# FlashAttention
This repository provides the official implementation of FlashAttention and
FlashAttention-2 from the
following papers.
**FlashAttention: Fast and Memory-Efficient Exact Attention with IO-Awareness**
Tri Dao, Daniel Y. Fu, Stefano Ermon, Atri Rudra, Christopher Ré
Paper: https://arxiv.org/abs/2205.14135
IEEE Spectrum [article](https://spectrum.ieee.org/mlperf-rankings-2022) about our submission to the MLPerf 2.0 benchmark using FlashAttention.
**FlashAttention-2: Faster Attention with Better Parallelism and Work Partitioning**
Tri Dao
Paper: https://tridao.me/publications/flash2/flash2.pdf
## Usage
We've been very happy to see FlashAttention being widely adopted in such a short
time after its release. This [page](https://github.com/Dao-AILab/flash-attention/blob/main/usage.md)
contains a partial list of places where FlashAttention is being used.
FlashAttention and FlashAttention-2 are free to use and modify (see LICENSE).
Please cite and credit FlashAttention if you use it.
## Installation and features
Requirements:
- CUDA 11.6 and above.
- PyTorch 1.12 and above.
- Linux. Might work for Windows starting v2.3.2 (we've seen a few positive [reports](https://github.com/Dao-AILab/flash-attention/issues/595)) but Windows compilation still requires more testing. If you have ideas on how to set up prebuilt CUDA wheels for Windows, please reach out via Github issue.
We recommend the
[Pytorch](https://catalog.ngc.nvidia.com/orgs/nvidia/containers/pytorch)
container from Nvidia, which has all the required tools to install FlashAttention.
To install:
1. Make sure that PyTorch is installed.
2. Make sure that `packaging` is installed (`pip install packaging`)
3. Make sure that `ninja` is installed and that it works correctly (e.g. `ninja
--version` then `echo $?` should return exit code 0). If not (sometimes `ninja
--version` then `echo $?` returns a nonzero exit code), uninstall then reinstall
`ninja` (`pip uninstall -y ninja && pip install ninja`). Without `ninja`,
compiling can take a very long time (2h) since it does not use multiple CPU
cores. With `ninja` compiling takes 3-5 minutes on a 64-core machine.
4. Then:
```sh
pip install flash-attn --no-build-isolation
```
Alternatively you can compile from source:
```sh
python setup.py install
```
If your machine has less than 96GB of RAM and lots of CPU cores, `ninja` might
run too many parallel compilation jobs that could exhaust the amount of RAM. To
limit the number of parallel compilation jobs, you can set the environment
variable `MAX_JOBS`:
```sh
MAX_JOBS=4 pip install flash-attn --no-build-isolation
```
Interface: `src/flash_attention_interface.py`
FlashAttention-2 currently supports:
1. Ampere, Ada, or Hopper GPUs (e.g., A100, RTX 3090, RTX 4090, H100). Support for Turing
GPUs (T4, RTX 2080) is coming soon, please use FlashAttention 1.x for Turing
GPUs for now.
2. Datatype fp16 and bf16 (bf16 requires Ampere, Ada, or Hopper GPUs).
3. All head dimensions up to 256. ~~Head dim > 192 backward requires A100/A800 or H100/H800~~. Head dim 256 backward now works on consumer GPUs (if there's no dropout) as of flash-attn 2.5.5.
## How to use FlashAttention
The main functions implement scaled dot product attention (softmax(Q @ K^T *
softmax_scale) @ V):
```python
from flash_attn import flash_attn_qkvpacked_func, flash_attn_func
```
```python
flash_attn_qkvpacked_func(qkv, dropout_p=0.0, softmax_scale=None, causal=False,
window_size=(-1, -1), alibi_slopes=None, deterministic=False):
"""dropout_p should be set to 0.0 during evaluation
If Q, K, V are already stacked into 1 tensor, this function will be faster than
calling flash_attn_func on Q, K, V since the backward pass avoids explicit concatenation
of the gradients of Q, K, V.
If window_size != (-1, -1), implements sliding window local attention. Query at position i
will only attend to keys between [i - window_size[0], i + window_size[1]] inclusive.
