Flash Linear Attention

This repo aims at providing a collection of efficient Triton-based implementations for state-of-the-art linear attention models. Any pull requests are welcome!

Table of Contents

• Models
• Installation
• Usage
  • Token Mixing
  • Fused Modules
  • Generation
  • Hybrid Models
• Evaluations
• Benchmarks
• Citation

Models

Roughly sorted according to the timeline supported in FLA

Date	Model	Title	Paper	Code	FLA impl
2023-07	RetNet	Retentive network: a successor to transformer for large language models	arxiv	official	code
2023-12	GLA	Gated Linear Attention Transformers with Hardware-Efficient Training	arxiv	official	code
2023-12	Based	An Educational and Effective Sequence Mixer	blog	official	code
2024-01	Rebased	Linear Transformers with Learnable Kernel Functions are Better In-Context Models	arxiv	official	code
2021-02	Delta Net	Linear Transformers Are Secretly Fast Weight Programmers	arxiv	official	code
2021-10	ABC	Attention with Bounded-memory Control	arxiv		code
2023-09	HGRN	Hierarchically Gated Recurrent Neural Network for Sequence Modeling	openreview	official	code
2024-04	HGRN2	HGRN2: Gated Linear RNNs with State Expansion	arxiv	official	code
2024-04	RWKV6	Eagle and Finch: RWKV with Matrix-Valued States and Dynamic Recurrence	arxiv	official	code
2024-06	Samba	Samba: Simple Hybrid State Space Models for Efficient Unlimited Context Language Modeling	arxiv	official	code
2024-05	Mamba2	Transformers are SSMs: Generalized Models and Efficient Algorithms Through Structured State Space Duality	arxiv	official	code
2024-09	GSA	Gated Slot Attention for Efficient Linear-Time Sequence Modeling	arxiv	official	code

Installation

The following requirements should be satisfied

• PyTorch >= 2.0
• Triton >=2.2
• einops

As fla is actively developed now, no released packages are provided at this time. If you do need to use fla ops/modules and contemplate further explorations, an alternative way is to install the package from source

pip install -U git+https://github.com/sustcsonglin/flash-linear-attention

or manage fla with submodules

git submodule add https://github.com/sustcsonglin/flash-linear-attention.git 3rdparty/flash-linear-attention
ln -s 3rdparty/flash-linear-attention/fla fla

Caution

If you're not working with Triton v2.2 or its nightly release, it's important to be aware of potential issues with the FusedChunk implementation, detailed in this issue. You can run the test python tests/test_fused_chunk.py to check if your version is affected by similar compiler problems. While we offer some fixes for Triton<=2.1, be aware that these may result in reduced performance.

For both Triton 2.2 and earlier versions (up to 2.1), you can reliably use the Chunk version (with hidden states materialized into HBMs). After careful optimization, this version generally delivers high performance in most scenarios.

Usage

Token Mixing

We provide ``token mixing'' linear attention layers in fla.layers for you to use. You can replace the standard multihead attention layer in your model with other linear attention layers. Example usage is as follows:

>>> import torch
>>> from fla.layers import MultiScaleRetention
>>> batch_size, num_heads, seq_len, hidden_size = 32, 4, 2048, 1024
>>> device, dtype = 'cuda:0', torch.bfloat16
>>> retnet = MultiScaleRetention(hidden_size=hidden_size, num_heads=num_heads).to(device=device, dtype=dtype)
>>> retnet
MultiScaleRetention(
  (q_proj): Linear(in_features=1024, out_features=1024, bias=False)
  (k_proj): Linear(in_features=1024, out_features=1024, bias=False)
  (v_proj): Linear(in_features=1024, out_features=2048, bias=False)
  (g_proj): Linear(in_features=1024, out_features=2048, bias=False)
  (o_proj): Linear(in_features=2048, out_features=1024, bias=False)
  (g_norm_swish_gate): FusedRMSNormSwishGate(512, eps=1e-05)
  (rotary): RotaryEmbedding()
)
>>> x = torch.randn(batch_size, seq_len, hidden_size).to(device=device, dtype=dtype)
>>> y, *_ = retnet(x)
>>> y.shape
torch.Size([32, 2048, 1024])

We provide the implementations of models that are compatible with 🤗 Transformers library. Here's an example of how to initialize a GLA model from the default configs in fla:

