v5.10.1
Breaking Changes📦 invokeaiView on GitHub →
⚠ 2 breaking✨ 3 features🐛 2 fixes🔧 3 symbols
Summary
This patch introduces VRAM optimization via partial loading for LLaVA and SigLIP models and fixes installation issues for older versions. Critical pre-update steps are required for users on versions older than v5.10.0.
⚠️ Breaking Changes
- If you are on v5.9.1 or older, updating may fail if you do not use the latest installer/launcher (v1.5.0). Fix: Download and use the latest installer/launcher from https://invoke.com/downloads.
- The Form Builder will be reset upon first launch after updating to v5.10.0, causing loss of unsaved changes. Fix: Save your current workflow before updating and manually reload it afterward.
Migration Steps
- If on v5.9.1 or older, update to the latest installer/launcher (v1.5.0) before updating Invoke.
- If the installation fails, retry the installation with **repair mode** enabled (tick the checkbox on the **Review** step of the install).
- Save any unsaved Form Builder changes before updating, as the Form Builder will reset on first launch post-update.
✨ New Features
- Support partial loading for LLaVA and SigLIP (FLUX Redux) models, reducing VRAM requirements for users with Nvidia GPUs.
- Reduced peak CPU RAM usage during initial load of LLaVA and SigLIP models, allowing users with at least 24GB CPU RAM to run LLaVA 7B model without crashing during load.
- The data viewer component now wraps lines when displaying JSON data (e.g., metadata, workflows, node outputs).
🐛 Bug Fixes
- Reverted a recent change to model installation that could cause some models to be misidentified as LoRAs.
- Fixed hard crash on initial load for LLaVA and SigLIP models via partial loading implementation.
🔧 Affected Symbols
LLaVASigLIPForm Builder