mobarski/aidapter

Adapter / facade for language models (OpenAI, Anthropic, Cohere, local transformers, etc)

aidapter

Simple adapter for many language models - remote (Hugging Face, OpenAI, AnthropicAI, CohereAI) and local (transformers library).

Facilitates loading of many new models (Guanaco, Falcon, Vicuna, etc) in 16/8/4 bit modes.

It also supports embedding models (OpenAI, CohereAI, Sentence Transformers).

Installation

🚧 This is experimental software. Anything can change without any notice.

pip install git+https://github.com/mobarski/aidapter.git

Note: each vendor API requires manual installation of dependencies.

Features

• simple, unified API to many models (remote and local)
• batching
• parallel calls
• caching
• usage tracking
• automatic retries
• response priming

Usage examples

completion:

>>> import aidapter
>>> model = aidapter.model('openai:gpt-3.5-turbo') # uses OPENAI_API_KEY env variable
>>> model.complete('2+2=')
4

>>> model.complete(['2+2=','7*6=']) # parallel calls
['4', '42']

embeddings:

>>> model = aidapter.model('sentence-transformers:multi-qa-mpnet-base-dot-v1')
>>> vector = model.embed('mighty indeed')
>>> vector[:5]
[-0.07946087, -0.2150347, -0.33358946, 0.18340564, 0.16403404]

>>> vectors = model.embed(['this is the way', 'so say we all']) # parallel / batch processing
>>> [x[:5] for x in vectors]
[[0.037638217, -0.30608281, -0.3064257, -0.46715638, -0.2608084],
 [-0.063842215, -0.16669855, -0.22363697, -0.2893797, 0.060464755]]

multiple models:

>>> m1 = aidapter.model('transformers:ehartford/Wizard-Vicuna-13B-Uncensored:4bit') # 4 bit mode
>>> m2 = aidapter.model('anthropic:claude-instant-v1') # uses ANTHROPIC_API_KEY env variable

persistent cache and usage tracking:

>>> import shelve
>>> model.cache = shelve.open('/tmp/aidapter.cache') # persistant disk cache
>>> model.usage = shelve.open('/tmp/aidapter.usage') # persistant usage tracking

>>> import diskcache as dc
>>> model.cache = dc.Cache('/tmp/aidapter.cache') # persistant disk cache
>>> model.usage = dc.Cache('/tmp/aidapter.usage') # persistant usage tracking

function calling interface*:

>>> def get_weather(city):
>>>     "get weather info for a city; city must be all caps after ISO country code and a : separator (e.g. FR:PARIS)"
>>>     ...
>>> model = aidapter.model('openai:gpt-3.5-turbo-0613')
>>> model.complete('Whats the weather in the capital of Poland?', functions=[get_weather])
{'function_name': 'get_weather', 'arguments': {'city': 'PL:WARSAW'}}

* currently, it works only with selected OpenAI models

use last_hidden_state from any transformer as an embedding*:

>>> model = aidapter.model('transformers:RWKV/rwkv-raven-1b5')
>>> model.raw_embed_one('mighty indeed')[:5]
[0.14850381016731262, -0.021324729546904564, 0.09214707463979721, 0.34308338165283203, -0.11288302391767502]

* requires additional normalization over a corpus, API will change

API

aidapter.model(model_id, **api_kwargs) -> model

• model_id - model identifier in the following format <vendor_name>:<model_name>
• api_kwargs - default API arguments

model.complete(prompt, system='', start='', stop=[], limit=100, temperature=0, functions=[], cache='use', debug=False) -> str | list | dict

• prompt - main prompt or list of prompts

• system - system prompt

• start - the text that will be appended to the start of the response and to the end of the prompt (aka response priming)

• stop - list of strings upon which to stop generating

• limit - maximum number of tokens to generate before stopping (aka max_new_tokens, max_tokens_to_sample)

• temperature - amount of randomness

• functions - list of functions available to the model (none of them will be executed - only the signatures are used)

• cache - cache usage:

