Skip to content

Latest commit

 

History

History
274 lines (184 loc) · 6.43 KB

CHANGELOG.md

File metadata and controls

274 lines (184 loc) · 6.43 KB
Raw

CHANGELOG

v0.3.0

Added:

  • llama-cpp-cffi server - compatible with llama.cpp cli options instead of OpenAI

Changed:

  • llama.cpp revision a813badbbdf0d38705f249df7a0c99af5cdee678
  • Refactored Options class into two separate classes: ModelOptions, CompletionsOptions

Removed:

  • Removed ambiguous Options class

v0.2.7

Changed:

  • In format_messages, optional options argument
  • llama.cpp revision 081b29bd2a3d91e7772e3910ce223dd63b8d7d26

v0.2.6

Changed:

  • llama.cpp revision 5437d4aaf5132c879acda0bb67f2f8f71da4c9fe

v0.2.5

Fixed:

  • Replaced tokenizer.decode(new_token_id) with custom _common_token_to_piece(context, new_token_id, True)

v0.2.4

Fixed:

  • sampler_init because llama_sampler_init_penalties in llama.cpp changed its behaviour

v0.2.3

Changed:

  • llama.cpp revision 4f51968aca049080dc77e26603aa0681ea77fe45
  • Build process now has global variable LLAMA_CPP_GIT_REF

Fixed:

  • Issue with Phi 3.5 based models, tokenizer.decode([new_token_id], clean_up_tokenization_spaces=False)

v0.2.2

Added:

  • Model.free

Changed:

  • Fixed revision of llama.cpp for all wheels
  • llama.cpp revision c27ac678dd393af0da9b8acf10266e760c8a0912
  • disabled llama_kv_cache_seq_cp in _decode_tokens

v0.2.1

Fixed:

  • Batch "decode" process. NOTE: Encode part is missing for encoder-decoder models.
  • Thread-safe calls to the most important functions of llama, llava, clip, ggml API.

Removed:

  • mllama_completions for low-level function for Mllama-based VLMs

v0.2.0

Added:

  • New high-level Python API
  • Low-level C API calls from llama.h, llava.h, clip.h, ggml.h
  • completions for high-level function for LLMs / VLMs
  • text_completions for low-level function for LLMs
  • clip_completions for low-level function for CLIP-based VLMs
  • WIP: mllama_completions for low-level function for Mllama-based VLMs

Changed:

  • All examples

Removed:

  • llama_generate function
  • llama_cpp_cli
  • llava_cpp_cli
  • minicpmv_cpp_cli

v0.1.23

Added:

  • Support and examples for llava and minicpmv models.

v0.1.22

Added:

  • llava high-level API calls
  • minicpmv high-level API support

Changed: - Updated llama.cpp.

v0.1.21

Changed: - Updated llama.cpp.

v0.1.20

Changed: - Updated llama.cpp.

v0.1.19

Changed: - Updated llama.cpp with RWKV 6 support.

v0.1.18

Changed: - Updated llama.cpp with RWKV support.

v0.1.17

Added: - LLAMA_CPP_BACKEND which can be cuda, vulkan or cpu.

Changed: - Updated llama.cpp. - Updated requirements. - CUDA backend imports only 12.6 library.

Fixed: - Options.top_p check using isinstance.

v0.1.16

Changed: - Updated llama.cpp.

v0.1.15

Added: - SmolLM-1.7B-Instruct-v0.2 examples.

Changed: - Updated llama.cpp.

v0.1.14

Fixed: - Vulkan detection.

v0.1.13

Fixed: - CUDA and Vulkan detection.

v0.1.12

Added: - Build vulkan_1_x for general GPU. - Build cuda 12.4.1 as default.

Changed: - Renamed examples for TinyLlama (chat, tool calling) and OpenAI. - Updated demo models definitions. - Updated examples (chat, tool calling). - get_special_tokens not supports parameter force_standard_special_tokens: bool=False which bypasses tokenizer's special tokens with standard/common ones. - Build cuda 12.5.1 as additional build target but packaged on PyPI. - Build cuda 12.6 as additional build target but packaged on PyPI. - Build openblas as additional build target but packaged on PyPI.

Fixed: - Handle Options.no_display_prompt on Python side.

v0.1.11

Changed: - OpenAI: allow import of routes and v1_chat_completions handler. - examples/demo_0.py, tool calling example.

v0.1.10

Added: - In openai, support for prompt and extra_body. Reference: https://github.com/openai/openai-python/blob/195c05a64d39c87b2dfdf1eca2d339597f1fce03/src/openai/resources/completions.py#L41 - Pass llama-cli options to openai. - util module with is_cuda_available function. - openai supports both prompt and messages. Reference: https://github.com/openai/openai-python/blob/195c05a64d39c87b2dfdf1eca2d339597f1fce03/src/openai/resources/completions.py#L45

v0.1.9

Added: - Support for default CPU tinyBLAS (llamafile, sgemm) builds. - Support for CPU OpenBLAS (GGML_OPENBLAS) builds.

Changed: - Build scripts now have separate step/function cuda_12_5_1_setup which setups CUDA 12.5.1 env for build-time.

Fixed: - Stop thread in llama_generate on GeneratorExit.

Removed: - callback parameter in llama_generate and dependent functions.

v0.1.8

Added: - Model.tokenizer_hf_repo as optional in case when Model.creator_hf_repo cannot be used to tokenize / format prompt/messages.

v0.1.7

Added: - Support for stop tokens/words.

Changed: - llama/llama_cli.py unified CPU and CUDA 12.5 modules into single module.

Removed: - Removed separate examples for CPU and CUDA 12.5 modules.

v0.1.6

Changed: - Updated huggingface-hub.

Fixed: - llama.__init__ now correctly imports submodules and handles CPU and CUDA backends. - OpenAI: ctx_size: int = config.max_position_embeddings if max_tokens is None else max_tokens.

v0.1.5

Fixed: - Build for linux, upx uses best compression option, 7z uses more aggressive compression. - Do not use UPX for shared/dynamic library compression.

v0.1.4

Added: - README: supported GPU Compute Capability for CUDA.

Fixed: - Cleaned up build.py. - Type annotations in OpenAI related code.

v0.1.3

Added: - Support for PyPy 3.10 versions.

Changed: - Disabled GitHub Actions. - Uses upx -9 to compress shared/dynamic libraries. - Repacks whl with better compression rate. - Auto-detect CUDA support.

Removed: - ctypes version and demos.

v0.1.2

Added: - Preparation for Chat Completions API by OpenAI © compatible server.

Fixed: - Argument options is deepcopy-ed when passed to llama_generate(options), so it can be reused.

Changed: - Build for manylinux_2_28 and musllinux_1_2. - Build for CUDA Compute Capability >= 6.1.

v0.1.1

Changed: - Updated: huggingface-hub = "^0.24.0", setuptools = "^71.0.3"

v0.1.0

Added: - Park first version.