Arguments:
qkv: (batch_size, seqlen, 3, nheads, headdim)
dropout_p: float. Dropout probability.
softmax_scale: float. The scaling of QK^T before applying softmax.
Default to 1 / sqrt(headdim).
causal: bool. Whether to apply causal attention mask (e.g., for auto-regressive modeling).
window_size: (left, right). If not (-1, -1), implements sliding window local attention.
alibi_slopes: (nheads,) or (batch_size, nheads), fp32. A bias of (-alibi_slope * |i - j|) is added to
the attention score of query i and key j.
deterministic: bool. Whether to use the deterministic implementation of the backward pass,
which is slightly slower and uses more memory. The forward pass is always deterministic.
Return:
out: (batch_size, seqlen, nheads, headdim).
"""
```
```python
flash_attn_func(q, k, v, dropout_p=0.0, softmax_scale=None, causal=False,
window_size=(-1, -1), alibi_slopes=None, deterministic=False):
"""dropout_p should be set to 0.0 during evaluation
Supports multi-query and grouped-query attention (MQA/GQA) by passing in KV with fewer heads
than Q. Note that the number of heads in Q must be divisible by the number of heads in KV.
For example, if Q has 6 heads and K, V have 2 heads, head 0, 1, 2 of Q will attention to head
0 of K, V, and head 3, 4, 5 of Q will attention to head 1 of K, V.
If window_size != (-1, -1), implements sliding window local attention. Query at position i
will only attend to keys between
[i + seqlen_k - seqlen_q - window_size[0], i + seqlen_k - seqlen_q + window_size[1]] inclusive.
Arguments:
q: (batch_size, seqlen, nheads, headdim)
k: (batch_size, seqlen, nheads_k, headdim)
v: (batch_size, seqlen, nheads_k, headdim)
dropout_p: float. Dropout probability.
softmax_scale: float. The scaling of QK^T before applying softmax.
Default to 1 / sqrt(headdim).
causal: bool. Whether to apply causal attention mask (e.g., for auto-regressive modeling).
window_size: (left, right). If not (-1, -1), implements sliding window local attention.
alibi_slopes: (nheads,) or (batch_size, nheads), fp32. A bias of
(-alibi_slope * |i + seqlen_k - seqlen_q - j|)
is added to the attention score of query i and key j.
deterministic: bool. Whether to use the deterministic implementation of the backward pass,
which is slightly slower and uses more memory. The forward pass is always deterministic.
Return:
out: (batch_size, seqlen, nheads, headdim).
"""
```
```python
def flash_attn_with_kvcache(
q,
k_cache,
v_cache,
k=None,
v=None,
rotary_cos=None,
rotary_sin=None,
cache_seqlens: Optional[Union[(int, torch.Tensor)]] = None,
cache_batch_idx: Optional[torch.Tensor] = None,
block_table: Optional[torch.Tensor] = None,
softmax_scale=None,
causal=False,
window_size=(-1, -1), # -1 means infinite context window
rotary_interleaved=True,
alibi_slopes=None,
):
"""
If k and v are not None, k_cache and v_cache will be updated *inplace* with the new values from
k and v. This is useful for incremental decoding: you can pass in the cached keys/values from
the previous step, and update them with the new keys/values from the current step, and do
attention with the updated cache, all in 1 kernel.
If you pass in k / v, you must make sure that the cache is large enough to hold the new values.
For example, the KV cache could be pre-allocated with the max sequence length, and you can use
cache_seqlens to keep track of the current sequence lengths of each sequence in the batch.
Also apply rotary embedding if rotary_cos and rotary_sin are passed in. The key @k will be
rotated by rotary_cos and rotary_sin at indices cache_seqlens, cache_seqlens + 1, etc.
If causal or local (i.e., window_size != (-1, -1)), the query @q will be rotated by rotary_cos
and rotary_sin at indices cache_seqlens, cache_seqlens + 1, etc.
If not causal and not local, the query @q will be rotated by rotary_cos and rotary_sin at
indices cache_seqlens only (i.e. we consider all tokens in @q to be at position cache_seqlens).
See tests/test_flash_attn.py::test_flash_attn_kvcache for examples of how to use this function.