  • use - use the cache if the temperature is 0 (default)
  • skip - don't use the cache
  • force - use the cache even if the temperature is not 0

• debug - if True, the function will return a dictionary (or a list of dictionaries) containing internal objects / values

  FULL_PROMPT = system + prompt + start

model.embed(input, limit=None) -> list | list[list]

• input - text or list of texts
• limit - limit the vector length to first n dimensions (default = None = no limit)

model configuration:

• model.workers - number of concurrent workers for parallel completion (default=4)

• model.show_progress - show progress bar when performing parallel completion (default=False)

• model.retry_tries - maximum number of retry attempts (default=5)

• model.retry_delay - initial delay between retry attempts (default=0.1)

• model.retry_backoff - multiplier applied to the delay between retry attempts (default=3)

Supported models

OpenAI

• openai:gpt-4

• openai:gpt-4-32k

• openai:gpt-3.5-turbo

• openai:text-davinci-003

• openai:code-davinci-002

• ...

API key env. variable: OPENAI_API_KEY

Anthropic

• anthropic:claude-v1

• anthropic:claude-instant-v1

• anthropic:claude-v1-100k

• anthropic:claude-instant-v1-100k

• ...

API key env. variable: ANTHROPIC_API_KEY

Cohere

• cohere:command

• cohere:command-light

• ...

API key env. variable: CO_API_KEY

Transformers

• transformers:TheBloke/guanaco-7B-HF

• transformers:tiiuae/falcon-7b

• transformers:RWKV/rwkv-raven-3b

• transformers:ehartford/Wizard-Vicuna-13B-Uncensored

• transformers:roneneldan/TinyStories-33M

• ...

Change log

0.6.4

• initial support for HF API models
• removed old HF implementation

0.6.3

• OpenAI's embeddings now use BaseModelV2

0.6.2

• as_iter option in BaseModelV2.transform
• removed BaseModelV2.register_progress

0.6.1

• handle cache=False in BaseModelV2.transform_many
• hf2 brand renamed to huggingface

0.6

• initial support for HF API embeddings
• BaseModelV2
  • cleaner code
  • diskcache support
  • batch + threads support
  • retry configuration
  • progress update

0.5.4

• initial support for the functions argument (works only with selected OpenAI models)

0.5.3

• initial support for raw_embed_one in transformers (for creating embeddings from ANY transformer models)

0.5.2

• fix: kw handling in get_cache_key

0.5.1

• limit option for embedding models

0.5

• initial support for embedding models (requires more work with batch / parallel processing):
  • OpenAI
  • Cohere
  • Sentence Transformers

0.4.4

• response priming (start option)

0.4.3

• stop option for transformers

0.4.2

• anthropic usage: tokens, characters

• transformers usage: tokens, characters

0.4.1

• remove prompt from transformers output
• removed kvdb
• usage['time']
• fixed pad_token_id
• fixed limit in transformer models

0.4

• initial support for local transformers models

  • float16 (add ":16bit" to the model name)

  • load_in_8bit (add ":8bit" to the model name)

  • load_in_4bit (add ":4bit" to the model name)

• cache = use | skip | force

• shelve based persistence (for cache and usage)

0.3.2

• kvdb import fix

0.3

• Cohere models
• disk cache

0.2

• OpenAI instruct models
• Anthropic models (ANTHROPIC_API_KEY env variable)
• complete: debug option
• BaseModel.RENAME_KWARGS
• pip install
• limit handling

0.1

• parallel calls / cache / usage tracking / retries
• OpenAI chat models

Next

• HF API text generation
• llama.cpp models (GGML!)
• strangulate BaseModel with BaseModelV2

Reference Materials

• https://github.com/kagisearch/pyllms
• https://chat.lmsys.org/?leaderboard
• https://huggingface.co/spaces/HuggingFaceH4/open_llm_leaderboard
• https://huggingface.co/spaces/mteb/leaderboard
• https://huggingface.co/blog/getting-started-with-embeddings