>>> from fla.models import GLAConfig
>>> from transformers import AutoModelForCausalLM
>>> config = GLAConfig()
>>> config
GLAConfig {
  "attn": null,
  "attn_mode": "chunk",
  "bos_token_id": 1,
  "clamp_min": null,
  "conv_size": 4,
  "elementwise_affine": true,
  "eos_token_id": 2,
  "expand_k": 0.5,
  "expand_v": 1,
  "feature_map": null,
  "fuse_cross_entropy": true,
  "fuse_norm": true,
  "hidden_act": "swish",
  "hidden_ratio": 4,
  "hidden_size": 2048,
  "initializer_range": 0.02,
  "intermediate_size": null,
  "max_position_embeddings": 2048,
  "model_type": "gla",
  "norm_eps": 1e-06,
  "num_heads": 4,
  "num_hidden_layers": 24,
  "num_kv_heads": null,
  "tie_word_embeddings": false,
  "transformers_version": "4.45.0",
  "use_cache": true,
  "use_gk": true,
  "use_gv": false,
  "use_output_gate": true,
  "use_short_conv": false,
  "vocab_size": 32000
}

>>> AutoModelForCausalLM.from_config(config)
GLAForCausalLM(
  (model): GLAModel(
    (embeddings): Embedding(32000, 2048)
    (layers): ModuleList(
      (0-23): 24 x GLABlock(
        (attn_norm): RMSNorm(2048, eps=1e-06)
        (attn): GatedLinearAttention(
          (q_proj): Linear(in_features=2048, out_features=1024, bias=False)
          (k_proj): Linear(in_features=2048, out_features=1024, bias=False)
          (v_proj): Linear(in_features=2048, out_features=2048, bias=False)
          (g_proj): Linear(in_features=2048, out_features=2048, bias=False)
          (gk_proj): Sequential(
            (0): Linear(in_features=2048, out_features=16, bias=False)
            (1): Linear(in_features=16, out_features=1024, bias=True)
          )
          (o_proj): Linear(in_features=2048, out_features=2048, bias=False)
          (g_norm_swish_gate): FusedRMSNormSwishGate(512, eps=1e-06)
        )
        (mlp_norm): RMSNorm(2048, eps=1e-06)
        (mlp): GLAMLP(
          (gate_proj): Linear(in_features=2048, out_features=11264, bias=False)
          (down_proj): Linear(in_features=5632, out_features=2048, bias=False)
          (act_fn): SiLU()
        )
      )
    )
    (norm): RMSNorm(2048, eps=1e-06)
  )
  (lm_head): Linear(in_features=2048, out_features=32000, bias=False)
)

Fused Modules

We offer a collection of fused modules in fla.modules to facilitate faster training:

• Rotary Embedding: rotary positional embeddings as adopted by the Llama architecture, a.k.a., Transformer++.
• Norm Layers:
  • RMSNorm, LayerNorm and GroupNorm
  • RMSNormLinear, LayerNormLinear and GroupNormLinear to reduce memory usage of intermediate tensors for improved memory efficiency.
• Norm Layers with Gating: combine norm layers with element-wise gating, as used by RetNet/GLA.
• Cross Entropy: faster Triton implementation of cross entropy loss.
• Linear Cross Entropy: fused linear layer and cross entropy loss to avoid the materialization of large logits tensors. Also refer to implementations by mgmalek and Liger-Kernel.
• Linear KL Divergence: fused linear layer and KL divergence loss in a similar vein as CE loss.

Generation

Upon successfully pretraining a model, it becomes accessible for generating text using the 🤗 text generation APIs. In the following, we give a generation example:

>>> import fla
>>> from transformers import AutoModelForCausalLM, AutoTokenizer
>>> name = 'fla-hub/gla-1.3B-100B'
>>> tokenizer = AutoTokenizer.from_pretrained(name)
>>> model = AutoModelForCausalLM.from_pretrained(name).cuda()
>>> input_prompt = "Power goes with permanence. Impermanence is impotence. And rotation is castration."
>>> input_ids = tokenizer(input_prompt, return_tensors="pt").input_ids.cuda()
>>> outputs = model.generate(input_ids, max_length=64)
>>> tokenizer.batch_decode(outputs, skip_special_tokens=True)[0]

We also provide a simple script here for benchmarking the generation speed. Simply run it by:

$ python -m benchmarks.benchmark_generation \
  --path 'fla-hub/gla-1.3B-100B' \
  --repetition_penalty 2. \
  --prompt="Hello everyone, I'm Songlin Yang"

Prompt:
Hello everyone, I'm Songlin Yang
Generated:
Hello everyone, I'm Songlin Yang.
I am a 20 year old girl from China who is currently studying in the United States of America for my Master degree and also working as an English teacher at school here on campus since last summer (1st semester). My main goal to be able do well with this course so that we can have

Prompt length: 10, generation length: 64
Total prompt processing + decoding time: 4593ms

All of the pretrained models currently available can be found in fla-hub.

