EasyDeL version 0.0.69

This release brings significant scalability improvements, new models, bug fixes, and usability enhancements to EasyDeL.

Highlights:

• Multi-host GPU Training: EasyDeL now scales seamlessly across multiple GPUs and hosts for demanding training workloads.
• New Models: Expand your NLP arsenal with the addition of Gemma2, OLMo, and Aya models.
• Improved KV Cache Quantization: Enjoy a substantial accuracy boost with enhanced KV cache quantization, achieving +21% accuracy compared to the previous version.
• Simplified Model Management: Load and save pretrained models effortlessly using the new model.from_pretrained and model.save_pretrained methods.
• Enhanced Generation Pipeline: The GenerationPipeLine now supports streaming token generation, ideal for real-time applications.
• Introducing the ApiEngine: Leverage the power of the new ApiEngine and engine_client for seamless integration with your applications.

Other Changes:

• Fixed GPU Flash Attention bugs for increased stability.
• Updated required jax version to >=0.4.28 for optimal performance. Versions 0.4.29 or higher are recommended if available.
• Streamlined the structure import process and resolved multi-host training issues.

Upgrade:

To upgrade to EasyDeL v0.0.69, use the following command:

pip install --upgrade easydel==0.0.69

EasyDeL - 0.0.67

• New Features

  • GenerationPipeLine was added for fast streaming and easy generation with JAX.
  • Using Int8Params instead of LinearBitKernel.
  • Better GPU support.
  • EasyDeLState is now better and supports more general options.
  • Trainers now support .save_pretrained(to_torch) and training logging.
  • EasyDeLState supports to_8bit.
  • All of the models support to_8bit for params.
  • imports are now 91x times faster in EasyDeL version 0.0.67.

• Removed API

  • JAXServe is no longer available.
  • PyTorchServe is no longer available.
  • EasyServe is no longer available.
  • LinearBitKernel is no longer available.
  • EasyDeL partitioners are no longer available.
  • Llama/Mistral/Falcon/Mpt static convertors or transforms are no longer available.

• Known Issues

  • Lora Kernel Sometimes Crash.
  • GenerationPipeLine has a compiling problem when the number of available devices is more than 4 and using 8_bit params.
  • Most of the features won't work for TPU-v3 and GPUs with compute capability lower than 7.5.
  • Kaggle session will crash after importing EasyDeL (Kaggle's latest environment is not stable it's not related to EasyDeL). (Fixed in EasyDeL version 0.0.67)

Pallas Fusion: GPU Turbocharged 🚀

EasyDeL version 0.0.65

• New Features

  • Pallas Flash Attention on CPU/GPU/TPU via FJFormer and supports bias.
  • ORPO Trainer is added and now it's in your bag.
  • WebSocket Serve Engine.
  • Now EasyDeL is 30% faster on GPUs.
  • No JAX-Triton is now needed to run GPU kernels.
  • Now you can specify the backward kernel implementation for Pallas Attention.
  • now you have to import EasyDeL as easydel instead of EasyDel.

• New Models

  • OpenELM model series are now present.
  • DeepseekV2 model series are now present.

• Fixed Bugs

  • CUDNN FlashAttention Bugs are now fixed.
  • Llama3 Model 8Bit quantization of parameters had a lot of improvements.
  • Splash Attention bugs on TPUs are now fixed .
  • Dbrx Model Bugs are fixed.
  • DPOTrainer Bugs are Fixed (creating dataset).

• Known Bugs

  • Splash Attention won't work on TPUv3.
  • Pallas Attention won't work on TPUv3.
  • You need to install flash_attn in order to convert HF DeepseekV2 to EasyDeL (bug in DeepseekV2 implementation from original authors).
  • Some Examples are out dated.

Full Changelog: 0.0.63...0.0.65

0.0.63

whats changed

• Phi3 Model Added.
• Dbrx Model Added.
• Arctic Model Added.
• Lora Fine-Tuning Bugs Fixed.
• Vanilla Attention is Optimized.
• Sharded Vanilla is the default attention mechanism now.

