QGML with Starcoder.cpp only needs less than 3GB GPU

#4
by DevElCuy - opened

First of all thanks for this great model!
I managed to convert it to GGML with https://github.com/bigcode-project/starcoder.cpp
And this is how I load it:

    from ctransformers import (
        AutoModelForCausalLM,
        AutoTokenizer
    )
    from transformers import pipeline
    from langchain.llms import HuggingFacePipeline

    model_name = "models/abacaj--starcoderbase-1b-sft-ggml.bin"
    model = AutoModelForCausalLM.from_pretrained(
        model_name,
        model_type="gpt_bigcode",
        gpu_layers=1024,
        hf=True
    )
    tokenizer = AutoTokenizer.from_pretrained(model)

    pipe = pipeline(
        "text-generation",
        model=model,
        tokenizer=tokenizer,
        max_length=2048,
    )

    llm = HuggingFacePipeline(
        pipeline=pipe,
    )

Dependencies (requirements.txt):

torch==2.0.1
transformers==4.33.1
langchain==0.0.285
ctransformers==0.2.26

Sign up or log in to comment