ChatLLaMA 0.0.4

Release Notes - ChatLLaMA 0.0.4

Major release of ChatLLaMA adding support to efficient training using LoRA.

New Features

• HF-based models can now be trained on both actor and RLHF using LoRA. This will help users in reducing the memory needed for training.

ChatLLaMA 0.0.3

Release Notes - ChatLLaMA 0.0.3

Major release of ChatLLaMA solving multiple bugs and expanding the support to distrubuted training.

New Features

• Training produces a log file containing also the training stats
• Add Template during the dataset creation
• Changed default training parameters to InstructGPT paper ones
• Implemented Cosine Scheduler for LR
• Add Conversation Logs during RLHF
• Improved management of checkpoints
• Add support for Accelerate backend for distributed training
• Defer import of LLaMA modules, avoiding the LLaMA dependecy while another model is used

Bug Fix

• Fix bug causing crashes during training due to long sequences
• Fix multiple bugs on deepspeed MultiGPU training
• Fix bugs on RLHF training
• Fix bug with different versions of LangChain during synthetic data generation

New Contributors

• @PierpaoloSorbellini made their first contribution in #185
• @theSekyi made their first contribution in #187
• @egrefen made their first contribution in #190
• @AAnirudh07 made their first contribution in #202
• @dentathor made their first contribution in #204
• @bzantium made their first contribution in #212
• @pgzhang made their first contribution in #230
• @sebastianschramm made their first contribution in #237
• @HuangLK made their first contribution in #253
• @zhzou2020 made their first contribution in #271

v0.9.0

Nebullvm 0.9.0 Release Notes

Major release for Nebullvm, adding support to diffusion model optimization.

New features

• Add support for diffusers UNet

Bug fixed

• Fix CI pipelines triggers

Speedster 0.3.0 Release Notes

Major release adding support to the diffusers library.

New features

• Add support to diffusers library. Speedster can now optimize diffusion models with a single line of code.
• Update readme for HF models.

Bug fixed

• Fix import error in Google Colab
• Fix few typos in the docs and update the benchmarks

New Contributors

• @mfumanelli made their first contribution in #178

v0.8.1

nebullvm 0.8.1 Release Notes

This is a minor release fixing multiple bugs.

New features

• Changed the Auto-Installer API
• Added support to onnxruntime TensorrtExecutionProvider

Bug fixed

• Fixed bug in torchscript casting integers to fp16.
• Optimized the memory usage for inference learners
• Now the tensorrt workspace size is dynamically computed according to the free memory available on the gpu device
• Fixed a bug in openvino when using static quantization

speedster 0.2.1 Release Notes

This is a minor release fixing multiple bugs.

New Features

• Added support for device selection when working multi-gpu environment
• Added support to input data with inconsistent batch size
• Implemented benchmark functions also for TensorFlow and ONNX

Bug Fixed

• Optimized the gpu memory usage during speedster optimization

Contributors

• @valeriosofi
• @cccntu
• @VinishUchiha

v0.8.0

nebullvm 0.8.0 Release Notes

This is a major release fixing multiple bugs and implementing two new functions for loading and saving the models.

New Features

• Implements two new functions for loading and saving inference learners.

Bug fixed

• Fixed bug for ONNXRuntime models being loaded on the wrong device after the optimization.
• Fixes TensorRT behaviour when using dynamic shape
• Improves significantly the performance of TensorRT with ONNX interface
• Limits the gpu memory used by tensorflow, to avoid memory issues during tensorflow models optimization
• Fixes some issues with unit tests, and adds additional controls to ensure that an optimized model with dynamic shape works properly.
• removes setuptools from tensorrt installation, it's not needed anymore by the newer version and it was causing issues with the azure pipelines.

speedster 0.2.0 Release Notes

Major release for Speedster adding the load_model and save_model functions.

New Features

• Improved the logs
• Save and load model functions can be imported directly from speedster
• Brand new documentation moved directly in the GitHub repository and built using mkdocs

Bug Fixed

• Fixed version compatibility issue on Colab
• adds missing port forwarding in the docker run command inside the notebooks readme.

Contributors

• Valerio Sofi (@valeriosofi)
• Diego Fiori (@diegofiori)
• Leonardo Zecchin (@ZecchinLeonardo)
• Arian Ghasemi (@arianGh1)

v0.7.3

nebullvm 0.7.3 Release Notes

This is a minor release modifying the metric_drop_ths behaviour.

New Features

• metric_drop_ths by default set to 0.001
• Half precision is activated for each value of metric_drop_ths>0
• Int8 quantization is activated just for values of metric_drop_ths>0.01

Speedster 0.1.2 Release Notes

This is a minor release editing the default value for metric_drop_ths and modifying the summarisation table at the end of the optimization.

New Features

• The summary at the end of the optimization now shows in a clear way the improvements.
• Warning with tips on how to improve the optimization result is now shown also for speedups lower than 2x.

v0.7.2

nebullvm 0.7.2 Release Notes

This is a minor release editing the logs style.

New Features

• Warning logs are now less intrusive.

Speedster 0.1.1 Release Notes

This is a minor release adding tips and improving the logs.

New Features

• Suggestions are provided to the user on how to increase model performance when optimization does not provide the required speedup.
• Improved the quality of the logs.

v0.7.1

nebullvm 0.7.1 Release Notes

This is a minor release fixing few bugs for supporting YOLOv8 models.

Bug fixed

• Fixed bugs with half-precision in PyTorch for supporting YOLOv8 models.

v0.7.0

nebullvm 0.7.0 Release Notes

This is a major release simplifying the backends installation and improving the UX.

New Features

• The auto-installer interface has been simplified and clarified.

Bug fixed

• Fixed problem with shell not found when installing TensorRT on specific Linux systems.

speedster 0.1.0 Release Notes

Major release for Speedster, now supporting TF backend for HuggingFace transformers.

New Features

• Add support for HuggingFace models having a TensorFlow backends.
• Improved the logs style.
• Optimization results are now summarised into a nice and informative table.

Contributors

• Valerio Sofi (@valeriosofi)

v0.6.0

nebullvm 0.6.0 Release Notes

This release of Nebullvm modifies the structure of the library. Nebullvm library is transformed into a framework which can be used for building Apps for AI optimization. The end-to-end optimization tool is moved into speedster, an App built using nebullvm as underlying framework.

New Features

• The API for model optimization has been moved in app/accelerate/speedster.
• The optimize_model function in nebullvm has been deprecated. It will be removed in the next major release.
• Added support for windows platforms.

speedster 0.0.1 Release Notes

Speedster is the new library replacing the previous nebullvm API optimizing DL models in a single line of code. Speedster keeps the same interface as the previous nebullvm API.

New Features

• Model optimization API moved into speedster. Now it is possible to import the optimize_model function running from speedster import optimize_model.

Contributors

• Diego Fiori (@morgoth95)
• Valerio Sofi (@valeriosofi)