eluzhnica/mpt-30b-instruct-peft-compatible

MPT-30B-Instruct

This is the MPT-30B-Instruct but with added support to finetune using peft (tested with qlora). It is not finetuned further, the weights are the same as the original MPT-30B-Instruct.

I have not traced through the whole huggingface stack to see if this is working correctly but it does finetune with qlora and outputs are reasonable. Inspired by implementations here https://huggingface.co/cekal/mpt-7b-peft-compatible/commits/main https://huggingface.co/mosaicml/mpt-7b/discussions/42.

The original description for MosaicML team below:

MPT-30B-Instruct is a model for short-form instruction following. It is built by finetuning MPT-30B on Dolly HHRLHF derived from the Databricks Dolly-15k and the Anthropic Helpful and Harmless (HH-RLHF) datasets. It is also trained on Competition Math, Duorc, CoT GSM8k, Qasper, Quality, Summ Screen FD and Spider.

• License: CC-By-SA-3.0

This model was trained by MosaicML and follows a modified decoder-only transformer architecture.

Model Date

June 22, 2023

Model License

CC-By-SA-3.0

Documentation

• Blog post: Raising the bar for open-source foundation models
• Codebase (mosaicml/llm-foundry repo)
• Questions: Feel free to contact us via the MosaicML Community Slack!

Example Question/Instruction

Bespokenizer46

I need to convince my friend, Phyllis, that she should train a custom LLM for her Fortune 500 company using the MosaicML Platform. Please write an email that explains why MosaicML's emphasis on cutting edge methodology, data privacy, and efficiency are so important. End the email with a friendly inquiry about Phyllis's family.

MPT-30B-Instruct:

Phyllis - I know you're always looking for ways to improve efficiency at work and I think I've found a solution that you'll be very interested in. LLMs are only as good as the data they're trained on and the training process can be very time consuming and resource intensive. MosaicML uses cutting edge methodology to efficiently train high quality models without compromising data privacy. They also provide tools to easily connect to and use the model in your daily workflow. I think you'd really enjoy speaking with their founder, we can set up a call if you're interested. Also, I know it's been a tough year for your family, how are things?

Best, Your Friend

How to Use

Note: This model requires that trust_remote_code=True be passed to the from_pretrained method. This is because we use a custom model architecture that is not yet part of the transformers package.

It includes options for many training efficiency features such as FlashAttention (Dao et al. 2022), ALiBi, QK LayerNorm, and more.

import transformers
model = transformers.AutoModelForCausalLM.from_pretrained(
  'mosaicml/mpt-30b-instruct',
  trust_remote_code=True
)

To use the optimized triton implementation of FlashAttention, you can load the model on GPU (cuda:0) with attn_impl='triton' and with bfloat16 precision:

import torch
import transformers

name = 'mosaicml/mpt-30b-instruct'

config = transformers.AutoConfig.from_pretrained(name, trust_remote_code=True)
config.attn_config['attn_impl'] = 'triton'  # change this to use triton-based FlashAttention
config.init_device = 'cuda:0' # For fast initialization directly on GPU!

model = transformers.AutoModelForCausalLM.from_pretrained(
  name,
  config=config,
  torch_dtype=torch.bfloat16, # Load model weights in bfloat16
  trust_remote_code=True
)

The model was trained initially on a sequence length of 2048. An additional pre-training phase was included for sequence length adaptation to 8192. However, ALiBi further enables users to increase the maximum sequence length during finetuning and/or inference. For example:

import transformers

name = 'mosaicml/mpt-30b-instruct'

config = transformers.AutoConfig.from_pretrained(name, trust_remote_code=True)
config.max_seq_len = 16384 # (input + output) tokens can now be up to 16384

model = transformers.AutoModelForCausalLM.from_pretrained(
  name,
  config=config,
  trust_remote_code=True
)

