ctranslate2 int8 float16

# Fast-Inference with Ctranslate2

Speedup inference while reducing memory by 2x-4x using int8 inference in C++ on CPU or GPU.

quantized version of togethercomputer/Pythia-Chat-Base-7B

pip install hf-hub-ctranslate2>=2.0.8

Converted on 2023-05-22 using

ct2-transformers-converter --model togethercomputer/Pythia-Chat-Base-7B --output_dir /home/michael/tmp-ct2fast-Pythia-Chat-Base-7B --force --copy_files tokenizer.json README.md tokenizer_config.json special_tokens_map.json .gitattributes --quantization float16

Checkpoint compatible to ctranslate2>=3.13.0 and hf-hub-ctranslate2>=2.0.6

from hf_hub_ctranslate2 import TranslatorCT2fromHfHub, GeneratorCT2fromHfHub
from transformers import AutoTokenizer

model_name = "michaelfeil/ct2fast-Pythia-Chat-Base-7B"
# use either TranslatorCT2fromHfHub or GeneratorCT2fromHfHub here, depending on model.
model = GeneratorCT2fromHfHub(
        # load in int8 on CUDA
        model_name_or_path=model_name, 
        device="cuda",
        compute_type="int8_float16",
        # tokenizer=AutoTokenizer.from_pretrained("togethercomputer/Pythia-Chat-Base-7B")
)
outputs = model.generate(
    text=["def print_hello_world():", "def hello_name(name:"],
    max_length=64
)
print(outputs)

Licence and other remarks:

This is just a quantized version. Licence conditions are intended to be idential to original huggingface repo.

Original description

<p style="font-size: 24px">Feel free to try out our OpenChatKit feedback app!</p>

Pythia-Chat-Base-7B-v0.16

TLDR: As part of OpenChatKit (codebase available here), Pythia-Chat-Base-7B-v0.16 is a 7B parameter language model, fine-tuned from EleutherAI’s Pythia 7B with over 40 million instructions on 100% carbon negative compute.

Pythia-Chat-Base-7B-v0.16 is based on ElutherAI’s Pythia-7B model, and is fine-tuned with data focusing on dialog-style interactions. We focused the tuning on several tasks such as question answering, classification, extraction, and summarization. We’ve fine-tuned the model with a collection of 43 million high-quality instructions. Together partnered with LAION and Ontocord.ai, who both helped curate the dataset the model is based on. You can read more about this process and the availability of this dataset in LAION’s blog post here.

In addition to the aforementioned fine-tuning, Pythia-Chat-Base-7B-v0.16 has also undergone further fine-tuning via a small amount of feedback data. This process allows the model to better adapt to human preferences in the conversations.

One of the notable features of Pythia-Chat-Base-7B-v0.16 is its ability to run inference on a 12GB GPU, thanks to the quantization technique. It helps maintain the dialogue capabilities while making the model more accessible to a wider range of users and hardware configurations.

Model Details

Quick Start

GPU Inference

This requires a GPU with 24GB memory.

from transformers import AutoTokenizer, AutoModelForCausalLM

# init
tokenizer = AutoTokenizer.from_pretrained("togethercomputer/Pythia-Chat-Base-7B-v0.16")
model = AutoModelForCausalLM.from_pretrained("togethercomputer/Pythia-Chat-Base-7B-v0.16", torch_dtype=torch.float16)
model = model.to('cuda:0')

# infer
inputs = tokenizer("<human>: Hello!\n<bot>:", return_tensors='pt').to(model.device)
outputs = model.generate(**inputs, max_new_tokens=10, do_sample=True, temperature=0.8)
output_str = tokenizer.decode(outputs[0])
print(output_str)

GPU Inference in Int8

This requires a GPU with 12GB memory.

from transformers import AutoTokenizer, AutoModelForCausalLM

# init
tokenizer = AutoTokenizer.from_pretrained("togethercomputer/Pythia-Chat-Base-7B-v0.16")
model = AutoModelForCausalLM.from_pretrained("togethercomputer/Pythia-Chat-Base-7B-v0.16", device_map="auto", load_in_8bit=True)

# infer
inputs = tokenizer("<human>: Hello!\n<bot>:", return_tensors='pt').to(model.device)
outputs = model.generate(**inputs, max_new_tokens=10, do_sample=True, temperature=0.8)
output_str = tokenizer.decode(outputs[0])
print(output_str)

CPU Inference

from transformers import AutoTokenizer, AutoModelForCausalLM

# init
tokenizer = AutoTokenizer.from_pretrained("togethercomputer/Pythia-Chat-Base-7B-v0.16")
model = AutoModelForCausalLM.from_pretrained("togethercomputer/Pythia-Chat-Base-7B-v0.16", torch_dtype=torch.bfloat16)

# infer
inputs = tokenizer("<human>: Hello!\n<bot>:", return_tensors='pt').to(model.device)
outputs = model.generate(**inputs, max_new_tokens=10, do_sample=True, temperature=0.8)
output_str = tokenizer.decode(outputs[0])
print(output_str)

Strengths of the model

There are several tasks that OpenChatKit excels at out of the box. This includes:

In addition, the model does well on few-shot prompts. For both classification and extraction, the model performs even better with few shots, as in most HELM tasks. Contact us if you’re interested in trying few-shot prompts with the model.

Weaknesses of the model

That said, there are several areas where we have more work to do, and we need your help! Some of these include:

We are excited to work with you to address these weaknesses by getting your feedback, bolstering data sets, and improving accuracy.

Uses

Direct Use

The model is intended for research purposes. Possible research areas and tasks include

Excluded uses are described below.

Misuse, Malicious Use, and Out-of-Scope Use

The OpenChatKit community provides Pythia-Chat-Base-7B-v0.16 as an open source tool for building chatbots. The community is not responsible for any misuse, malicious use, or out-of-scope use of the model. It is the responsibility of the end user to ensure that the model is used in a responsible and ethical manner.

Out-of-Scope Use

Pythia-Chat-Base-7B-v0.16 is designed for use in chatbot applications and may not perform well for other use cases outside of its intended scope. For example, it may not be suitable for use in safety-critical applications or for making decisions that have a significant impact on individuals or society. It is important to consider the limitations of the model and to only use it for its intended purpose.

Misuse and Malicious Use

Pythia-Chat-Base-7B-v0.16 is designed for use in chatbot applications and should not be used for any other purpose. Misuse of the model, such as using it to engage in illegal or unethical activities, is strictly prohibited and goes against the principles of the OpenChatKit community project.

Using the model to generate content that is cruel to individuals is a misuse of this model. This includes, but is not limited to:

Limitations

Pythia-Chat-Base-7B-v0.16, like other language model-based chatbots, has limitations that should be taken into consideration. For example, the model may not always provide accurate or relevant answers, particularly for questions that are complex, ambiguous, or outside of its training data. We therefore welcome contributions from individuals and organizations, and encourage collaboration towards creating a more robust and inclusive chatbot.

Training

Training Data

Please refer to togethercomputer/OpenDataHub

Training Procedure

Community

Join us on Together Discord