Supports multi-query and grouped-query attention (MQA/GQA) by passing in KV with fewer heads
than Q. Note that the number of heads in Q must be divisible by the number of heads in KV.
For example, if Q has 6 heads and K, V have 2 heads, head 0, 1, 2 of Q will attention to head
0 of K, V, and head 3, 4, 5 of Q will attention to head 1 of K, V.
If causal=True, the causal mask is aligned to the bottom right corner of the attention matrix.
For example, if seqlen_q = 2 and seqlen_k = 5, the causal mask (1 = keep, 0 = masked out) is:
1 1 1 1 0
1 1 1 1 1
If seqlen_q = 5 and seqlen_k = 2, the causal mask is:
0 0
0 0
0 0
1 0
1 1
If the row of the mask is all zero, the output will be zero.
If window_size != (-1, -1), implements sliding window local attention. Query at position i
will only attend to keys between
[i + seqlen_k - seqlen_q - window_size[0], i + seqlen_k - seqlen_q + window_size[1]] inclusive.
Note: Does not support backward pass.
Arguments:
q: (batch_size, seqlen, nheads, headdim)
k_cache: (batch_size_cache, seqlen_cache, nheads_k, headdim) if there's no block_table,
or (num_blocks, page_block_size, nheads_k, headdim) if there's a block_table (i.e. paged KV cache)
page_block_size must be a multiple of 256.
v_cache: (batch_size_cache, seqlen_cache, nheads_k, headdim) if there's no block_table,
or (num_blocks, page_block_size, nheads_k, headdim) if there's a block_table (i.e. paged KV cache)
k [optional]: (batch_size, seqlen_new, nheads_k, headdim). If not None, we concatenate
k with k_cache, starting at the indices specified by cache_seqlens.
v [optional]: (batch_size, seqlen_new, nheads_k, headdim). Similar to k.
rotary_cos [optional]: (seqlen_ro, rotary_dim / 2). If not None, we apply rotary embedding
to k and q. Only applicable if k and v are passed in. rotary_dim must be divisible by 16.
rotary_sin [optional]: (seqlen_ro, rotary_dim / 2). Similar to rotary_cos.
cache_seqlens: int, or (batch_size,), dtype torch.int32. The sequence lengths of the
KV cache.
block_table [optional]: (batch_size, max_num_blocks_per_seq), dtype torch.int32.
cache_batch_idx: (batch_size,), dtype torch.int32. The indices used to index into the KV cache.
If None, we assume that the batch indices are [0, 1, 2, ..., batch_size - 1].
If the indices are not distinct, and k and v are provided, the values updated in the cache
might come from any of the duplicate indices.
softmax_scale: float. The scaling of QK^T before applying softmax.
Default to 1 / sqrt(headdim).
causal: bool. Whether to apply causal attention mask (e.g., for auto-regressive modeling).
window_size: (left, right). If not (-1, -1), implements sliding window local attention.
rotary_interleaved: bool. Only applicable if rotary_cos and rotary_sin are passed in.
If True, rotary embedding will combine dimensions 0 & 1, 2 & 3, etc. If False,
rotary embedding will combine dimensions 0 & rotary_dim / 2, 1 & rotary_dim / 2 + 1
(i.e. GPT-NeoX style).
alibi_slopes: (nheads,) or (batch_size, nheads), fp32. A bias of
(-alibi_slope * |i + seqlen_k - seqlen_q - j|)
is added to the attention score of query i and key j.
Return:
out: (batch_size, seqlen, nheads, headdim).
"""
```
To see how these functions are used in a multi-head attention layer (which
includes QKV projection, output projection), see the MHA [implementation](https://github.com/Dao-AILab/flash-attention/blob/main/flash_attn/modules/mha.py).
