llm-rs-metal 0.2.15

Unofficial python bindings for llm-rs. 🐍❤️🦀

llm-rs-python: Python Bindings for Rust's llm Library

Welcome to llm-rs, an unofficial Python interface for the Rust-based llm library, made possible through PyO3. Our package combines the convenience of Python with the performance of Rust to offer an efficient tool for your machine learning projects. 🐍❤️🦀

With llm-rs, you can operate a variety of Large Language Models (LLMs) including LLama and GPT-NeoX directly on your CPU or GPU.

For a detailed overview of all the supported architectures, visit the llm project page.

Integrations:

• 🦜️🔗 LangChain
• 🌾🔱 Haystack

Installation

Simply install it via pip: pip install llm-rs

Installation with GPU Acceleration Support

llm-rs incorporates support for various GPU-accelerated backends to facilitate enhanced inference times. To enable GPU-acceleration the use_gpu parameter of your SessionConfig must be set to True. The llm documentation lists all model architectures, which are currently accelerated. We distribute prebuilt binaries for the following operating systems and graphics APIs:

MacOS (Using Metal)

For MacOS users, the Metal-supported version of llm-rs can be easily installed via pip:

pip install llm-rs-metal

Windows/Linux (Using CUDA for Nvidia GPUs)

Due to the significant file size, CUDA-supported packages cannot be directly uploaded to pip. To install them, download the appropriate *.whl file from the latest Release and install it using pip as follows:

pip install [wheelname].whl

Windows/Linux (Using OpenCL for All GPUs)

For universal GPU support on Windows and Linux, we offer an OpenCL-supported version. It can be installed via pip:

pip install llm-rs-opencl

Usage

Running local GGML models:

Models can be loaded via the AutoModel interface.

from llm_rs import AutoModel, KnownModels

#load the model
model = AutoModel.from_pretrained("path/to/model.bin",model_type=KnownModels.Llama)

#generate
print(model.generate("The meaning of life is"))

Streaming Text

Text can be yielded from a generator via the stream function:

from llm_rs import AutoModel, KnownModels

#load the model
model = AutoModel.from_pretrained("path/to/model.bin",model_type=KnownModels.Llama)

#generate
for token in model.stream("The meaning of life is"):
    print(token)

Running GGML models from the Hugging Face Hub

GGML converted models can be directly downloaded and run from the hub.

from llm_rs import AutoModel

model = AutoModel.from_pretrained("rustformers/mpt-7b-ggml",model_file="mpt-7b-q4_0-ggjt.bin")

If there are multiple models in a repo the model_file has to be specified. If you want to load repositories which were not created throught this library, you have to specify the model_type parameter as the metadata files needed to infer the architecture are missing.

Running Pytorch Transfomer models from the Hugging Face Hub

llm-rs supports automatic conversion of all supported transformer architectures on the Huggingface Hub.

To run covnersions additional dependencies are needed which can be installed via pip install llm-rs[convert].

The models can then be loaded and automatically converted via the from_pretrained function.

from llm_rs import AutoModel

model = AutoModel.from_pretrained("mosaicml/mpt-7b")

Convert Huggingface Hub Models

The following example shows how a Pythia model can be covnverted, quantized and run.

from llm_rs.convert import AutoConverter
from llm_rs import AutoModel, AutoQuantizer
import sys

#define the model which should be converted and an output directory
export_directory = "path/to/directory" 
base_model = "EleutherAI/pythia-410m"

#convert the model
converted_model = AutoConverter.convert(base_model, export_directory)

#quantize the model (this step is optional)
quantized_model = AutoQuantizer.quantize(converted_model)

#load the quantized model
model = AutoModel.load(quantized_model,verbose=True)

#generate text
def callback(text):
    print(text,end="")
    sys.stdout.flush()

model.generate("The meaning of life is",callback=callback)

🦜️🔗 LangChain Usage

Utilizing llm-rs-python through langchain requires additional dependencies. You can install these using pip install llm-rs[langchain]. Once installed, you gain access to the RustformersLLM model through the llm_rs.langchain module. This particular model offers features for text generation and embeddings.

Consider the example below, demonstrating a straightforward LLMchain implementation with MPT-Instruct:

from llm_rs.langchain import RustformersLLM
from langchain import PromptTemplate
from langchain.chains import LLMChain
from langchain.callbacks.streaming_stdout import StreamingStdOutCallbackHandler

template="""Below is an instruction that describes a task. Write a response that appropriately completes the request.
### Instruction:
{instruction}
### Response:
Answer:"""

prompt = PromptTemplate(input_variables=["instruction"],template=template,)

llm = RustformersLLM(model_path_or_repo_id="rustformers/mpt-7b-ggml",model_file="mpt-7b-instruct-q5_1-ggjt.bin",callbacks=[StreamingStdOutCallbackHandler()])

chain = LLMChain(llm=llm, prompt=prompt)

chain.run("Write a short post congratulating rustformers on their new release of their langchain integration.")

🌾🔱 Haystack Usage

Utilizing llm-rs-python through haystack requires additional dependencies. You can install these using pip install llm-rs[haystack]. Once installed, you gain access to the RustformersInvocationLayer model through the llm_rs.haystack module. This particular model offers features for text generation.

Consider the example below, demonstrating a straightforward Haystack-Pipeline implementation with OpenLLama-3B:

from haystack.nodes import PromptNode, PromptModel
from llm_rs.haystack import RustformersInvocationLayer

model = PromptModel("rustformers/open-llama-ggml",
                    max_length=1024,
                    invocation_layer_class=RustformersInvocationLayer,
                    model_kwargs={"model_file":"open_llama_3b-q5_1-ggjt.bin"})

pn = PromptNode(
    model,
    max_length=1024
)

pn("Write me a short story about a lama riding a crab.",stream=True)

Documentation

For in-depth information on customizing the loading and generation processes, refer to our detailed documentation.