v0.31.2-rc1
Breaking Changes📦 ollamaView on GitHub →
⚠ 2 breaking✨ 2 features🐛 3 fixes🔧 1 symbols
Summary
This release focuses on improving hardware acceleration compatibility, particularly for CUDA 6.x GPUs, and cleans up experimental features and dead code. It also addresses CUDA toolkit lookup issues.
⚠️ Breaking Changes
- The environment variable OLLAMA_EXPERIMENT=client2 has been removed, which might break workflows relying on this experimental feature.
- ROCm support has been removed for devices that are no longer supported.
Migration Steps
- Remove any usage or reliance on the OLLAMA_EXPERIMENT=client2 environment variable.
- Review ROCm device support if using older hardware.
✨ New Features
- Enabled Flash Attention (FA) support on CUDA Compute Capability 6.x GPUs.
- Introduced agent core harnessing capabilities.
🐛 Bug Fixes
- Fixed CUDA toolkit lookup and parallel execution issues.
- Resolved an issue where the JetPack runner absence prevented fallback to standard CUDA.
- Updated file opening to use UTF-8-safe encoding.