• Parallelformers, which is based on Megatron LM, is designed to make model parallelization easier.
• You can parallelize various models in HuggingFace Transformers on multiple GPUs with a single line of code.
• Currently, Parallelformers only supports inference. Training features are NOT included.

What's New:

• October 24, 2021 Docker support.
• July 28, 2021 Released a tech blog.
• July 18, 2021 Released Parallelformers 1.0.

Why Parallelformers?

You can load a model that is too large for a single GPU. For example, using Parallelformers, you can load a model of 12GB on two 8 GB GPUs. In addition, you can save your precious money because usually multiple smaller size GPUs are less costly than a single larger size GPU.

Installation

Parallelformers can be easily installed using the pip package manager. All the dependencies such as torch, transformers, and dacite should be installed automatically with the following command. Be careful that the name is plural.

pip install parallelformers

Getting Started

1. Create a HuggingFace transformers model.

You don't need to call .half() or .cuda() as those functions will be invoked automatically. It is more memory efficient to start parallelization on the CPU.

from transformers import AutoModelForCausalLM, AutoTokenizer

model = AutoModelForCausalLM.from_pretrained("EleutherAI/gpt-neo-2.7B")
tokenizer = AutoTokenizer.from_pretrained("EleutherAI/gpt-neo-2.7B")

2. Put the model in the parallelize() function.

from parallelformers import parallelize

parallelize(model, num_gpus=2, fp16=True, verbose='detail')

Since nvidia-smi shows the reserved cache area, it is difficult to check the exact allocated memory. To check the allocated memory state well, you can set the verbose option as 'detail' or 'simple'. (default is None). If you want to set a random seed value, input the seed value using parallelize(..., seed=YOUR_SEED).

|===========================================================================|
|                  PyTorch CUDA memory summary, device ID 0                 |
|---------------------------------------------------------------------------|
|            CUDA OOMs: 0            |        cudaMalloc retries: 0         |
|===========================================================================|
|        Metric         | Cur Usage  | Peak Usage | Tot Alloc  | Tot Freed  |
|---------------------------------------------------------------------------|
| Allocated memory      |    2721 MB |    2967 MB |    2967 MB |  251905 KB |
|       from large pool |    2720 MB |    2966 MB |    2966 MB |  251904 KB |
|       from small pool |       1 MB |       1 MB |       1 MB |       1 KB |
|---------------------------------------------------------------------------|

GPU:0 => 2.72GB

|===========================================================================|
|                  PyTorch CUDA memory summary, device ID 1                 |
|---------------------------------------------------------------------------|
|            CUDA OOMs: 0            |        cudaMalloc retries: 0         |
|===========================================================================|
|        Metric         | Cur Usage  | Peak Usage | Tot Alloc  | Tot Freed  |
|---------------------------------------------------------------------------|
| Allocated memory      |    2721 MB |    2967 MB |    2967 MB |  251905 KB |
|       from large pool |    2720 MB |    2966 MB |    2966 MB |  251904 KB |
|       from small pool |       1 MB |       1 MB |       1 MB |       1 KB |
|---------------------------------------------------------------------------|

GPU:1 => 2.72GB

3. Do Inference as usual.

You don't have to call .cuda() when creating input tokens. Note that you should input both input tokens and attention masks to the model. (**inputs is the recommended way for this)

inputs = tokenizer("Parallelformers is", return_tensors="pt")

outputs = model.generate(
    **inputs,
    num_beams=5,
    no_repeat_ngram_size=4,
    max_length=15,
)

print(f"Output: {tokenizer.batch_decode(outputs)[0]}")

Output: Parallelformers is an open-source library for parallel programming ...

4. Deploy the model to the server as usual.

The parallelization process does not affect the web server because they are automatically synchronized.

from flask import Flask

app = Flask(__name__)


@app.route("/generate_text/<text>")
def generate_text(text):
    inputs = tokenizer(text, return_tensors="pt")
    
    outputs = model.generate(
        **inputs,
        num_beams=5,
        no_repeat_ngram_size=4,
        max_length=15,
    )
    
    outputs = tokenizer.batch_decode(
        outputs,
        skip_special_tokens=True,
    )
    
    return {
        "inputs": text,
        "outputs": outputs[0],
    }


app.run(host="0.0.0.0", port=5000)

You can send a request to the web server as follows:

$ curl -X get "YOUR_IP:5000/generate_text/Messi"

And the following result should be returned.

{"inputs": "Messi", "outputs": "Messi is the best player in the world right now. He is the"}

5. Check the current GPU states.

You can check GPU states using .memory_allocated(), .memory_reserved() and .memory_chached() to make sure the parallelization is successful.

model.memory_allocated()
model.memory_reserved()
model.memory_chached()

{'cuda:0':XXXXXX, 'cuda:1':XXXXXX}

6. Manage the model parallelization states.

You can manage model parallelization states using .cuda(), .cpu() and .to(). The model parallelization process ends if you call those functions.

model.cuda()

print(torch.cuda.memory_summary(0))
print(torch.cuda.memory_summary(1))

Check the allocated memory status using torch.cuda.memory_summary().

|===========================================================================|
|                  PyTorch CUDA memory summary, device ID 0                 |
|---------------------------------------------------------------------------|
|            CUDA OOMs: 0            |        cudaMalloc retries: 0         |
|===========================================================================|
|        Metric         | Cur Usage  | Peak Usage | Tot Alloc  | Tot Freed  |
|---------------------------------------------------------------------------|
| Allocated memory      |    5121 MB |    5121 MB |    5121 MB |    1024 B  |
|       from large pool |    5120 MB |    5120 MB |    5120 MB |       0 B  |
|       from small pool |       1 MB |       1 MB |       1 MB |    1024 B  |
|---------------------------------------------------------------------------|