## Changelog
### 2.0: Complete rewrite, 2x faster
Upgrading from FlashAttention (1.x) to FlashAttention-2
These functions have been renamed:
- `flash_attn_unpadded_func` -> `flash_attn_varlen_func`
- `flash_attn_unpadded_qkvpacked_func` -> `flash_attn_varlen_qkvpacked_func`
- `flash_attn_unpadded_kvpacked_func` -> `flash_attn_varlen_kvpacked_func`
If the inputs have the same sequence lengths in the same batch, it is simpler
and faster to use these functions:
```python
flash_attn_qkvpacked_func(qkv, dropout_p=0.0, softmax_scale=None, causal=False)
```
```python
flash_attn_func(q, k, v, dropout_p=0.0, softmax_scale=None, causal=False)
```
### 2.1: Change behavior of causal flag
If seqlen_q != seqlen_k and causal=True, the causal mask is aligned to the
bottom right corner of the attention matrix, instead of the top-left corner.
For example, if seqlen_q = 2 and seqlen_k = 5, the causal mask (1 = keep, 0 =
masked out) is:
v2.0:
1 0 0 0 0
1 1 0 0 0
v2.1:
1 1 1 1 0
1 1 1 1 1
If seqlen_q = 5 and seqlen_k = 2, the causal mask is:
v2.0:
1 0
1 1
1 1
1 1
1 1
v2.1:
0 0
0 0
0 0
1 0
1 1
If the row of the mask is all zero, the output will be zero.
### 2.2: Optimize for inference
Optimize for inference (iterative decoding) when query has very small sequence
length (e.g., query sequence length = 1). The bottleneck here is to load KV
cache as fast as possible, and we split the loading across different thread
blocks, with a separate kernel to combine results.
See the function `flash_attn_with_kvcache` with more features for inference
(perform rotary embedding, updating KV cache inplace).
Thanks to the xformers team, and in particular Daniel Haziza, for this
collaboration.
### 2.3: Local (i.e., sliding window) attention
Implement sliding window attention (i.e., local attention). Thanks to [Mistral
AI](https://mistral.ai/) and in particular Timothée Lacroix for this
contribution. Sliding window was used in the [Mistral 7B](https://mistral.ai/news/announcing-mistral-7b/) model.
### 2.4: ALiBi (attention with linear bias), deterministic backward pass.
Implement ALiBi (Press et al., 2021). Thanks to Sanghun Cho from Kakao Brain for this contribution.
Implement deterministic backward pass. Thanks to engineers from [Meituan](www.meituan.com) for this contribution.
### 2.5: Paged KV cache.
Support paged KV cache (i.e., [PagedAttention](https://arxiv.org/abs/2309.06180)).
Thanks to @beginlner for this contribution.
## Performance
We present expected speedup (combined forward + backward pass) and memory savings from using FlashAttention against PyTorch standard attention, depending on sequence length, on different GPUs (speedup depends on memory bandwidth - we see more speedup on slower GPU memory).
We currently have benchmarks for these GPUs:
* [A100](#a100)
* [H100](#h100)
<!-- * [RTX 3090](#rtx-3090) -->
<!-- * [T4](#t4) -->
### A100
We display FlashAttention speedup using these parameters:
* Head dimension 64 or 128, hidden dimension 2048 (i.e. either 32 or 16 heads).
* Sequence length 512, 1k, 2k, 4k, 8k, 16k.
* Batch size set to 16k / seqlen.
#### Speedup
#### Memory
We show memory savings in this graph (note that memory footprint is the same no matter if you use dropout or masking).
Memory savings are proportional to sequence length -- since standard attention has memory quadratic in sequence length, whereas FlashAttention has memory linear in sequence length.
We see 10X memory savings at sequence length 2K, and 20X at 4K.
As a result, FlashAttention can scale to much longer sequence lengths.
### H100
## Full model code and training script
We have released the full GPT model
[implementation](https://github.com/Dao-AILab/flash-attention/blob/main/flash_attn/models/gpt.py).
We also provide optimized implementations of other layers (e.g., MLP, LayerNorm,
cross-entropy loss, rotary embedding). Overall this speeds up training by 3-5x
compared to the baseline implementation from Huggingface, reaching up to 225
TFLOPs/sec per A100, equivalent to 72% model FLOPs utilization (we don't need
any activation checkpointing).
We also include a training
[script](https://github.com/Dao-AILab/flash-attention/tree/main/training) to
train GPT2 on Openwebtext and GPT3 on The Pile.
