• Use AWS Neuron sdk 2.18 by @dacorvo in #547
• Cache utils related cleanup by @michaelbenayoun in #553

New Contributors

• @glegendre01 made their first contribution in #541
• @Wauplin made their first contribution in #536

Full Changelog: v0.0.20...v0.0.21

v0.0.20: Multi-node training, SD Lora, sentence transformers clip, TGI improvements

What's Changed

Training

• Multi-node training support by @michaelbenayoun (#440)

TGI

• optimize continuous batching and improve export (#506)

Inference

• Add Lora support to stable diffusion by @JingyaHuang (#483)
• Support sentence transformers clip by @JingyaHuang (#495)
• Inference compile cache script by @philschmid and @dacorvo (#496, #504)

Doc

• Update Inference supported models list by @JingyaHuang (#501)

Bug fixes

• inference cache: omit irrelevant config parameters in lookup dy @dacorvo (#494)
• Optimize disk usage when fetching model checkpoints by @dacorvo (#505)

Full Changelog: v0.0.19...v0.0.20

v0.0.19: AWS Neuron SDK 2.17.0, training cache system, TGI improved batching

What's Changed

Training

• Integrate new cache system for training by @michaelbenayoun in #472

TGI

• Support higher batch sizes using transformers-neuronx continuous batching by @dacorvo in #488
• Lift max-concurrent-request limitation usingTGI 1.4.1 by @dacorvo in #488

AMI

• Add packer support for building AWS AMI by @shub-kris in #441
• [AMI] Updates base ami to new id by @philschmid in #482

Major bugfixes

• Fix sdxl inpaint pipeline for diffusers 0.26.* by @JingyaHuang in #458
• TGI: update to controller version 1.4.0 & bug fixes by @dacorvo in #470
• Fix optimum-cli export for inf1 by @JingyaHuang in #474

Other changes

• Add TGI tests and CI workflow by @dacorvo in #355
• Bump to optimum 1.17 - Adapt to optimum exporter refactoring by @JingyaHuang in #414
• [Training] Support for Transformers 4.37 by @michaelbenayoun in #459
• Add contribution guide for Neuron exporter by @JingyaHuang in #461
• Fix path, update versions by @shub-kris in #462
• Add issue and PR templates & build optimum env cli for Neuron by @JingyaHuang in #463
• Fix trigger for actions by @philschmid in #468
• TGI: bump rust version by @dacorvo in #477
• [documentation] Add Container overview page. by @philschmid in #481
• Bump to Neuron sdk 2.17.0 by @JingyaHuang in #487

New Contributors

• @shub-kris made their first contribution in #441

Full Changelog: v0.0.18...v0.0.19

v0.0.18: AWS Neuron SDK 2.16.1, NeuronX TGI improvements, PP for Training

What's Changed

AWS SDK

• Use AWS Neuron SDK 2.16.1 (#449)

Inference

• Preliminary support for neff/weights decoupling by @JingyaHuang (#402)
• Allow exporting decoder models using optimum-cli by @dacorvo (#422)
• Add Neuron X cache registry by @dacorvo (#442)
• Add StoppingCriteria to generate() of NeuronModelForCausalLM by @dacorvo (#454)

Training

• Initial support for pipeline parallelism by @michaelbenayoun (#279)

TGI

• TGI: support vanilla transformer models whose configuration is cached by @dacorvo (#445)

Tutorials and doc improvement

• Various fixes by @jimburtoft @michaelbenayoun @JingyaHuang (#428 #429 #432)
• Improve Stable Diffusion Notebooks by @JingyaHuang (#431)
• Add Sentence Transformers Guide and Notebook by @philschmid (#434)
• Add benchmark section by @dacorvo (#435)

Major bugfixes

• TGI: correctly identify special tokens during generation by @dacorvo (#438)
• TGI: do not include the input_text in generated text by @dacorvo (#454)

Other changes

• API change to be compatible to Optimum by @JingyaHuang (#421)

New Contributors

• @jimburtoft made their first contribution in #432

Full Changelog: v0.0.17...v0.0.18