>>> from huggingface_hub import list_models
>>> for model in list_models(author='fla-hub'): print(model.id)

Hybrid Models

fla provides a flexible method to incorporate standard attention layers into existing linear attention models. This is easily achieved by specifying the attn argument in the model configuration.

For example, to create a 2-layer Samba model with interleaved Mamba and local attention layers, using a sliding window size of 2048:

>>> from fla.models import SambaConfig
>>> from transformers import AutoModelForCausalLM
>>> config = SambaConfig(num_hidden_layers=2)
>>> config.attn = { 
  'layers': [1], 
  'num_heads': 18, 
  'num_kv_heads': 18,
  'window_size': 2048
}
>>> config
SambaConfig {
  "attn": {
    "layers": [
      1
    ],
    "num_heads": 18,
    "num_kv_heads": 18,
    "window_size": 2048
  },
  "bos_token_id": 1,
  "conv_kernel": 4,
  "eos_token_id": 2,
  "expand": 2,
  "fuse_cross_entropy": true,
  "fuse_norm": true,
  "hidden_act": "silu",
  "hidden_ratio": 4,
  "hidden_size": 2304,
  "initializer_range": 0.02,
  "intermediate_size": 4608,
  "max_position_embeddings": 2048,
  "model_type": "samba",
  "norm_eps": 1e-05,
  "num_hidden_layers": 2,
  "pad_token_id": 0,
  "rescale_prenorm_residual": false,
  "residual_in_fp32": false,
  "state_size": 16,
  "tie_word_embeddings": false,
  "time_step_floor": 0.0001,
  "time_step_init_scheme": "random",
  "time_step_max": 0.1,
  "time_step_min": 0.001,
  "time_step_rank": 144,
  "time_step_scale": 1.0,
  "transformers_version": "4.45.0",
  "use_bias": false,
  "use_cache": true,
  "use_conv_bias": true,
  "vocab_size": 32000
}

>>> AutoModelForCausalLM.from_config(config)
SambaForCausalLM(
  (backbone): SambaModel(
    (embeddings): Embedding(32000, 2304)
    (layers): ModuleList(
      (0): SambaBlock(
        (mixer_norm): RMSNorm(2304, eps=1e-05)
        (mixer): MambaMixer(
          (conv1d): Conv1d(4608, 4608, kernel_size=(4,), stride=(1,), padding=(3,), groups=4608)
          (act): SiLU()
          (in_proj): Linear(in_features=2304, out_features=9216, bias=False)
          (x_proj): Linear(in_features=4608, out_features=176, bias=False)
          (dt_proj): Linear(in_features=144, out_features=4608, bias=True)
          (out_proj): Linear(in_features=4608, out_features=2304, bias=False)
        )
        (mlp_norm): RMSNorm(2304, eps=1e-05)
        (mlp): SambaMLP(
          (gate_proj): Linear(in_features=2304, out_features=12288, bias=False)
          (down_proj): Linear(in_features=6144, out_features=2304, bias=False)
          (act_fn): SiLU()
        )
      )
      (1): SambaBlock(
        (mixer_norm): RMSNorm(2304, eps=1e-05)
        (mixer): Attention(
          (q_proj): Linear(in_features=2304, out_features=2304, bias=False)
          (k_proj): Linear(in_features=2304, out_features=2304, bias=False)
          (v_proj): Linear(in_features=2304, out_features=2304, bias=False)
          (o_proj): Linear(in_features=2304, out_features=2304, bias=False)
          (rotary): RotaryEmbedding()
        )
        (mlp_norm): RMSNorm(2304, eps=1e-05)
        (mlp): SambaMLP(
          (gate_proj): Linear(in_features=2304, out_features=12288, bias=False)
          (down_proj): Linear(in_features=6144, out_features=2304, bias=False)
          (act_fn): SiLU()
        )
      )
    )
    (norm_f): RMSNorm(2304, eps=1e-05)
  )
  (lm_head): Linear(in_features=2304, out_features=32000, bias=False)
)

During inference, you DO NOT need to revise anything for generation! The model will produce output as-is, without any need for additional configurations or modifications.

