https://github.com/xrsrke/pipegoose
Large scale 4D parallelism pre-training for 🤗 transformers in Mixture of Experts *(still work in progress)*
https://github.com/xrsrke/pipegoose
3d-parallelism data-parallelism distributed-optimizers huggingface-transformers large-scale-language-modeling megatron megatron-lm mixture-of-experts model-parallelism moe pipeline-parallelism sequence-parallelism tensor-parallelism transformers zero-1
Last synced: 6 months ago
JSON representation
Large scale 4D parallelism pre-training for 🤗 transformers in Mixture of Experts *(still work in progress)*
- Host: GitHub
- URL: https://github.com/xrsrke/pipegoose
- Owner: xrsrke
- License: mit
- Created: 2023-06-14T06:14:50.000Z (almost 2 years ago)
- Default Branch: main
- Last Pushed: 2023-12-14T06:20:37.000Z (over 1 year ago)
- Last Synced: 2024-04-25T06:20:51.384Z (about 1 year ago)
- Topics: 3d-parallelism, data-parallelism, distributed-optimizers, huggingface-transformers, large-scale-language-modeling, megatron, megatron-lm, mixture-of-experts, model-parallelism, moe, pipeline-parallelism, sequence-parallelism, tensor-parallelism, transformers, zero-1
- Language: Python
- Homepage:
- Size: 1.26 MB
- Stars: 68
- Watchers: 4
- Forks: 16
- Open Issues: 32
-
Metadata Files:
- Readme: README.md
- License: LICENSE
Awesome Lists containing this project
README
# 🚧 pipegoose: Large-scale 4D parallelism multi-modal pre-training for 🤗 `transformers` in Mixture of Experts
[
](https://github.com/xrsrke/pipegoose) [](https://github.com/xrsrke/pipegoose/actions/workflows/tests.yaml) [
](https://discord.gg/s9ZS9VXZ3p) [](https://github.com/psf/black) [
](https://app.codecov.io/gh/xrsrke/pipegoose) [](https://pycqa.github.io/isort/)
We're building an end-to-end library for **training multi-modal MoE in a decentralized way, as proposed by the paper [DiLoCo](https://arxiv.org/abs/2311.08105)**. The core papers that we are replicating are:
- DiLoCo: Distributed Low-Communication Training of Language Models [[link]](https://arxiv.org/abs/2311.08105)
- Pipeline MoE: A Flexible MoE Implementation with Pipeline Parallelism [[link]](https://arxiv.org/abs/2304.11414)
- Switch Transformers: Scaling to Trillion Parameter Models with Simple and Efficient Sparsity [[link]](https://arxiv.org/abs/2101.03961)
- Flamingo: a Visual Language Model for Few-Shot Learning [[link]](https://arxiv.org/abs/2204.14198)
- Megatron-LM: Training Multi-Billion Parameter Language Models Using Model Parallelism [[link]](https://arxiv.org/abs/1909.08053)
**If you're interested in contributing, check out [[CONTRIBUTING.md]](./CONTRIBUTING.md) [[good first issue]](https://github.com/xrsrke/pipegoose/issues?q=is%3Aopen+is%3Aissue+label%3A%22good+first+issue%22) [[roadmap]](https://github.com/users/xrsrke/projects/5). Come join us: [[discord link]](https://discord.gg/s9ZS9VXZ3p)**
⚠️ **Currently only parallelize `transformers`'s `bloom` is supported.**
```diff
from torch.utils.data import DataLoader
+ from torch.utils.data.distributed import DistributedSampler
from torch.optim import Adam
from transformers import AutoModelForCausalLM, AutoTokenizer
from datasets import load_dataset
+ from pipegoose.distributed import ParallelContext, ParallelMode
+ from pipegoose.nn import DataParallel, TensorParallel
+ from pipegoose.optim import DistributedOptimizer
model = AutoModelForCausalLM.from_pretrained("bigscience/bloom-560m")