GPU0 => 5.12GB

|===========================================================================|
|                  PyTorch CUDA memory summary, device ID 1                 |
|---------------------------------------------------------------------------|
|            CUDA OOMs: 0            |        cudaMalloc retries: 0         |
|===========================================================================|
|        Metric         | Cur Usage  | Peak Usage | Tot Alloc  | Tot Freed  |
|---------------------------------------------------------------------------|
| Allocated memory      |       0 B  |    1024 B  |    1024 B  |    1024 B  |
|       from large pool |       0 B  |       0 B  |       0 B  |       0 B  |
|       from small pool |       0 B  |    1024 B  |    1024 B  |    1024 B  |
|---------------------------------------------------------------------------|

GPU1 => 0.00GB

If you switch to the CPU mode, it works like this.

model.cpu()

print(torch.cuda.memory_summary(0))
print(torch.cuda.memory_summary(1))

|===========================================================================|
|                  PyTorch CUDA memory summary, device ID 0                 |
|---------------------------------------------------------------------------|
|            CUDA OOMs: 0            |        cudaMalloc retries: 0         |
|===========================================================================|
|        Metric         | Cur Usage  | Peak Usage | Tot Alloc  | Tot Freed  |
|---------------------------------------------------------------------------|
| Allocated memory      |       0 B  |    5121 MB |    5121 MB |    5121 MB |
|       from large pool |       0 B  |    5120 MB |    5120 MB |    5120 MB |
|       from small pool |       0 B  |       1 MB |       1 MB |       1 MB |
|---------------------------------------------------------------------------|

GPU0 => 0.00GB

|===========================================================================|
|                  PyTorch CUDA memory summary, device ID 1                 |
|---------------------------------------------------------------------------|
|            CUDA OOMs: 0            |        cudaMalloc retries: 0         |
|===========================================================================|
|        Metric         | Cur Usage  | Peak Usage | Tot Alloc  | Tot Freed  |
|---------------------------------------------------------------------------|
| Allocated memory      |       0 B  |    1024 B  |    1024 B  |    1024 B  |
|       from large pool |       0 B  |       0 B  |       0 B  |       0 B  |
|       from small pool |       0 B  |    1024 B  |    1024 B  |    1024 B  |
|---------------------------------------------------------------------------|

GPU1 => 0.00GB

Do your processes die or stop working?

Many issues have pointed this out. And I've found that running code inside the context of if __name__ == '__main__' solves a lot of problems. So if you run have some problems about processes, try writing your code inside the context of it.

Are you getting some errors in docker container?

I recently found out that ALL errors that occur in environments with limited resources such as docker containers are due to shared memory size. So, if you want to use larger models with parallelformers in docker containers, INCREASE the size of shared memory by --shm-size=?gb or REMOVE the limitation of shared memory size by --ipc=host. the larger shared memory size is required if you want to use larger model.

Supported Models

Currently, most models in Huggingface transformers are supported. All layers in the models listed below can be parallelized. They include vision models like ViT, CLIP and speech models like Wav2Vec2 as well as language models.

Fully Supported Models

• ALBERT
• BART
• BARThez (=BERT)
• BERT
• BERTweet (=BERT)
• BertJapanese (=BERT)
• BertGeneration
• Blenderbot
• Blenderbot Samll
• BORT (=BERT)
• CamemBERT (=RoBERTa)
• CLIP
• CPM
• CTRL
• DeBERTa
• DeBERTa-v2
• DeiT
• DETR
• DialoGPT (=GPT2)
• DistilBERT
• DPR (=BERT)
• ELECTRA
• FlauBERT (=XLM)
• FSMT
• Funnel Transformer
• herBERT (=RoBERTa)
• I-BERT
• LayoutLM
• LED
• Longformer
• LUKE
• LXMERT
• MarianMT
• M2M100
• MBart
• Mobile BERT
• MPNet
• MT5 (=T5)
• Megatron BERT (=BERT)
• Megatron GPT2 (=GPT2)
• OpenAI GPT
• OpenAI GPT2
• OPT
• GPTNeo
• GPTJ
• Hubert
• Pegasus
• PhoBERT (=RoBERTa)
• Reformer
• RetriBERT
• RoBERTa
• RoFormer
• Speech2Text
• T5
• ByT5 (=T5)
• TAPAS
• TransformerXL
• ViT
• VisualBERT
• Wav2Vec2
• XLM
• XLM-RoBERTa (=RoBERTa)
• XLNet
• XLSR-Wave2Vec2

At present the following models are partly supported or not supported.

Partly Supported Models

• BigBird
• BigBirdPegasus
• ConvBERT
• ProphetNet
• XLM-ProphetNet

Unsupported Models

• SqueezeBERT
• RAG

Advanced Usage

Refer to POLICY.md

FAQ

Refer to FAQ.md.

Contributing

Refer to CONTRIBUTING.md

Documentation

For more detailed information, see full documentation

Citation

If you find this library useful, please consider citing:

@misc{parallelformers,
  author       = {Ko, Hyunwoong},
  title        = {Parallelformers: An Efficient Model Parallelization Toolkit for Deployment},
  howpublished = {\url{https://github.com/tunib-ai/parallelformers}},
  year         = {2021},
}

This library is cited by:

• Few-Shot Bot: Prompt-Based Learning for Dialogue Systems, Madotto et al, 2021
• AMMUS : A Survey of Transformer-based Pretrained Models in Natural Language Processing, Kalyan et al, 2021

LICENSE

Parallelformers is licensed under the terms of the Apache License 2.0.

Copyright 2021 TUNiB inc. http://www.tunib.ai. All Rights Reserved.