●Change8

b7668

Breaking Changes
📦 llama-cppView on GitHub →
⚠ 2 breaking✨ 1 features🐛 5 fixes🔧 5 symbols

Summary

This release introduces a new `use_direct_io` flag for model loading and includes several fixes related to raw I/O operations, particularly on Windows, alongside internal type and function signature adjustments.

⚠️ Breaking Changes

  • The type for off_t was changed to size_t on Windows, which may affect code relying on the previous type definition.
  • Functions were renamed, requiring updates to any code that directly called the old function names.

Migration Steps

  1. Review and update any code that relies on the previous type definition of off_t on Windows to use size_t.
  2. Update any direct calls to functions that were renamed during this release.

✨ New Features

  • Added a `use_direct_io` flag for model loading via the `--direct-io` command-line argument.

🐛 Bug Fixes

  • Fixed issues with `read_raw()` calls.
  • Fixed `read_raw_at` implementation on Windows.
  • Disabled direct I/O when mmap is explicitly enabled.
  • Implemented fallback to `std::fread` if O_DIRECT fails due to a bad address.
  • Removed `const` keywords and unused functions on Windows.

🔧 Affected Symbols

read_raw()read_raw_atoff_tsize_tsrc/llama-mmap.cpp