FlagAttention

中文版

FlagAttention is a project for memory-efficient attention operators implemented in the Triton language. Motivated by the need for non-standard attention operators in language modeling, it starts as an extension of multi-head attention.

It saves memory footprint and traffic like FlashAttention and FlashAttention v2. Implemented in the Triton language, it is easier to understand and modify. The original implementation of FlashAttention in CUDA(flash-attention) provides a good example of how to design an algorithm that takes different levels of memory into account. By tiling and re-computation, FlashAttention avoids materializing the attention scores, whose capacity is proportional to the square of the sequence length. However, custom transformation to the attention scores is not possible when using FlashAttention, unless it is supported by FlashAttention out-of-the-box. While extending FlashAttention requires proficiency in CUDA programming, FlagAttention implemented in the Triton language is easier to modify.

FlagAttention now offers two operators.

1. flash_attention: FlashAttention implemented in the Triton language.
2. piecewise_attention. Currently employed for NLPE(Non-Linearized position embedding) in both training and inference of the Aquila-2-34B model.

When further customization is required, FlagAttention servers as an example.

Changelog

v0.1

Add piecewise_attention & flash_attention.

v0.2

Optimization of operators.

1. applying mask only when needed.
2. use a separate kernel to compute the gradien of q to avoid atomic RMW to global memory.

Requirements

FlagAttention requires Pytorch and Triton. To use the new features of Triton, a nightly release is recommended.

# install a nightly release of Triton
pip install -U --index-url https://aiinfra.pkgs.visualstudio.com/PublicPackages/_packaging/Triton-Nightly/pypi/simple/ triton-nightly

FlagAttention requires Ampere Nvidia GPUs(e.g. A100, RTX-3090, ...) and CUDA Toolkit 11.6 or above. Other GPUs may work but have not been tested yet.

Installation

FlagAttention can be installed in either way below.

1. Editable Installation. Changes to the code in the local source tree are effective without re-installation.
2. Build a distribution and then install. Only the package is installed.

Editable Installation

Editable installation with pip.

git clone https://github.com/FlagOpen/FlagAttention && cd FlagAttention
pip install -e .

Build a Distribution & Install

Following modern Python packaging convention(PEP-517), FlagAttention is configured by pyproject.toml, and no setup.py is provided. To build a distribution, either a source distribution or a binary distribution, python package build is recommended.

First, install build package via pip.

pip install build

Then build the package.

git clone https://github.com/FlagOpen/FlagAttention && cd FlagAttention
# to build in `no-isolation` mode requires installing build requirements manually
pip install -U setuptools setuptools-scm
python -m build --no-isolation

The built package is in dist/ for installation.

pip install dist/flag_attn-xxx.whl

Usage

FlagAttention provides customized operators for attention. When an operator is equivalent to a torch function, it can be used as a drop-in replacement.

Run the Tests

A recent version of pytest(>=7.1.0) is required to run the tests in tests/. Operators in FlagAttention are tested against reference implementations in Pytorch provided by flag_attn.testing, both for the forward and backward operators. For operators with support for inputs of float16 or bfloat16, three different implementations are included for numerical accuracy testing.

1. Reference Implementation in Pytorch: This implementation upcasts the inputs to float32 and performs the computations in float32 all the way through before casting the outputs to float16 or bfloat16.
2. Triton Implementation: The Triton implementation uses float16 or bfloat16 for MMA(matrix multiplication accumulation) inputs and float32 for MMA outputs and other computations.
3. Pytorch Implementation: This implementation mirrors the computations in the reference implementation, except that the precision is the same as the Triton implementation.

The tests for numerical accuracy enforce that the maximum difference between the Triton implementation and reference implementation is not greater than twice the maximanum difference between the Pytorch implementation and reference implementation.

pytest .

Run the Benchmark

Benchmarks are included to quantify the achieved TFLOP/s, which serves as a metric of speed operators. The calculation of FLOPs for an operator considers only the matmul operation. The resulting FLOPs are then divided by the median runtime to determine the achieved FLOPs/s.

The benchmarking process involves comparing the Triton implementations with counterparts in Pytorch. When the input size is large, resulting in memory exhaustion in the Pytorch implementation, the FLOP/s is considered zero.

cd benchmarks/
python flash_benchmark.py
python piecewise_benchmark.py

Operators

flash_attention

The implementation of FlashAttention in the Triton language. The interface is.

flash_attention(q, k, v, causal=False, sm_scale=None, return_log_normalizer=False, return_total_attention=False)

In addition to the attention outputs, it can return some extra outputs dependes on return_log_normalizer and return_total_attention.