Evaluations

The lm-evaluation-harness library allows you to easily perform (zero-shot) model evaluations. Follow the steps below to use this library:

1. Install lm_eval following their instructions.

2. Run evaluation with:

$ PATH='fla-hub/gla-1.3B-100B'
$ python -m evals.harness --model hf \
    --model_args pretrained=$PATH,dtype=bfloat16 \
    --tasks wikitext,lambada_openai,piqa,hellaswag,winogrande,arc_easy,arc_challenge,boolq,sciq,copa,openbookqa \
    --batch_size 64 \
    --num_fewshot 0 \
    --device cuda \
    --show_config                  

We've made fla compatible with hf-style evaluations, you can call evals.harness to finish the evaluations. Running the command above will provide the task results reported in the GLA paper.

Tip

If you are using lm-evaluation-harness as an external library and can't find (almost) any tasks available, before calling lm_eval.evaluate() or lm_eval.simple_evaluate(), simply run the following to load the library's stock tasks!

>>> from lm_eval.tasks import TaskManager; TaskManager().initialize_tasks()

Benchmarks

We compared our Triton-based RetNet implementation with CUDA-based FlashAttention2, using a batch size of 8, 32 heads, and a head dimension of 128, across different sequence lengths. These tests were conducted on a single A100 80GB GPU, as illustrated in the following graph

# you might have to first install `fla` to enable its import via `pip install -e .`
$ python benchmark_retention.py
Performance:
   seq_len  fused_chunk_fwd  chunk_fwd  parallel_fwd  fused_chunk_fwdbwd  chunk_fwdbwd  parallel_fwdbwd  flash_fwd  flash_fwdbwd
0    128.0         0.093184   0.185344      0.067584            1.009664      1.591296         1.044480   0.041984      0.282624
1    256.0         0.165888   0.219136      0.126976            1.024000      1.596928         1.073152   0.074752      0.413696
2    512.0         0.308224   0.397312      0.265216            1.550336      1.603584         1.301504   0.156672      0.883712
3   1024.0         0.603136   0.747520      0.706560            3.044864      3.089408         3.529728   0.467968      2.342912
4   2048.0         1.191424   1.403904      2.141184            6.010880      6.059008        11.009024   1.612800      7.135232
5   4096.0         2.377728   2.755072      7.392256           11.932672     11.938816        37.792770   5.997568     24.435200
6   8192.0         4.750336   5.491712     26.402817           23.759359     23.952385       141.014023  22.682114     90.619904
7  16384.0         9.591296  10.870784    101.262337           47.666176     48.745472       539.853821  91.346947    346.318848

Citation

If you find this repo useful, please consider citing our works:

@inproceedings{yang2024gla,
  title     = {Gated Linear Attention Transformers with Hardware-Efficient Training},
  author    = {Yang, Songlin and Wang, Bailin and Shen, Yikang and Panda, Rameswar and Kim, Yoon},
  booktitle = {Proceedings of ICML},
  year      = {2024}
}

@software{yang2024fla,
  title  = {FLA: A Triton-Based Library for Hardware-Efficient Implementations of Linear Attention Mechanism},
  author = {Yang, Songlin and Zhang, Yu},
  url    = {https://github.com/sustcsonglin/flash-linear-attention},
  month  = jan,
  year   = {2024}
}

@inproceedings{yang2024parallelizing,
  title     = {Parallelizing Linear Transformers with the Delta Rule over Sequence Length},
  author    = {Yang, Songlin and Wang, Bailin and Zhang, Yu and Shen, Yikang and Kim, Yoon},
  booktitle = {Proceedings of NeurIPS},
  year      = {2024}
}

@inproceedings{zhang2024gsa,
  title     = {Gated Slot Attention for Efficient Linear-Time Sequence Modeling},
  author    = {Zhang, Yu and Yang, Songlin and Zhu, Ruijie and Zhang, Yue and Cui, Leyang and Wang, Yiqiao and Wang, Bolun and Shi, Freda and Wang, Bailin and Bi, Wei and Zhou, Peng and Fu, Guohong},
  booktitle = {Proceedings of NeurIPS},
  year      = {2024}
}