## Triton implementation of FlashAttention
Phil Tillet (OpenAI) has an experimental implementation of FlashAttention in Triton:
https://github.com/openai/triton/blob/master/python/tutorials/06-fused-attention.py
As Triton is a higher-level language than CUDA, it might be easier to understand
and experiment with. The notations in the Triton implementation are also closer
to what's used in our paper.
We also have an experimental implementation in Triton that support attention
bias (e.g. ALiBi):
https://github.com/Dao-AILab/flash-attention/blob/main/flash_attn/flash_attn_triton.py
## Tests
We test that FlashAttention produces the same output and gradient as a reference
implementation, up to some numerical tolerance. In particular, we check that the
maximum numerical error of FlashAttention is at most twice the numerical error
of a baseline implementation in Pytorch (for different head dimensions, input
dtype, sequence length, causal / non-causal).
To run the tests:
```sh
pytest -q -s tests/test_flash_attn.py
```
## When you encounter issues
This new release of FlashAttention-2 has been tested on several GPT-style
models, mostly on A100 GPUs.
If you encounter bugs, please open a GitHub Issue!
## Citation
If you use this codebase, or otherwise found our work valuable, please cite:
```
@inproceedings{dao2022flashattention,
title={Flash{A}ttention: Fast and Memory-Efficient Exact Attention with {IO}-Awareness},
author={Dao, Tri and Fu, Daniel Y. and Ermon, Stefano and Rudra, Atri and R{\'e}, Christopher},
booktitle={Advances in Neural Information Processing Systems},
year={2022}
}
@article{dao2023flashattention2,
title={Flash{A}ttention-2: Faster Attention with Better Parallelism and Work Partitioning},
author={Dao, Tri},
year={2023}
}
```
Raw data
{
"_id": null,
"home_page": "https://github.com/Dao-AILab/flash-attention",
"name": "flash-attn",
"maintainer": null,
"docs_url": null,
"requires_python": ">=3.7",
"maintainer_email": null,
"keywords": null,
"author": "Tri Dao",
"author_email": "trid@cs.stanford.edu",
"download_url": "https://files.pythonhosted.org/packages/72/94/06f618bb338ec7203b48ac542e73087362b7750f9c568b13d213a3f181bb/flash_attn-2.5.8.tar.gz",
"platform": null,
"description": "# FlashAttention\nThis repository provides the official implementation of FlashAttention and\nFlashAttention-2 from the\nfollowing papers.\n\n**FlashAttention: Fast and Memory-Efficient Exact Attention with IO-Awareness** \nTri Dao, Daniel Y. Fu, Stefano Ermon, Atri Rudra, Christopher R\u00e9 \nPaper: https://arxiv.org/abs/2205.14135 \nIEEE Spectrum [article](https://spectrum.ieee.org/mlperf-rankings-2022) about our submission to the MLPerf 2.0 benchmark using FlashAttention.\n\n\n**FlashAttention-2: Faster Attention with Better Parallelism and Work Partitioning** \nTri Dao\n\nPaper: https://tridao.me/publications/flash2/flash2.pdf\n\n\n\n\n## Usage\n\nWe've been very happy to see FlashAttention being widely adopted in such a short\ntime after its release. This [page](https://github.com/Dao-AILab/flash-attention/blob/main/usage.md)\ncontains a partial list of places where FlashAttention is being used.\n\nFlashAttention and FlashAttention-2 are free to use and modify (see LICENSE).\nPlease cite and credit FlashAttention if you use it.\n\n## Installation and features\n\nRequirements:\n- CUDA 11.6 and above.\n- PyTorch 1.12 and above.\n- Linux. Might work for Windows starting v2.3.2 (we've seen a few positive [reports](https://github.com/Dao-AILab/flash-attention/issues/595)) but Windows compilation still requires more testing. If you have ideas on how to set up prebuilt CUDA wheels for Windows, please reach out via Github issue.\n\nWe recommend the\n[Pytorch](https://catalog.ngc.nvidia.com/orgs/nvidia/containers/pytorch)\ncontainer from Nvidia, which has all the required tools to install FlashAttention.