1. log_normalizer: shape (batch_size, num_heads, seqlen_q). The log normalizer of the softmax inside attention operation.
2. total_attention: shape (batch_size, num_heads, seqlen_k). The sum of attention weights along q's sequence axis.

piecewise_attention

The first extension to FlashAttention is piecewise_attention. This operator enhances FlashAttention by using two q's and two k's to calculate the attention scores(S) before applying softmax to obtain the attention weights(P).

The rationale behind this design is rooted in the observations that a transformer with rotary position embedding struggles with predicting sequences longer than the maximum sequence length it is trained on. Pairs of (q, k) yield unexpectedly high attention scores when the distance exceeds the maximum sequence length in the training set.

To address this issue, BAAI proposes NLPE(Non-Linearized Position Embedding), which applies two different position embeddings to q and k based on whether the distance between q and k exceeds a pre-defined threshold, producing q1, q2 and k1, k2. Then the attention score is computed as the dot product of q1, k1 or q2, k2 depending on the distance between q and k.

The interface is shown below.

piecewise_attention(q1, k1, q2, k2, v, dist_threshold, causal=False, sm_scale=None)

It splices two attention scores(S) in the forward computation and splits the gradient of S in the backward computation.

Usage

# piecewise_attention
import torch
from flag_attn import piecewise_attention

B, H, T, D = 2, 16, 8192, 128
dist_threshold = T // 2

q1 = torch.randn((B, H, T, D), dtype=torch.float16, device="cuda:0").requires_grad_()
q2 = torch.randn((B, H, T, D), dtype=torch.float16, device="cuda:0").requires_grad_()
k1 = torch.randn((B, H, T, D), dtype=torch.float16, device="cuda:0").requires_grad_()
k2 = torch.randn((B, H, T, D), dtype=torch.float16, device="cuda:0").requires_grad_()
v = torch.randn((B, H, T, D), dtype=torch.float16, device="cuda:0").requires_grad_()
o = piecewise_attention(q1, k1, q2, k2, v, dist_threshold, causal=True)
print(o)

go = torch.randn((B, H, T, D), dtype=torch.float16, device="cuda:0")
gq1, gk1, gq2, gk2, gv = torch.autograd.grad(
    o, (q1, k1, q2, k2, v), go
)
print(gq1)

# flash_attention
import torch
from flag_attn import flash_attention

B, H, T, D = 2, 16, 8192, 128

q = torch.randn((B, H, T, D), dtype=torch.float16, device="cuda:0").requires_grad_()
k = torch.randn((B, H, T, D), dtype=torch.float16, device="cuda:0").requires_grad_()
v = torch.randn((B, H, T, D), dtype=torch.float16, device="cuda:0").requires_grad_()
o = flash_attention(q, k, v, causal=True)
print(o)

go = torch.randn((B, H, T, D), dtype=torch.float16, device="cuda:0")
gq, gk, gv = torch.autograd.grad(
    o, (q, k, v), go
)
print(gq)

Performance

Benchmark is performed under such conditions.

1. seqlen in [512, 1k, 2k, 4k, 16k, 32k];
2. batch size: 32k / seqlen;
3. headdim in[64, 128]；
4. num_heads: 2048 / headdim.

flash_attention

The performance of flash_attention with causal masking is shown below.

The forward operator runs as fast as, and in some cases, faster than FlashAttention(CUDA), but the backward operator is generally slower than FlashAttention. We first follow the paper and update the gradient of Q with atomic addition in the backward operator, which runs extremely slowly. Then we split the backward operator into two kernels, one to compute the gradient of k and v, the other to compute the gradient of q. This alternation avoids atomic additions but introduces more re-computation. Although this strategy yields a 4x to 5x speedup in the backward operator, it is still slower than FlashAttention(CUDA).

The same split-kernel trick is also applied to piecewise_attention for efficiency.

piecewise_attention

The performance of piecewise_attention has improved compared to that in v0.1. In the case where the head dim is 128 and causal masking is applied, the forward and backward operator is faster than that in v0.1 by 36% and 9%, respectively.

Features

• support for Nvidia Ampere GPU(Tested on RTX-3090 and A100)；
• support for Iluvatar CoreX GPU(Tested on Iluvatar CoreX MR-V100)；
• datatype support, float16 and bfloat16 for Ampere Nvidia GPUs;
• support causal and non-causal modes;
• support forward & backward modes;
• the sequence length of k/v can be different from that of q;
• support computation of total attention of each k gets from all q's;
• supports returning accumulative attention of each keys.
• supports MQA and GQA.
• supports dropout of attention weights.

Limitations

• headdim should be in [16, 32, 64, 128].

TODOs

1. Test on other GPUs;
2. Test on more versions of triton；
3. Improve performance of attention operators(especially for the backward op);
4. Support other extensions to flash attention.

More

For more about the open source system for large models from BAAI, please with BAAI/FlagOpen.