Implementation of a Transformer, but completely in Triton

Related tags

Deep Learning deep-learning transformers artificial-intelligence attention-mechanism gpu-programming
Overview

Transformer in Triton (wip)

Implementation of a Transformer, but completely in Triton. I'm completely new to lower-level neural net code, so this repository will mostly be a learning experience, with the end-goal being a vanilla transformer that is faster and more efficient to train.

Install

$ pip install triton-transformer

Usage

import torch
from triton_transformer import Transformer

model = Transformer(
    num_tokens = 256,
    max_seq_len = 1024,
    dim = 512,
    depth = 6,
    heads = 8,
    dim_head = 64
)

x = torch.randint(0, 256, (1, 1024))
mask = torch.ones(1, 1024).bool()

logits = model(x, mask = mask) # (1, 1024, 256)

Citations

@article{Tillet2019TritonAI,
    title   = {Triton: an intermediate language and compiler for tiled neural network computations},
    author  = {Philippe Tillet and H. Kung and D. Cox},
    journal = {Proceedings of the 3rd ACM SIGPLAN International Workshop on Machine Learning and Programming Languages},
    year    = {2019}
}
@misc{vaswani2017attention,
    title   = {Attention Is All You Need}, 
    author  = {Ashish Vaswani and Noam Shazeer and Niki Parmar and Jakob Uszkoreit and Llion Jones and Aidan N. Gomez and Lukasz Kaiser and Illia Polosukhin},
    year    = {2017},
    eprint  = {1706.03762},
    archivePrefix = {arXiv},
    primaryClass = {cs.CL}
}
You might also like...
A Pytorch implementation of CVPR 2021 paper "RSG: A Simple but Effective Module for Learning Imbalanced Datasets"

RSG: A Simple but Effective Module for Learning Imbalanced Datasets (CVPR 2021) A Pytorch implementation of our CVPR 2021 paper "RSG: A Simple but Eff

null 120 Dec 12, 2022
A concise but complete implementation of CLIP with various experimental improvements from recent papers
A concise but complete implementation of CLIP with various experimental improvements from recent papers

x-clip (wip) A concise but complete implementation of CLIP with various experimental improvements from recent papers Install $ pip install x-clip Usag

Phil Wang 515 Dec 26, 2022
A concise but complete implementation of CLIP with various experimental improvements from recent papers
A concise but complete implementation of CLIP with various experimental improvements from recent papers

x-clip (wip) A concise but complete implementation of CLIP with various experimental improvements from recent papers Install $ pip install x-clip Usag

Phil Wang 115 Dec 9, 2021
Implementation of a protein autoregressive language model, but with autoregressive infilling objective (editing subsequences capability)
Implementation of a protein autoregressive language model, but with autoregressive infilling objective (editing subsequences capability)

Protein GLM (wip) Implementation of a protein autoregressive language model, but with autoregressive infilling objective (editing subsequences capabil

Phil Wang 17 May 6, 2022
Third party Pytorch implement of Image Processing Transformer (Pre-Trained Image Processing Transformer arXiv:2012.00364v2)

ImageProcessingTransformer Third party Pytorch implement of Image Processing Transformer (Pre-Trained Image Processing Transformer arXiv:2012.00364v2)

null 61 Jan 1, 2023
Episodic Transformer (E.T.) is a novel attention-based architecture for vision-and-language navigation. E.T. is based on a multimodal transformer that encodes language inputs and the full episode history of visual observations and actions.
Episodic Transformer (E.T.) is a novel attention-based architecture for vision-and-language navigation. E.T. is based on a multimodal transformer that encodes language inputs and the full episode history of visual observations and actions.

Episodic Transformers (E.T.) Episodic Transformer for Vision-and-Language Navigation Alexander Pashevich, Cordelia Schmid, Chen Sun Episodic Transform

Alex Pashevich 62 Dec 24, 2022
CSWin Transformer: A General Vision Transformer Backbone with Cross-Shaped
CSWin Transformer: A General Vision Transformer Backbone with Cross-Shaped

CSWin-Transformer This repo is the official implementation of "CSWin Transformer: A General Vision Transformer Backbone with Cross-Shaped Windows". Th

Microsoft 409 Jan 6, 2023
nnFormer: Interleaved Transformer for Volumetric Segmentation Code for paper "nnFormer: Interleaved Transformer for Volumetric Segmentation "

nnFormer: Interleaved Transformer for Volumetric Segmentation Code for paper "nnFormer: Interleaved Transformer for Volumetric Segmentation ". Please

jsguo 610 Dec 28, 2022
3D-Transformer: Molecular Representation with Transformer in 3D Space

3D-Transformer: Molecular Representation with Transformer in 3D Space

null 55 Dec 19, 2022
Comments
  • Question concerning PyTorch build

    Question concerning PyTorch build

    Hello. I find your project very interesting and I have seen your comparison between PyTorch and Triton implementations.

    However, I am curious whether your PyTorch environment is a source build optimized for your machine or a pip/conda install.

    Source building has faster runtimes and if a conda install is being used for comparison, the difference in speed may simply be due to Triton optimizing CUDA for the run environment.

