This repository contains bunch of autoregressive transformer language models trained on a huge dataset of russian language.
Russian GPT-3 models (ruGPT3XL, ruGPT3Large, ruGPT3Medium, ruGPT3Small) trained with 2048 sequence length with sparse and dense attention blocks. We also provide Russian GPT-2 large model (ruGPT2Large) trained with 1024 sequence length.
We suggest using ruGPT2Large or ruGPT3XL because this models are well tested and achieve the best perplexity.
Usage examples are described in detail here.
Old version of code you can find here
- Setup and usage
- Pretraining details
- Advanced
Models can be used for inference or finetuning with two ways: 🤗HuggingFace interface or our code based on this implementation.
For both ways install transformers:
pip install transformers==3.5.0We support 🤗HuggingFace interface only for ruGPT3Large, ruGPT3Medium, ruGPT3Small and ruGPT2Large models. For RuGPT3XL please use code in this repo because RuGPT3XL model was trained with sparse attention.
Here we can obtain examples of finetuning or generation.
Also this examples is adapted for google colab:
- finetuning:
- generation:
Basic usage:
from transformers import GPT2LMHeadModel, GPT2Tokenizer
model_name_or_path = "sberbank-ai/rugpt3large_based_on_gpt2"
tokenizer = GPT2Tokenizer.from_pretrained(model_name_or_path)
model = GPT2LMHeadModel.from_pretrained(model_name_or_path).cuda()
text = "Александр Сергеевич Пушкин родился в "
input_ids = tokenizer.encode(text, return_tensors="pt").cuda()
out = model.generate(input_ids.cuda())
generated_text = list(map(tokenizer.decode, out))[0]
print(generated_text)
# Output should be like this:
# Александр Сергеевич Пушкин родился в \n1799 году. Его отец был крепостным крестьянином, а мать – крепостной крестьянкой. Детство и юность Пушкина прошли в деревне Михайловское под Петербургом. В 1820-х годах семья переехалаFor more information about 🤗HuggingFace interface please follow this documentation.
For training pass single txt file.
For using our code for finetuning without deepspeed (not recommended) we should install apex:
%%writefile setup.sh
export CUDA_HOME=/usr/local/cuda-10.1
git clone https://github.com/NVIDIA/apex
pip install -v --no-cache-dir --global-option="--cpp_ext" --global-option="--cuda_ext" ./apex
sh setup.shExample of finetuning, generating and loading/convert megatron checkpoints here or
Note! This way is valid for all RuGPTs models except RuGPT3XL.
For using our code for finetuning with deepspeed (recommended) we should install apex (see previous section) and deepspeed:
pip install deepspeed==0.3.7Example of finetuning, generating and loading/convert megatron checkpoints here or
Note! For using deepspeed we should specify environ variable before all your python scripts and run with torch.distributed or mpi:
USE_DEEPSPEED=1 python -m torch.distributed.launch --nproc_per_node 1 ru-gpts/pretrain_gpt3.py \
--train-data-path "train.list" \
--test-data-path "valid.list" \
--max-files-per-process 100 \
--save model \
--load-huggingface sberbank-ai/rugpt3small_based_on_gpt2 \
--model-parallel-size 1 \
--num-layers 12 \
--hidden-size 768 \
--num-attention-heads 12 \
--seq-length 2048 \
--max-position-embeddings 2048 \
--fp16 \
--checkpoint-activations \
--deepspeed-activation-checkpointing \
--deepspeed \
--deepspeed_config ru-gpts/src/deepspeed_config/gpt3_small_2048.json
We use custom implementation of distributed dataset. For training and evaluating we should specify file file.list with list of paths to txt files. All files from file.list will be splitted between aviable GPUs. The logic of splitting is described by the following code:
shard_size = len(files) // world_size
shard_start = rank * shard_size
shard_end = (rank + 1) * shard_size
files = files[shard_start:shard_end]For more details please see full code of dataset: src.dataset_rugpt3.RuGpt3TextDataset and example.
Note! This way is valid for all RuGPTs models except RuGPT3XL.
This section is used mostly for usage of RuGPT3XL model and training models with sparse attention.
apt-get install llvm-9-dev
pip install cpufeature
pip install triton==0.2.3
DS_BUILD_CPU_ADAM=1 DS_BUILD_SPARSE_ATTN=1 pip install deepspeed==0.3.7Test installation of deepspeed you can with the following command: ds_report.
Example of inference of RuGPT3XL here or
For using sparse layers in model use --sparse-mode <mode> and specify key "sparse_attention" at deepspeed_config (RuGPT3XL config example). Modes can be: fixed, bigbird, bslongformer, variable, dense.
More information about sparse attention here.
All pretraining was done on Nvidia Tesla V100-SXM3 32 Gb GPUs on a Christofari Cluster. Following are the details of pretraining for each model.
Model was trained with 512 sequence length using Deepspeed and Megatron code by SberDevices team, on 80B tokens dataset for 4 epochs. After that model was finetuned 1 epoch with sequence length 2048.
Note! Model has sparse attention blocks.
Total training time was around 10 days on 256 GPUs.
Final perplexity on test set is 12.05.
🤗HuggingFace model card link.
See more details for generation here or 
Our pretraining script here
Model was trained with sequence length 1024 using transformers lib by SberDevices team on 80B tokens for 3 epochs. After that model was finetuned 1 epoch with sequence length 2048.
Total training time was around 14 days on 128 GPUs for 1024 context and few days on 16 GPUs for 2048 context.
Final perplexity on test set is 13.6.
You can obtain this model by using transformers with model name sberbank-ai/rugpt3large_based_on_gpt2.
🤗HuggingFace model card link
Our pretraining script here
Model was trained with sequence length 1024 using transformers lib by SberDevices team on 80B tokens for 3 epoch. After that model was finetuned on 2048 context.
Total training time was around 16 days on 64 GPUs.
Final perplexity on test set is 17.4.
You can obtain this model by using transformers with model name sberbank-ai/rugpt3medium_based_on_gpt2.
🤗HuggingFace model card link
Our pretraining script here
Model was trained with sequence length 1024 using transformers by SberDevices team on 80B tokens around 3 epoch. After that model was finetuned on 2048 context.
Total training time took around one week on 32 GPUs.
You can obtain this model by using transformers with model name sberbank-ai/rugpt3small_based_on_gpt2.
🤗HuggingFace model card link
Our pretraining script here
Model was trained with sequence length 1024 using transformers by SberDevices team on 170Gb data on 64 GPUs 3 weeks.
You can obtain this model by using transformers with model name sberbank-ai/rugpt2large.
🤗HuggingFace model card link
Also we add pretraining scripts for all models (except RuGPT2Large). See scripts dir.
Note! All training params (such as lr, wd, ...) may was different while real training. This is just for example.
For converting megatron checkpoint to HuggingFace format use the following script (example for RuGPT3Small):
python convert2huggingface.py \
--load /path/to/save/dir/ \
--model-parallel-size 1 \
--num-layers 12 \
--hidden-size 768 \
--num-attention-heads 12 \
--max-position-embeddings 2048 \
--tokenizer-path sberbank-ai/rugpt3small_based_on_gpt2 \
--no-load-optim \
--export-huggingface /path/to/converted/checkpointAfter converting we can use HuggingFace model:
from transformers import GPT2LMHeadModel
model = GPT2LMHeadModel.from_pretrained("/path/to/converted/checkpoint")Note! Conversion is worked for all models except RuGPT3XL. For using of RuGPT3XL see example of inference of RuGPT3XL here or