\n\nTo install:\n1. Make sure that PyTorch is installed.\n2. Make sure that `packaging` is installed (`pip install packaging`)\n3. Make sure that `ninja` is installed and that it works correctly (e.g. `ninja\n--version` then `echo $?` should return exit code 0). If not (sometimes `ninja\n--version` then `echo $?` returns a nonzero exit code), uninstall then reinstall\n`ninja` (`pip uninstall -y ninja && pip install ninja`). Without `ninja`,\ncompiling can take a very long time (2h) since it does not use multiple CPU\ncores. With `ninja` compiling takes 3-5 minutes on a 64-core machine.\n4. Then:\n```sh\npip install flash-attn --no-build-isolation\n```\nAlternatively you can compile from source:\n```sh\npython setup.py install\n```\n\nIf your machine has less than 96GB of RAM and lots of CPU cores, `ninja` might\nrun too many parallel compilation jobs that could exhaust the amount of RAM. To\nlimit the number of parallel compilation jobs, you can set the environment\nvariable `MAX_JOBS`:\n```sh\nMAX_JOBS=4 pip install flash-attn --no-build-isolation\n```\n\nInterface: `src/flash_attention_interface.py`\n\nFlashAttention-2 currently supports:\n1. Ampere, Ada, or Hopper GPUs (e.g., A100, RTX 3090, RTX 4090, H100). Support for Turing\n GPUs (T4, RTX 2080) is coming soon, please use FlashAttention 1.x for Turing\n GPUs for now.\n2. Datatype fp16 and bf16 (bf16 requires Ampere, Ada, or Hopper GPUs).\n3. All head dimensions up to 256. ~~Head dim > 192 backward requires A100/A800 or H100/H800~~. Head dim 256 backward now works on consumer GPUs (if there's no dropout) as of flash-attn 2.5.5.\n\n\n## How to use FlashAttention\n\nThe main functions implement scaled dot product attention (softmax(Q @ K^T *\nsoftmax_scale) @ V):\n```python\nfrom flash_attn import flash_attn_qkvpacked_func, flash_attn_func\n```\n\n```python\nflash_attn_qkvpacked_func(qkv, dropout_p=0.0, softmax_scale=None, causal=False,\n window_size=(-1, -1), alibi_slopes=None, deterministic=False):\n\"\"\"dropout_p should be set to 0.0 during evaluation\nIf Q, K, V are already stacked into 1 tensor, this function will be faster than\ncalling flash_attn_func on Q, K, V since the backward pass avoids explicit concatenation\nof the gradients of Q, K, V.\nIf window_size != (-1, -1), implements sliding window local attention. Query at position i\nwill only attend to keys between [i - window_size[0], i + window_size[1]] inclusive.\nArguments:\n qkv: (batch_size, seqlen, 3, nheads, headdim)\n dropout_p: float. Dropout probability.\n softmax_scale: float. The scaling of QK^T before applying softmax.\n Default to 1 / sqrt(headdim).\n causal: bool. Whether to apply causal attention mask (e.g., for auto-regressive modeling).\n window_size: (left, right). If not (-1, -1), implements sliding window local attention.\n alibi_slopes: (nheads,) or (batch_size, nheads), fp32. A bias of (-alibi_slope * |i - j|) is added to\n the attention score of query i and key j.\n deterministic: bool. Whether to use the deterministic implementation of the backward pass,\n which is slightly slower and uses more memory. The forward pass is always deterministic.\nReturn:\n out: (batch_size, seqlen, nheads, headdim).\n\"\"\"\n```\n\n```python\nflash_attn_func(q, k, v, dropout_p=0.0, softmax_scale=None, causal=False,\n window_size=(-1, -1), alibi_slopes=None, deterministic=False):\n\"\"\"dropout_p should be set to 0.0 during evaluation\nSupports multi-query and grouped-query attention (MQA/GQA) by passing in KV with fewer heads\nthan Q. Note that the number of heads in Q must be divisible by the number of heads in KV.\nFor example, if Q has 6 heads and K, V have 2 heads, head 0, 1, 2 of Q will attention to head\n0 of K, V, and head 3, 4, 5 of Q will attention to head 1 of K, V.