    Thank you again for your interesting project.

    opened by veritas9872 13
  • _layernorm implementation forward result not equal F.layer_norm

    _layernorm implementation forward result not equal F.layer_norm

    I have a try on your triton-transformer and test the layernorm module alone. It's very weird that the forward result is different while the backward result is equal.

    code: from triton_transformer.layernorm import layernorm import torch import torch.nn as nn

    torch.manual_seed(0) x = torch.randn(2,5).cuda() x.requires_grad_(True) dy = .1*torch.randn_like(x).cuda() dim = 5 norm = nn.LayerNorm(dim).cuda()

    y1 = layernorm(x, norm.weight, norm.bias, use_triton = True) y2 = layernorm(x, norm.weight, norm.bias, use_triton = False) print(y1, y2) print(torch.allclose(y1, y2))

    y1.backward(dy, retain_graph=True) dx_y1 = x.grad.clone()

    x.grad = None

    y2.backward(dy, retain_graph=True) dx_y2 = x.grad.clone() print(dx_y1, dx_y2) print(torch.allclose(dx_y1, dx_y2))

    result: `tensor([[ 0.9492, -0.0021, -0.9797, 0.4449, -0.4123], [-0.7624, 0.4399, 0.7299, -0.3091, -0.0983]], device='cuda:0', grad_fn=<_layernormBackward>) tensor([[ 1.4217, -0.0031, -1.4674, 0.6663, -0.6175], [-1.4342, 0.8276, 1.3732, -0.5815, -0.1850]], device='cuda:0', grad_fn=) False

    tensor([[-0.0706, 0.0288, -0.0813, 0.0446, 0.0785], [ 0.0218, -0.0152, 0.0141, -0.0522, 0.0315]], device='cuda:0') tensor([[-0.0706, 0.0288, -0.0813, 0.0446, 0.0785], [ 0.0218, -0.0152, 0.0141, -0.0522, 0.0315]], device='cuda:0') True`

    opened by Tengxu-Sun 1
  • Current state of benchmarking & contributing?

    Current state of benchmarking & contributing?

    Hey @lucidrains - hope you're doing well! I have some time to hack the next couple weeks, just wanted to get a sense of:

    • Current state of benchmarking (what Triton kernels provide how much lift, aggregate lift over a "vanilla Transformer implementation"
    • If there's anything I could help with, especially as I learn Triton!
    opened by siddk 0
  • Official layer norm added

    Official layer norm added

    Hi @lucidrains , in Triton layer norm was just added in examples, https://github.com/openai/triton/commit/d4baad426db72b83c5222e1c83c929c1860cae54 I tested it, it's twice as fast as Torch, often faster then Apex.

    I'm looking forward for your implementation of attention, so far the Torch implementation is the fastest with 12.3 / 14.5 (forw / back) vs the other Triton implementation in DeepSpeed which is 17.3/ 23.0 on my data.

    opened by olegklimov 2
Releases(0.1.1)
Owner
Phil Wang
Working with Attention. It's all we need
Phil Wang
GitHub
Deploy optimized transformer based models on Nvidia Triton server

?? Hugging Face Transformer submillisecond inference ?? and deployment on Nvidia Triton server Yes, you can perfom inference with transformer based mo

Lefebvre Sarrut Services 1.2k Jan 5, 2023
VSR-Transformer - This paper proposes a new Transformer for video super-resolution (called VSR-Transformer).

VSR-Transformer By Jiezhang Cao, Yawei Li, Kai Zhang, Luc Van Gool This paper proposes a new Transformer for video super-resolution (called VSR-Transf

Jiezhang Cao 225 Nov 13, 2022
A simple but complete full-attention transformer with a set of promising experimental features from various papers

x-transformers A concise but fully-featured transformer, complete with a set of promising experimental features from various papers. Install $ pip ins

Phil Wang 2.3k Jan 3, 2023
Implementation of Transformer in Transformer, pixel level attention paired with patch level attention for image classification, in Pytorch

Transformer in Transformer Implementation of Transformer in Transformer, pixel level attention paired with patch level attention for image c

Phil Wang 272 Dec 23, 2022
This is an official implementation for "Swin Transformer: Hierarchical Vision Transformer using Shifted Windows" on Object Detection and Instance Segmentation.

Swin Transformer for Object Detection This repo contains the supported code and configuration files to reproduce object detection results of Swin Tran

Swin Transformer 1.4k Dec 30, 2022
Implementation of Vision Transformer, a simple way to achieve SOTA in vision classification with only a single transformer encoder, in Pytorch

Implementation of Vision Transformer, a simple way to achieve SOTA in vision classification with only a single transformer encoder, in Pytorch

Phil Wang 12.6k Jan 9, 2023
The implementation of "Shuffle Transformer: Rethinking Spatial Shuffle for Vision Transformer"

Shuffle Transformer The implementation of "Shuffle Transformer: Rethinking Spatial Shuffle for Vision Transformer" Introduction Very recently, window-

null 87 Nov 29, 2022
Unofficial implementation of "Swin Transformer: Hierarchical Vision Transformer using Shifted Windows" (https://arxiv.org/abs/2103.14030)

Swin-Transformer-Tensorflow A direct translation of the official PyTorch implementation of "Swin Transformer: Hierarchical Vision Transformer using Sh

null 52 Dec 29, 2022
Implementation of the Transformer variant proposed in "Transformer Quality in Linear Time"

FLASH - Pytorch Implementation of the Transformer variant proposed in the paper Transformer Quality in Linear Time Install $ pip install FLASH-pytorch

Phil Wang 209 Dec 28, 2022
Official implementation of "Not only Look, but also Listen: Learning Multimodal Violence Detection under Weak Supervision" ECCV2020

XDVioDet Official implementation of "Not only Look, but also Listen: Learning Multimodal Violence Detection under Weak Supervision" ECCV2020. The proj

peng 64 Dec 12, 2022