tokenizer = AutoTokenizer.from_pretrained("bigscience/bloom-560m")
tokenizer.pad_token = tokenizer.eos_token
BATCH_SIZE = 4
+ DATA_PARALLEL_SIZE = 2
+ parallel_context = ParallelContext.from_torch(
+ tensor_parallel_size=2,
+ data_parallel_size=2,
+ pipeline_parallel_size=1
+ )
+ model = TensorParallel(model, parallel_context).parallelize()
+ model = DataParallel(model, parallel_context).parallelize()
model.to("cuda")
+ device = next(model.parameters()).device
optim = Adam(model.parameters(), lr=1e-3)
+ optim = DistributedOptimizer(optim, parallel_context)
dataset = load_dataset("imdb", split="train")
+ dp_rank = parallel_context.get_local_rank(ParallelMode.DATA)
+ sampler = DistributedSampler(dataset, num_replicas=DATA_PARALLEL_SIZE, rank=dp_rank, seed=42)
+ dataloader = DataLoader(dataset, batch_size=BATCH_SIZE // DATA_PARALLEL_SIZE, shuffle=False, sampler=sampler)
for epoch in range(100):
+ sampler.set_epoch(epoch)
for batch in dataloader:
inputs = tokenizer(batch["text"], padding=True, truncation=True, max_length=1024, return_tensors="pt")
inputs = {name: tensor.to(device) for name, tensor in inputs.items()}
labels = inputs["input_ids"]
outputs = model(**inputs, labels=labels)
optim.zero_grad()
outputs.loss.backward()
optim.step()
```
**Installation and try it out**
You can install the package through the following command:
```bash
git clone https://github.com/xrsrke/pipegoose.git
cd pipegoose && pip install -e .
```
And try out a hybrid tensor and data parallelism training script (You must have at least 4 GPUs in order to try hybrid 2D parallelism).
```bash
cd pipegoose/examples
torchrun --standalone --nnodes=1 --nproc-per-node=4 hybrid_parallelism.py
```
We did a small scale correctness test by comparing the validation losses between a paralleized transformer and one kept by default, starting at identical checkpoints and training data. We will conduct rigorous large scale convergence and weak scaling law benchmarks against Megatron and DeepSpeed in the near future if we manage to make it.
- Data Parallelism [[link]](https://wandb.ai/xariusdrake/pipegoose/runs/t5cr56xd?workspace)
- ~~Tensor Parallelism [[link]](https://wandb.ai/xariusdrake/pipegoose/runs/iz17f50n)~~ (We've found a bug in convergence, and we are fixing it)
- ~~Hybrid 2D Parallelism (TP+DP) [[link]](https://wandb.ai/xariusdrake/pipegoose/runs/us31p3q1)~~
- Distributed Optimizer ZeRO-1 Convergence: [[sgd link]](https://wandb.ai/xariusdrake/pipegoose/runs/fn4t9as4?workspace) [[adam link]](https://wandb.ai/xariusdrake/pipegoose/runs/yn4m2sky)
- Mixture of Experts [[link]](https://wandb.ai/xariusdrake/pipegoose/jobs/QXJ0aWZhY3RDb2xsZWN0aW9uOjExOTU2MTU5MA==/version_details/v20)
**Features**
- End-to-end multi-modal including in 3D parallelism including distributed CLIP..
- Sequence parallelism and Mixture of Experts that work in 3D parallelism
- ZeRO-1: Distributed Optimizer
- Kernel fusion
- ...
**Appreciation**
- Big thanks to 🤗 [Hugging Face](https://huggingface.co/) for sponsoring this project with GPUs for testing!
- The library's APIs are inspired by [OSLO](https://github.com/EleutherAI/oslo)'s and [ColossalAI](https://github.com/hpcaitech/ColossalAI)'s APIs.
**Citation**
```
@software{pipegoose,
title = {{pipegoose: Large-scale 4D parallelism pre-training for `transformers`}},
author = {},
url = {https://github.com/xrsrke/pipegoose},
doi = {},
month = {},
year = {2024},
version = {},
}
```