b8320
📦 llama-cppView on GitHub →
✨ 4 features🐛 7 fixes🔧 2 symbols
Summary
This release focuses heavily on improving the test-backend-ops framework by enabling JSON loading, adding operator extraction tools, and fixing numerous build and runtime issues, particularly around tensor views and internal API usage.
Migration Steps
- Replace internal API calls with the public llama_graph_reserve call.
- Move the llama_graph_reserve function to the new llama-ext header.
- Move export-graph-ops to the tests/ directory.
✨ New Features
- Enabled loading of test-backend-ops tests from JSON files.
- Added error threshold configuration based on operator execution.
- Introduced a tool for extracting graph operators into JSON format.
- Added 'nb' parameter support for non-contiguous input tensors.
🐛 Bug Fixes
- Fixed an issue related to view checks.
- Ensured view operations are only used when tensors are non-contiguous or permuted, and switched to C++ random instead of rand().
- Fixed an issue where nb[0] was not being set correctly for view operations.
- Resolved an inplace error.
- Fixed a missing declaration issue.
- Fixed indentation issues.
- Fixed Windows build failures.