Note: This model requires that trust_remote_code=True be passed to the from_pretrained method.
This is because we use a custom MPT model architecture that is not yet part of the Hugging Face transformers package.
MPT includes options for many training efficiency features such as FlashAttention, ALiBi, QK LayerNorm, and more.
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:
python
1import torch
2import transformers
34name ='mosaicml/mpt-7b-chat-8k'56config = transformers.AutoConfig.from_pretrained(name, trust_remote_code=True)7config.attn_config['attn_impl']='triton'# change this to use triton-based FlashAttention8config.init_device ='cuda:0'# For fast initialization directly on GPU!910model = transformers.AutoModelForCausalLM.from_pretrained(11 name,12 config=config,13 torch_dtype=torch.bfloat16,# Load model weights in bfloat1614 trust_remote_code=True15)
The model was trained initially with a sequence length of 2048 with an additional pretraining stage for sequence length adapation up to 8192. However, ALiBi enables users to increase the maximum sequence length even further during finetuning and/or inference. For example:
python
1import transformers
23name ='mosaicml/mpt-7b-chat-8k'45config = transformers.AutoConfig.from_pretrained(name, trust_remote_code=True)6config.max_seq_len =16384# (input + output) tokens can now be up to 1638478model = transformers.AutoModelForCausalLM.from_pretrained(9 name,10 config=config,11 trust_remote_code=True12)
This model was trained with the MPT-7B-chat tokenizer which is based on the EleutherAI/gpt-neox-20b tokenizer and includes additional ChatML tokens.
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.
python
1from transformers import pipeline
23with torch.autocast('cuda', dtype=torch.bfloat16):4 inputs = tokenizer('Here is a recipe for vegan banana bread:\n', return_tensors="pt").to('cuda')5 outputs = model.generate(**inputs, max_new_tokens=100)6print(tokenizer.batch_decode(outputs, skip_special_tokens=True))78# or using the HF pipeline9pipe = pipeline('text-generation', model=model, tokenizer=tokenizer, device='cuda:0')10with torch.autocast('cuda', dtype=torch.bfloat16):11print(12 pipe('Here is a recipe for vegan banana bread:\n',13 max_new_tokens=100,14 do_sample=True,15 use_cache=True))
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:
"LongConversations" is a GPT3.5/4-generated dataset, details of which will be released at a later date.
Training Configuration
This model was trained on 192 H100s for about 48 minutes using the MosaicML Platform.
The model was trained with sharded data parallelism using FSDP and used the AdamW optimizer.
MPT-7B-Chat-8k can produce factually incorrect output, and should not be relied on to produce factually accurate information.
MPT-7B-Chat-8k 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 the MosaicML NLP team
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.
