TheBloke/bloomz-176B-GPTQ

Chat & support: my new Discord server

Want to contribute? TheBloke's Patreon page

BigScience's BLOOMZ 176B GPTQ

These files are GPTQ 4bit model files for BigScience's BLOOMZ.

It is the result of quantising to 4bit using AutoGPTQ.

This is a BIG model! 2 x 80GB or 3 x 48GB GPUs are required

Important note: files must be joined before use

It is not currently possible to shard GPTQ files, therefore the model file is one single 94 GB safetensors file.

Huggingface Hub has a 50GB per-file limit. I have therefore been forced to split the file in to three parts for upload.

I did this using the simple *nix command split.

To join the files on any *nix system, run:

cat gptq_model-4bit--1g.JOINBEFOREUSE.split-*.safetensors > gptq_model-4bit--1g.safetensors

To join the files on Windows, open a Command Prompt and run:

COPY /B gptq_model-4bit--1g.JOINBEFOREUSE.split-a.safetensors + gptq_model-4bit--1g.JOINBEFOREUSE.split-b.safetensors + gptq_model-4bit--1g.JOINBEFOREUSE.split-c.safetensors gptq_model-4bit--1g.safetensors

Or for Python code for joining the files, see the Python section below.

The SHA256SUM of the joined file will be:

50baeab9859362d22df6f822f158b9ba75b44ffc6605b715992fe6245aa6e93a  gptq_model-4bit--1g.safetensors

Once you have the joined file, you can safely delete gptq_model-4bit--1g.JOINBEFOREUSE.split-*.safetensors.

Repositories available

• 4-bit GPTQ model for GPU inference
• Unquantised fp16 model in pytorch format, for GPU inference and for further conversions

Two files provided - separate branches

• Main branch: gptq_model-4bit--1g.safetensors
  • Group Size = None
  • Desc Act (act-order) = True
  • This version will use the least possible VRAM, and should have higher inference performance in CUDA mode
• Branch group_size_128g: gptq_model-4bit-128g.safetensors
  • Group Size = 128g
  • Desc Act (act-oder) = True
  • This version will use more VRAM, which shouldn't be a problem as it shouldn't exceed 2 x 80GB or 3 x 48GB cards.
  • However CUDA inference performance is likely to be a lot slower, possibly necessitating the use of Triton mode.

By default you will download the first file, unless you choose to download from branch group_size_128g.

Prompt template: none

Translate to English: Je t’aime.
Translation:

How to easily download and use this model in text-generation-webui

Please make sure you're using the latest version of text-generation-webui.

Note 1: this is a non-Llama model which cannot be used with ExLlama. Use Loader: AutoGPTQ.

Note 2: As described above, you must join the files after downloading and before loading in text-generation-webui.

1. Click the Model tab.
2. Under Download custom model or LoRA, enter TheBloke/bloomz-176B-GPTQ.

• If you would rather download the group_size 128g version, enter TheBloke/bloomz-176B-GPTQ:group_size_128g

3. Click Download.
4. The model will start downloading. Once it's finished it will say "Done". This is a huge model so it may take a while!
5. Now follow the steps described above to join the model to get a single .safetensors file.
6. Untick Autoload model
7. In the top left, click the refresh icon next to Model.
8. In the Model dropdown, choose the model you just downloaded: bloomz-176B-GPTQ
9. Make sure Loader is set to AutGPTQ.
10. This model cannot load on one GPU, so you should set GPU Memory accordingly.

• If using two 80GB GPUs, try: GPU0 = 60GB, GPU1 = 79GB
• If using three 48GB GPUs, try: GPU0 = 30GB, GPU1 = 47GB, GPU2 = 47GB

11. Click Save settings to save your settings, and then Reload to load the model.
12. The model will load, and is now ready for use!
13. Once you're ready, click the Text Generation tab and enter a prompt to get started!

How to use this GPTQ model from Python code

First make sure you have AutoGPTQ installed:

GITHUB_ACTIONS=true pip install auto-gptq

Because this model has to be joined locally, you must first download it. Example download code:

from huggingface_hub import snapshot_download
snapshot_download(repo_id="TheBloke/bloomz-176B-GPTQ",
  local_dir="/workspace/models/bloomz-176GB-GPTQ",
  local_dir_use_symlinks=False)

If you want to download the group_size 128g file instead, add revision="group_size_128g" to the above command.

