dilated-attention-pytorch

(Unofficial) Implementation of DilatedAttention from LongNet: Scaling Transformers to 1,000,000,000 Tokens in PyTorch.

Install

NOTE: This library depends on facebookresearch/xformers. If you're not using torch>=2.0.0, you may need to install it from source. See their installation instructions.

PyPI:

pip install dilated-attention-pytorch

From source:

pip install "dilated-attention-pytorch @ git+ssh://[email protected]/fkodom/dilated-attention-pytorch.git"

For contributors:

# Install all dev dependencies (tests etc.)
pip install "dilated-attention-pytorch[all] @ git+ssh://[email protected]/fkodom/dilated-attention-pytorch.git"
# Setup pre-commit hooks
pre-commit install

Benchmark

I follow the benchmarking procedure from the LongNet paper (Section 3.1) as best I can. They tested in a distributed, multi-GPU setting (and by my estimation, with much better GPUs), and I test on a single GTX 2080 Ti, but the same general scaling trends still apply. Rather than 1B tokens, I scale the batch size so that the total number of tokens is 32M, which is the largest sequence that fits in memory on my GPU when running dilated attention.

See: benchmark.py

NOTE: Clearly, there are some inefficiencies in my DilatedAttention implementation for shorter sequence lengths. I'm not sure what's causing this. If you have any insights, please let me know!

Usage

DilatedAttention

The LongNet paper introduces a new attention mechanism called DilatedAttention. It is a drop-in replacement (see below) for "vanilla" attention that allows for much longer sequences to be processed.

NOTE: DilatedAttention only supports batch_first=True. This is different from "vanilla" attention in PyTorch, which supports both batch_first=True and batch_first=False.

Arguments:

• segment_lengths (required, list[int]): Length of each attention segment. This is usually a geometric sequence increasing in powers of 2, such as [2048, 4096, 8192].
• dilation_rates (required, list[int]): Dilation rate for each segment. Like with segment_lengths, this is usually a geometric sequence increasing in powers of 2, such as [1, 2, 4].

import torch
from dilated_attention_pytorch.dilated_attention import DilatedAttention

dilated_attention = DilatedAttention(
    segment_lengths=[2048, 4096, 8192],
    dilation_rates=[1, 2, 4],
)

# shape: (batch_size, seq_len, num_heads, embed_dim)
# NOTE: 'seq_len' must be a multiple of 8192 (the largest segment length)
# NOTE: For best performance, use 'dtype=torch.float16' or `dtype=torch.bfloat16`
query = torch.randn(1, 8192, 8, 64, device="cuda", dtype=torch.float16)
key = torch.randn(1, 8192, 8, 64, device="cuda", dtype=torch.float16)
value = torch.randn(1, 8192, 8, 64, device="cuda", dtype=torch.float16)

out = dilated_attention(query, key, value, is_causal=False)  # default: causal=False
print(out.shape)
# torch.Size([1, 8192, 8, 64])

MultiheadDilatedAttention

MultiheadDilatedAttention is a drop-in replacement (see below) for nn.MultiheadAttention that uses DilatedAttention instead of "vanilla" attention. It also incorporates improvements from the MAGNETO architecture (nn.LayerNorm placements), as mentioned in the LongNet paper.

NOTE: MultiheadDilatedAttention only supports batch_first=True. This is different from nn.MultiheadAttention, which supports both batch_first=True and batch_first=False.

Arguments:

• segment_lengths (required, list[int]): Length of each attention segment. This is usually a geometric sequence increasing in powers of 2, such as [2048, 4096, 8192].
• dilation_rates (required, list[int]): Dilation rate for each segment. Like with segment_lengths, this is usually a geometric sequence increasing in powers of 2, such as [1, 2, 4].
• Many of the same arguments from nn.MultiheadAttention. See the MultiheadDilatedAttention class for more details.

from dilated_attention_pytorch.dilated_attention import MultiheadDilatedAttention

device = torch.device("cuda")
dtype = torch.float16
embed_dim = 512

# NOTE: Omitting most of the optional arguments for brevity
mhda = MultiheadDilatedAttention(
    embed_dim=embed_dim,
    num_heads=8,
    segment_lengths=[2048, 4096, 8192],
    dilation_rates=[1, 2, 4],
    device=device,  # optional
    dtype=dtype,  # optional
)

# shape: (batch_size, seq_len, embed_dim)
# NOTE: 'seq_len' must be a multiple of 8192 (the largest segment length)
x = torch.randn(1, 8192, embed_dim, device=device, dtype=dtype)
y = mhda(x, x, x, is_causal=False)  # default: is_causal=False
print(y.shape)
# torch.Size([1, 8192, 512])

LongNet

The LongNet paper culminates in a transformer architecture, which can be trained for language modeling with very long context windows. I have implemented two LongNet variants, based on the base configurations from the paper:

• LongNetLM - designed specifically for language modeling
• LongNet - a more general encoder-decoder architecture, which is not specific to language modeling

Based on these implementations, it is fairly straightforward to adapt LongNet to encoder- or decoder-only architectures, as needed for specific applications.

from dilated_attention_pytorch.long_net import LongNetLM, LongNet

device = torch.device("cuda")
dtype = torch.float16

# NOTE: Showing all default values, which are described in the paper.
net = LongNet(
    d_model=768,
    nhead=12,
    num_encoder_layers=12,
    num_decoder_layers=12,
    dim_feedforward=3072,
    segment_lengths=[2048, 4096, 8192, 16384, 32768],
    dilation_rates=[1, 2, 4, 6, 12],
    dropout=0.0,
    activation="relu",
    layer_norm_eps=1e-5,
    device=device,
    dtype=dtype,
)
# shape: (batch_size, seq_len, d_model)
x = torch.randn(1, 32768, 768, device=device, dtype=dtype)
with torch.no_grad():
    y = net.forward(x, is_causal=True)  # default: is_causal=True
print(y.shape)
# torch.Size([1, 32768, 768])

num_tokens = 10000  # (required) usually obtained from the tokenizer
lm = LongNetLM(
    num_tokens=num_tokens,
    d_model=768,
    nhead=12,
    num_encoder_layers=12,
    num_decoder_layers=12,
    dim_feedforward=3072,
    segment_lengths=[2048, 4096, 8192, 16384, 32768],
    dilation_rates=[1, 2, 4, 6, 12],
    dropout=0.0,
    activation="relu",
    layer_norm_eps=1e-5,
    device=device,
    dtype=dtype,
)
# shape: (batch_size, seq_len)
x = torch.randint(0, num_tokens, (1, 32768), device=device, dtype=torch.long)
with torch.no_grad():
    y = lm.forward(x, is_causal=True)  # default: is_causal=True
print(y.shape)
# torch.Size([1, 32768, num_tokens])

Citations

@misc{ding2023longnet,
      title={LongNet: Scaling Transformers to 1,000,000,000 Tokens}, 
      author={Jiayu Ding and Shuming Ma and Li Dong and Xingxing Zhang and Shaohan Huang and Wenhui Wang and Nanning Zheng and Furu Wei},
      year={2023},
      eprint={2307.02486},
      archivePrefix={arXiv},
      primaryClass={cs.CL}
}