Full Changelog: 0.0.61...0.0.63

EasyDeL-0.0.61 Dynamic Changes

What's Changed

• Add support for iterable dataset loading by @yhavinga in #138
• SFTTrainer bugs are fixed.
• Parameter quantization is now available for all of the models.
• AutoEasyDeLModelForCausalLM now supports load_in_8bit.
• Memory Management improved.
• Gemma Models Generation Issue is now Fixed.
• Trainers are now 2~8% faster.
• Attention Operation is improved.
• The Cohere Model is now present.
• JAXServer is improved.
• Due to recent changes a lot of examples of documentation have changed and will be changed soon.

Full Changelog: 0.0.60...0.0.61

EasyDeL Version 0.0.60

What's Changed

• SFTTrainer is now available.
• VideoCausalLanguageModelTrainer is now available.
• New models such as Grok-1, Qwen2Moe, Mamba, Rwkv, and Whisper are available.
• MoE models had some speed improvements.
• Training Speed is now 18%~42% faster.
• Normal Attention is now faster by 12%~30% #131 .
• DPOTrainer Bugs Fixed.
• CausalLanguageModelTrainer is now more customizable.
• WANDB logging has improved.
• Performace Mode is added to Training Arguments.
• Model configs pass attributes to PretrainedConfig to prevent override… by @yhavinga in #122
• Ignore token label smooth z loss by @yhavinga in #123
• Time the whole train loop instead of only call to train step function by @yhavinga in #124
• Add save_total_limit argument to delete older checkpoints by @yhavinga in #127
• Add gradient norm logging, fix metric collection on multi-worker setup by @yhavinga in #135

Full Changelog: 0.0.55...0.0.60

EasyDeL Version 0.0.55

EasyDeL Version 0.0.55

• JAX DPOTrainer Bugs Fixed
• StableLM Models are supported with FlashAttention and RING-Attention
• RingAttention is supported for Up to 512K or 1M token training and inference
• chunk MLP Is Supported for Up to 512K or 1M token training and inference
• now all the Models support shared key and value caching for high context length interface and can be accessed via use_sharded_kv_caching=True in model config (see examples).
• EasyDeL successfully passed 1256000 Context Length Inference on TPUs (Llama Model Tested)
• Vision Trainer is added, you might except some bugs from that.

Full Changelog: 0.0.50...0.0.55

0.0.50 Mixture of EasyDeL experts

What's Changed

• Optimize mean loss and accuracy calculation by @yhavinga in #100
• Mixtral Models are fully supported and they are PJIT-compatible
• A Wider range of models now support FlashAttention on TPU
• Qwen 1, Qwen 2, PHI 2, Robert is new Added Models which support FlashAttention on TPU and EasyBIT
• LoRA support for the trainer is now Added (EasyDeLXRapTureConfig)
• Adding EasyDel Serve Engine APIs
• Adding Prompter (Beta and might be removed in future updates)
• The Training Process is now 21 % Faster in 0.0.50 than 0.0.42.
• Transform Functions are now Automated for all the models (Except MosaicMPT for this one you still have to use static methods)
• The Trainer APIs have changed and now it's faster, more dynamic, and more hackable.
• Default Version of the JAX now changed to 0.4.22 for FJFormer custom Pallas kernels usage.

New Contributors

• @yhavinga made their first contribution at #100

Full Changelog: 0.0.42...0.0.50

Version 0.0.42 Easy State

New Features:

• EasyDelState is added
• Auto Convertors from torch > huggingface > jax > flax > EasyDel are added
• Trainer has a lot of improvements

Full Changelog: 0.0.41...0.0.42

0.0.41

what has changed so far in 0.0.41

• API Changes
• making CausalLanguageModel Trainer separated from others
• Custom Errors added
• Timer bugs fixed
• AutoEasyDelForCasualLM is now more automated and falcon bugs have been fixed
• 4D Mesh being used for better partitioning
• And many many more

Full Changelog: 0.0.40...0.0.41