Now join the three split files, which can be done with the following Python code:

import glob

# Get the list of all files matching the pattern
files = sorted(glob.glob('gptq_model-4bit--1g.JOINBEFOREUSE.split-*.safetensors'))

# Open the output file in binary write mode
with open('gptq_model-4bit--1g.safetensors', 'wb') as outfile:
    for filename in files:
        with open(filename, 'rb') as infile:
            outfile.write(infile.read())

Then try the following example code:

from transformers import AutoTokenizer, pipeline, logging
from auto_gptq import AutoGPTQForCausalLM, BaseQuantizeConfig
import argparse

# Use the local path you downloaded the model to and joined the split files in
model_name_or_path = "/workspace/models/bloomz-176GB-GPTQ"
model_basename = "gptq_model-4bit--1g"

use_triton = False

tokenizer = AutoTokenizer.from_pretrained(model_name_or_path, use_fast=True)

model = AutoGPTQForCausalLM.from_quantized(model_name_or_path,
        model_basename=model_basename,
        max_memory={0: '60GiB', 1: '79GiB'} # max_memory is for 2 x 80GB GPUs; adjust if your config is different!
        use_safetensors=True,
        trust_remote_code=False,
        use_triton=use_triton,
        quantize_config=None)

prompt = "Translate this to French: AI is the future of computing"
prompt_template=f'''{prompt}
Translation:
'''

print("\n\n*** Generate:")

input_ids = tokenizer(prompt_template, return_tensors='pt').input_ids.cuda()
output = model.generate(inputs=input_ids, temperature=0.7, max_new_tokens=512)
print(tokenizer.decode(output[0]))

# Inference can also be done using transformers' pipeline

# Prevent printing spurious transformers error when using pipeline with AutoGPTQ
logging.set_verbosity(logging.CRITICAL)

print("*** Pipeline:")
pipe = pipeline(
    "text-generation",
    model=model,
    tokenizer=tokenizer,
    max_new_tokens=512,
    temperature=0.7,
    top_p=0.95,
    repetition_penalty=1.15
)

print(pipe(prompt_template)[0]['generated_text'])

Provided files

Main branch:

gptq_model-4bit--1g.safetensors

This will work with AutoGPTQ. It is untested with GPTQ-for-LLaMa. It will not work with ExLlama.

It was created with group_size none (-1) to reduce VRAM usage, and with --act-order (desc_act) to improve accuracy of responses.

• gptq_model-4bit-128g.safetensors
  • Works with AutoGPTQ in CUDA or Triton modes.
  • Does NOT work with ExLlama as it's not a Llama model.
  • Untested with GPTQ-for-LLaMa.
  • Works with text-generation-webui, including one-click-installers.
  • Parameters: Groupsize = -1. Act Order / desc_act = True.

Branch group_size_128g

gptq_model-4bit-128g.safetensors

This will work with AutoGPTQ. It is untested with GPTQ-for-LLaMa. It will not work with ExLlama.

It was created with both group_size 128g and --act-order (desc_act) for even higher inference accuracy, at the cost of increased VRAM usage. Because we already need 2 x 80GB or 3 x 48GB GPUs, I don't expect the increased VRAM usage to change the GPU requirements.

Note Using group_size + desc_act together can significantly lower performance in AutoGPTQ CUDA. You might want to try AutoGPTQ Triton mode instead (Linux only.)

• gptq_model-4bit-128g.safetensors
  • Works with AutoGPTQ in CUDA or Triton modes.
  • Does NOT work with ExLlama as it's not a Llama model.
  • Untested with GPTQ-for-LLaMa.
  • Works with text-generation-webui, including one-click-installers.
  • Parameters: Groupsize = 128. Act Order / desc_act = True.

Discord

For further support, and discussions on these models and AI in general, join us at:

TheBloke AI's Discord server

Thanks, and how to contribute.

Thanks to the chirper.ai team!

