v0.3.0
v0.2.5
v0.2.4
v0.2.4 Latest - 2025-04-06
SDK Fix
-
In simple-conversation program, #include <locale.h> changed to #include <clocale>. Former was causing compiler errors on macOS. f782666
-
llama kv cache methods names are updated. 0ab7b66
-
Fixed a problem in io_file.h file which were causing compiler errors on windows. e64c869
-
Phi model template fixed. 124e9d2
SDK Added
v0.2.3
v0.2.2
v0.2.1
Release Notes
- In decode behavior description, remove the halt delay and work with mutexes. -> f3e924f
- Fixed a major bug in vector and string erase method in which was causing std::bad_alloc exception.
About Pre-compiled Binaries
The precompiled binaries can't be provided for now due to some incorrectness in project configuration.
Refer to compiling from source: https://docs.mbasesoftware.com/inference/quickstart/setup/compile_from_source
v0.2.0
Release Notes
- Write safer is_float, is_number ... methods in mbase::string -> 7d385b1
- Implement processor unregistration. -> 3267bb1
- Processor LoRA implementation is missing. -> a8f4406
API Changes
- mbase::vector<inf_lora_adapter> get_adapters() const; method implemented in InfModelTextToText.
- I32 get_cache_token_count() const;, method implemented in InfProcessorTextToText. By this, users will be able to see how many tokens are in the kv cache.
- LoRA operations are implemented.
About Pre-compiled Binaries
The precompiled binaries can't be provided for now due to some incorrectness in project configuration.
Refer to compiling from source: https://docs.mbasesoftware.com/inference/quickstart/setup/compile_from_source