0.19.1

Highlights

• support for schedulers when using diffusion
• support for minP as a generation parameter
• several fixes for docker support, and we now use a single Dockerfile for all platforms
• trains will now use the first unused split in a dataset if there is no "train" split
• upgrades to MLX LoRA trainer to better support Qwen models
• fix for embedding dataset creation using Synthesizer Docs plugin
• run a model on an inference engine even if the architecture is not listed as supported

Details

• schedulers drop-down in Diffusion tab by @Sourenm in #539
• Add Minimum P sampling for all generations by @deep1401 in #538
• Model architecture override by @mina-parham in #519
• Fix issues related to unsupported engine checkbox by @mina-parham in #546
• Add model architecture on server start by @deep1401 in #540
• Moved entire Workflows router under Experiment by @aahaanmaini in #531
• Fix default inference engine logic for new experiment by @deep1401 in #554
• Fix errors showing up in console from Import Models modal by @dadmobile in #548
• Add Alert Message on Workflow Start by @aahaanmaini in #553

Full Changelog: v0.19.0...v0.19.1

0.19.0

This release introduces support for diffusion models including inference, training and improved management of image datasets.

What's Changed

• Make Workflows tab visible by @aahaanmaini in #517
• V1 of adaptors search and install for models by @Sourenm in #514
• Added creation and preview of local image datasets with in place caption editing and searching abilities by @Sourenm in #498
• fix for drag and drop for image datasets by @Sourenm in #523
• Workflows belong to Experiments by @aahaanmaini in #525
• Track job queue events by @aliasaria in #516
• Fix errant API calls on Model Zoo by @dadmobile in #524
• Export experiment to recipe JSON by @aahaanmaini in #520
• Fix information for GPU metrics on Header in Mac by @deep1401 in #532
• Added the same dataset preview logic that exists in the Dataset tab to the Generate tab by @Sourenm in #533
• Add Diffusion by @deep1401 in #488
• Fix/separate-use-analytics by @aliasaria in #536

Full Changelog: v0.18.0...v0.19.0

0.18.0

The most exciting new feature is support for multimodal model training. This includes a VLM trainer and the ability to work with vision datasets.

Notable bug fixes:

• Canceling downloads is a lot more responsive
• LlamaIndex plugin setup error causing issues for some RAG users
• removed limit on the number of samples being used for trains on STF Llama Factory plugin
• Eval and Generate jobs no longer show as COMPLETE when they were STOPPED

What's Changed

• v1 of the Recipes Modal (hidden in DEV_MODE) by @aliasaria in #495
• Preview Image Data in Datasets Tab by @deep1401 in #470
• Changed the visuals in Download Progress Bar by @Sourenm in #491
• Mark plugin as deprecated if supported_hardware_architectures is an empty list by @deep1401 in #494
• Add MCP Server support in Tools tab within Interact tab by @deep1401 in #468
• Recipe navigates to Notes page by @aahaanmaini in #509

Full Changelog: v0.17.1...v0.18.0

0.17.1

What's Changed

• Fix apple compatibility machine type by @deep1401 in #497

Full Changelog: v0.17.0...v0.17.1

0.17.0

• AMD GPU Support!
• New Model Zoo groups models to make it easier to navigate
• Export screen updated to match style of other job pages
• Dataset fixes: Better viewing of ChatML datasets, fixed issues using some of our gallery datasets

What's Changed

• Add a tab for running Sweeps by @deep1401 in #382
• Add/model-groups-with-tags-search-bar by @Sourenm in #477
• Fixed the Preview Dataset feature for ChatML style datasets by @Sourenm in #479
• Fix tab for model zoo sidebar by @deep1401 in #482
• Recipe Modal by @aliasaria in #484
• fix model store option default in foundation section by @mina-parham in #486
• Make the 'New' button dropdown in the Train section scrollable by @mina-parham in #487
• Make Exporter Page Match Others by @aahaanmaini in #467
• receipes: fake loading of dependencies by @aliasaria in #490
• Fix Local Models Download bar, inference destructuring and experimental plugin card overflows by @deep1401 in #489
• Use datasets.path instead of datasets.name while creating a task from recipe by @deep1401 in #492
• Add support for the new Claude 4 models by @deep1401 in #493
• Add support for AMD GPUs by @deep1401 in #425
• One liner change for correct machine type by @deep1401 in #496

New Contributors

• @mina-parham made their first contribution in #486

Full Changelog: v0.16.4...v0.17.0

0.16.4

What's Changed

• Github build upset about outputs in publish.yml by @dadmobile in #475

Full Changelog: v0.16.3...v0.16.4

0.16.1

What's Changed

• Fix creating a new experiment by @deep1401 in #458

Full Changelog: v0.16.0...v0.16.1

0.16.0

• Major update to CUDA and torch in order to add support for NVIDIA 50 series GPUS
• new MLX PPO Trainer plugin!
• ability to run 4-bit bitsandbytes models in fastchat

Bug fixes

• local and generated datasets viewing doesn’t throw an error
• ability to manually set fallback context length used by MLX
• names of task templates update correctly

What's Changed

• Fix internal error on Datasets tab due to name field by @deep1401 in #444
• Add/further sdk cleanup by @aliasaria in #447
• Add logic to filter experimental plugins by @deep1401 in #451
• Fix rendering of eval and generate configs on update by @deep1401 in #455

Full Changelog: v0.15.2...v0.16.0

0.15.2

Highlights

This build fixes an installer bug that was caused by the latest API update.

What's Changed

• Simplify transformerlab api sdk by @aliasaria in #418
• Fixed dropdown dark mode background by @aahaanmaini in #434
• If Plugin install fails, show a message by @dadmobile in #430
• Remove unneeded API call on startup by @dadmobile in #435
• Split up transformerlab-api-sdk into components by @aliasaria in #438
• Add/first-example-of-using-new-sdk by @aliasaria in #441
• Added input field to CSS by @aahaanmaini in #437
• All exporter plugins showing 'Exporting...' when one runs by @aahaanmaini in #442
• Fix info icon for training jobs by @deep1401 in #445
• Change key dependency name to transformerlab-inference to fix broken installer by @deep1401 in #446

New Contributors

• @aahaanmaini made their first contribution in #434

Full Changelog: v0.15.1...v0.15.2

0.15.1

This is a patch release with a number of bug fixes:

• Web app will auto-connect to API running on the same machine
• Installer no longer halts if local_server.log is not accessible/writable
• Console automatically opens to show error if inference server fails to start correctly
• Console logger scrolls way faster
• Plugins no longer always say they are unsupported for non-GPU/mac users
• if you upload docs before installing RAG plugin it will index the docs for you before running your query
• You can now edit the setup script for plugins inside Transformer Lab (good for debugging or making custom modifications to a plugin’s venv)

git Details

• Open console logger upon a failed inference server run by @deep1401 in #415
• Fix supported arch message for local plugins by @deep1401 in #417
• Fix error caused by deleting plugins by @deep1401 in #419
• If you are on the webapp, correctly find the local server we should first try to connect to by @aliasaria in #421
• Don't hang or crash if local_server.log is not writeable by @dadmobile in #423
• Reduce console scroll delay from 100ms to 10ms by @dadmobile in #428

Full Changelog: v0.15.0...v0.15.1