I've had a lot of people ask if they can contribute. I enjoy providing models and helping people, and would love to be able to spend even more time doing it, as well as expanding into new projects like fine tuning/training.

If you're able and willing to contribute it will be most gratefully received and will help me to keep providing more models, and to start work on new AI projects.

Donaters will get priority support on any and all AI/LLM/model questions and requests, access to a private Discord room, plus other benefits.

• Patreon: https://patreon.com/TheBlokeAI
• Ko-Fi: https://ko-fi.com/TheBlokeAI

Special thanks to: Luke from CarbonQuill, Aemon Algiz, Dmitriy Samsonov.

Patreon special mentions: zynix , ya boyyy, Trenton Dambrowitz, Imad Khwaja, Alps Aficionado, chris gileta, John Detwiler, Willem Michiel, RoA, Mano Prime, Rainer Wilmers, Fred von Graf, Matthew Berman, Ghost , Nathan LeClaire, Iucharbius , Ai Maven, Illia Dulskyi, Joseph William Delisle, Space Cruiser, Lone Striker, Karl Bernard, Eugene Pentland, Greatston Gnanesh, Jonathan Leane, Randy H, Pierre Kircher, Willian Hasse, Stephen Murray, Alex , terasurfer , Edmond Seymore, Oscar Rangel, Luke Pendergrass, Asp the Wyvern, Junyu Yang, David Flickinger, Luke, Spiking Neurons AB, subjectnull, Pyrater, Nikolai Manek, senxiiz, Ajan Kanaga, Johann-Peter Hartmann, Artur Olbinski, Kevin Schuppel, Derek Yates, Kalila, K, Talal Aujan, Khalefa Al-Ahmad, Gabriel Puliatti, John Villwock, WelcomeToTheClub, Daniel P. Andersen, Preetika Verma, Deep Realms, Fen Risland, trip7s trip, webtim, Sean Connelly, Michael Levine, Chris McCloskey, biorpg, vamX, Viktor Bowallius, Cory Kujawski.

Thank you to all my generous patrons and donaters!

Original model card: BigScience's BLOOMZ 176B

Table of Contents

1. Model Summary
2. Use
3. Limitations
4. Training
5. Evaluation
6. Citation

Model Summary

We present BLOOMZ & mT0, a family of models capable of following human instructions in dozens of languages zero-shot. We finetune BLOOM & mT5 pretrained multilingual language models on our crosslingual task mixture (xP3) and find the resulting models capable of crosslingual generalization to unseen tasks & languages.

• Repository: bigscience-workshop/xmtf
• Paper: Crosslingual Generalization through Multitask Finetuning
• Point of Contact: Niklas Muennighoff
• Languages: Refer to bloom for pretraining & xP3 for finetuning language proportions. It understands both pretraining & finetuning languages.
• BLOOMZ & mT0 Model Family:

Multitask finetuned on xP3. Recommended for prompting in English.
Parameters	300M	580M	1.2B	3.7B	13B	560M	1.1B	1.7B	3B	7.1B	176B
Finetuned Model	mt0-small	mt0-base	mt0-large	mt0-xl	mt0-xxl	bloomz-560m	bloomz-1b1	bloomz-1b7	bloomz-3b	bloomz-7b1	bloomz
Multitask finetuned on xP3mt. Recommended for prompting in non-English.
Finetuned Model					mt0-xxl-mt					bloomz-7b1-mt	bloomz-mt
Multitask finetuned on P3. Released for research purposes only. Strictly inferior to above models!
Finetuned Model					mt0-xxl-p3					bloomz-7b1-p3	bloomz-p3
Original pretrained checkpoints. Not recommended.
Pretrained Model	mt5-small	mt5-base	mt5-large	mt5-xl	mt5-xxl	bloom-560m	bloom-1b1	bloom-1b7	bloom-3b	bloom-7b1	bloom

Use

Intended use

We recommend using the model to perform tasks expressed in natural language. For example, given the prompt "Translate to English: Je t’aime.", the model will most likely answer "I love you.". Some prompt ideas from our paper:

• 一个传奇的开端，一个不灭的神话，这不仅仅是一部电影，而是作为一个走进新时代的标签，永远彪炳史册。你认为这句话的立场是赞扬、中立还是批评?
• Suggest at least five related search terms to "Mạng neural nhân tạo".
• Write a fairy tale about a troll saving a princess from a dangerous dragon. The fairy tale is a masterpiece that has achieved praise worldwide and its moral is "Heroes Come in All Shapes and Sizes". Story (in Spanish):
• Explain in a sentence in Telugu what is backpropagation in neural networks.