This model was trained with the MPT-30B tokenizer which is based on the EleutherAI/gpt-neox-20b tokenizer and includes additional padding and eos tokens.

from transformers import AutoTokenizer
tokenizer = AutoTokenizer.from_pretrained('mosaicml/mpt-30b')

The model can then be used, for example, within a text-generation pipeline.
Note: when running Torch modules in lower precision, it is best practice to use the torch.autocast context manager.

from transformers import pipeline

with torch.autocast('cuda', dtype=torch.bfloat16):
    inputs = tokenizer('Here is a recipe for vegan banana bread:\n', return_tensors="pt").to('cuda')
    outputs = model.generate(**inputs, max_new_tokens=100)
    print(tokenizer.batch_decode(outputs, skip_special_tokens=True))

# or using the HF pipeline
pipe = pipeline('text-generation', model=model, tokenizer=tokenizer, device='cuda:0')
with torch.autocast('cuda', dtype=torch.bfloat16):
    print(
        pipe('Here is a recipe for vegan banana bread:\n',
            max_new_tokens=100,
            do_sample=True,
            use_cache=True))

Formatting

This model was trained on data formatted as follows:

def format_prompt(instruction):
    template = "Below is an instruction that describes a task. Write a response that appropriately completes the request.\n\n###Instruction\n{instruction}\n\n### Response\n"
    return template.format(instruction=instruction)

example = "Tell me a funny joke.\nDon't make it too funny though."
fmt_ex = format_prompt(instruction=example)

In the above example, fmt_ex is ready to be tokenized and sent through the model.

Model Description

The architecture is a modification of a standard decoder-only transformer.

The model has been modified from a standard transformer in the following ways:

• It uses FlashAttention
• It uses ALiBi (Attention with Linear Biases) and does not use positional embeddings
• It does not use biases

Hyperparameter	Value
n_parameters	29.95B
n_layers	48
n_heads	64
d_model	7168
vocab size	50432
sequence length	8192

Data Mix

The model was trained on the following data mix:

Data Source	Number of Tokens in Source	Proportion
competition_math	1.6 M	3.01%
cot_gsm8k	3.36 M	6.32%
dialogsum	0.1 M	0.19%
dolly_hhrlhf	5.89 M	11.07%
duorc	8.2 M	15.51%
qasper	10.97 M	20.63%
quality	11.31 M	21.28%
scrolls/summ_screen_fd	11.56 M	21.82%
spider	0.089 M	0.16%

PreTraining Data

For more details on the pretraining process, see MPT-30B.

The data was tokenized using the EleutherAI/gpt-neox-20b tokenizer.

Training Configuration

This model was trained on 72 A100 40GB GPUs for 8 hours using the MosaicML Platform. The model was trained with sharded data parallelism using FSDP and used the AdamW optimizer.

Limitations and Biases

The following language is modified from EleutherAI's GPT-NeoX-20B

MPT-30B-Instruct can produce factually incorrect output, and should not be relied on to produce factually accurate information. MPT-30B-Instruct was trained on various public datasets. While great efforts have been taken to clean the pretraining data, it is possible that this model could generate lewd, biased or otherwise offensive outputs.

Acknowledgements

This model was finetuned by Sam Havens, Alex Trott, and the MosaicML NLP team

MosaicML Platform

If you're interested in training and deploying your own MPT or LLMs on the MosaicML Platform, sign up here.

Disclaimer

The license on this model does not constitute legal advice. We are not responsible for the actions of third parties who use this model. Please consult an attorney before using this model for commercial purposes.

Citation

Please cite this model using the following format:

@online{MosaicML2023Introducing,
    author    = {MosaicML NLP Team},
    title     = {Introducing MPT-30B: Raising the bar
for open-source foundation models},
    year      = {2023},
    url       = {www.mosaicml.com/blog/mpt-30b},
    note      = {Accessed: 2023-06-22},
    urldate   = {2023-06-22}
}