3.0.0-beta.45 (2024-09-19)

Bug Fixes

• improve performance of parallel evaluation from multiple contexts (#309) (4b3ad61)
• Llama 3.1 chat wrapper standard chat history (#309) (4b3ad61)
• adapt to llama.cpp sampling refactor (#309) (4b3ad61)
• Llama 3 Instruct function calling (#309) (4b3ad61)
• don't preload prompt in the chat command when using --printTimings or --meter (#309) (4b3ad61)
• more stable Jinja template matching (#309) (4b3ad61)

Features

• inspect estimate command (#309) (4b3ad61)
• move seed option to the prompt level (#309) (4b3ad61)
• Functionary v3 support (#309) (4b3ad61)
• Mistral chat wrapper (#309) (4b3ad61)
• improve Llama 3.1 chat template detection (#309) (4b3ad61)
• change autoDisposeSequence default to false (#309) (4b3ad61)
• move download, build and clear commands to be subcommands of a source command (#309) (4b3ad61)
• simplify TokenBias (#309) (4b3ad61)
• better threads default value (#309) (4b3ad61)
• make LlamaEmbedding an object (#309) (4b3ad61)
• HF_TOKEN env var support for reading GGUF file metadata (#309) (4b3ad61)
• TemplateChatWrapper: custom history template for each message role (#309) (4b3ad61)
• more helpful inspect gpu command (#309) (4b3ad61)
• all tokenizer tokens iterator (#309) (4b3ad61)
• failed context creation automatic remedy (#309) (4b3ad61)
• abort generation support in CLI commands (#309) (4b3ad61)
• --gpuLayers max and --contextSize max flag support for inspect estimate command (#309) (4b3ad61)
• extract all prebuilt binaries to external modules (#309) (4b3ad61)
• updated docs (#309) (4b3ad61)
• combine model downloaders (#309) (4b3ad61)
• feat(electron example template): update badge, scroll anchoring, table support (#309) (4b3ad61)

Shipped with llama.cpp release b3785

To use the latest llama.cpp release available, run npx -n node-llama-cpp source download --release latest. (learn more)