Feel free to share your generations in the Community tab!

How to use

CPU

Click to expand

# pip install -q transformers
from transformers import AutoModelForCausalLM, AutoTokenizer

checkpoint = "bigscience/bloomz"

tokenizer = AutoTokenizer.from_pretrained(checkpoint)
model = AutoModelForCausalLM.from_pretrained(checkpoint)

inputs = tokenizer.encode("Translate to English: Je t’aime.", return_tensors="pt")
outputs = model.generate(inputs)
print(tokenizer.decode(outputs[0]))

GPU

Click to expand

# pip install -q transformers accelerate
from transformers import AutoModelForCausalLM, AutoTokenizer

checkpoint = "bigscience/bloomz"

tokenizer = AutoTokenizer.from_pretrained(checkpoint)
model = AutoModelForCausalLM.from_pretrained(checkpoint, torch_dtype="auto", device_map="auto")

inputs = tokenizer.encode("Translate to English: Je t’aime.", return_tensors="pt").to("cuda")
outputs = model.generate(inputs)
print(tokenizer.decode(outputs[0]))

GPU in 8bit

Click to expand

# pip install -q transformers accelerate bitsandbytes
from transformers import AutoModelForCausalLM, AutoTokenizer

checkpoint = "bigscience/bloomz"

tokenizer = AutoTokenizer.from_pretrained(checkpoint)
model = AutoModelForCausalLM.from_pretrained(checkpoint, device_map="auto", load_in_8bit=True)

inputs = tokenizer.encode("Translate to English: Je t’aime.", return_tensors="pt").to("cuda")
outputs = model.generate(inputs)
print(tokenizer.decode(outputs[0]))

Limitations

Prompt Engineering: The performance may vary depending on the prompt. For BLOOMZ models, we recommend making it very clear when the input stops to avoid the model trying to continue it. For example, the prompt "Translate to English: Je t'aime" without the full stop (.) at the end, may result in the model trying to continue the French sentence. Better prompts are e.g. "Translate to English: Je t'aime.", "Translate to English: Je t'aime. Translation:" "What is "Je t'aime." in English?", where it is clear for the model when it should answer. Further, we recommend providing the model as much context as possible. For example, if you want it to answer in Telugu, then tell the model, e.g. "Explain in a sentence in Telugu what is backpropagation in neural networks.".

Training

Model

• Architecture: Same as bloom, also refer to the config.json file
• Finetuning steps: 498
• Finetuning tokens: 2.09 billion
• Finetuning layout: 72x pipeline parallel, 1x tensor parallel, 4x data parallel
• Precision: bfloat16

Hardware

• CPUs: AMD CPUs with 512GB memory per node
• GPUs: 288 A100 80GB GPUs with 8 GPUs per node (36 nodes) using NVLink 4 inter-gpu connects, 4 OmniPath links
• Communication: NCCL-communications network with a fully dedicated subnet

Software

• Orchestration: Megatron-DeepSpeed
• Optimizer & parallelism: DeepSpeed
• Neural networks: PyTorch (pytorch-1.11 w/ CUDA-11.5)
• FP16 if applicable: apex

Evaluation

We refer to Table 7 from our paper & bigscience/evaluation-results for zero-shot results on unseen tasks. The sidebar reports zero-shot performance of the best prompt per dataset config.

Citation

@article{muennighoff2022crosslingual,
  title={Crosslingual generalization through multitask finetuning},
  author={Muennighoff, Niklas and Wang, Thomas and Sutawika, Lintang and Roberts, Adam and Biderman, Stella and Scao, Teven Le and Bari, M Saiful and Shen, Sheng and Yong, Zheng-Xin and Schoelkopf, Hailey and others},
  journal={arXiv preprint arXiv:2211.01786},
  year={2022}
}