\nIf window_size != (-1, -1), implements sliding window local attention. Query at position i\nwill only attend to keys between\n[i + seqlen_k - seqlen_q - window_size[0], i + seqlen_k - seqlen_q + window_size[1]] inclusive.\n\nArguments:\n q: (batch_size, seqlen, nheads, headdim)\n k: (batch_size, seqlen, nheads_k, headdim)\n v: (batch_size, seqlen, nheads_k, headdim)\n dropout_p: float. Dropout probability.\n softmax_scale: float. The scaling of QK^T before applying softmax.\n Default to 1 / sqrt(headdim).\n causal: bool. Whether to apply causal attention mask (e.g., for auto-regressive modeling).\n window_size: (left, right). If not (-1, -1), implements sliding window local attention.\n alibi_slopes: (nheads,) or (batch_size, nheads), fp32. A bias of\n (-alibi_slope * |i + seqlen_k - seqlen_q - j|)\n is added to the attention score of query i and key j.\n deterministic: bool. Whether to use the deterministic implementation of the backward pass,\n which is slightly slower and uses more memory. The forward pass is always deterministic.\nReturn:\n out: (batch_size, seqlen, nheads, headdim).\n\"\"\"\n```\n\n```python\ndef flash_attn_with_kvcache(\n q,\n k_cache,\n v_cache,\n k=None,\n v=None,\n rotary_cos=None,\n rotary_sin=None,\n cache_seqlens: Optional[Union[(int, torch.Tensor)]] = None,\n cache_batch_idx: Optional[torch.Tensor] = None,\n block_table: Optional[torch.Tensor] = None,\n softmax_scale=None,\n causal=False,\n window_size=(-1, -1), # -1 means infinite context window\n rotary_interleaved=True,\n alibi_slopes=None,\n):\n \"\"\"\n If k and v are not None, k_cache and v_cache will be updated *inplace* with the new values from\n k and v. This is useful for incremental decoding: you can pass in the cached keys/values from\n the previous step, and update them with the new keys/values from the current step, and do\n attention with the updated cache, all in 1 kernel.\n\n If you pass in k / v, you must make sure that the cache is large enough to hold the new values.\n For example, the KV cache could be pre-allocated with the max sequence length, and you can use\n cache_seqlens to keep track of the current sequence lengths of each sequence in the batch.\n\n Also apply rotary embedding if rotary_cos and rotary_sin are passed in. The key @k will be\n rotated by rotary_cos and rotary_sin at indices cache_seqlens, cache_seqlens + 1, etc.\n If causal or local (i.e., window_size != (-1, -1)), the query @q will be rotated by rotary_cos\n and rotary_sin at indices cache_seqlens, cache_seqlens + 1, etc.\n If not causal and not local, the query @q will be rotated by rotary_cos and rotary_sin at\n indices cache_seqlens only (i.e. we consider all tokens in @q to be at position cache_seqlens).\n\n See tests/test_flash_attn.py::test_flash_attn_kvcache for examples of how to use this function.\n\n Supports multi-query and grouped-query attention (MQA/GQA) by passing in KV with fewer heads\n than Q. Note that the number of heads in Q must be divisible by the number of heads in KV.\n For example, if Q has 6 heads and K, V have 2 heads, head 0, 1, 2 of Q will attention to head\n 0 of K, V, and head 3, 4, 5 of Q will attention to head 1 of K, V.\n\n If causal=True, the causal mask is aligned to the bottom right corner of the attention matrix.\n For example, if seqlen_q = 2 and seqlen_k = 5, the causal mask (1 = keep, 0 = masked out) is:\n 1 1 1 1 0\n 1 1 1 1 1\n If seqlen_q = 5 and seqlen_k = 2, the causal mask is:\n 0 0\n 0 0\n 0 0\n 1 0\n 1 1\n If the row of the mask is all zero, the output will be zero.\n\n If window_size != (-1, -1), implements sliding window local attention. Query at position i\n will only attend to keys between\n [i + seqlen_k - seqlen_q - window_size[0], i + seqlen_k - seqlen_q + window_size[1]] inclusive.\n\n Note: Does not support backward pass.\n\n Arguments:\n q: (batch_size, seqlen, nheads, headdim)\n k_cache: (batch_size_cache, seqlen_cache, nheads_k, headdim) if there's no block_table,\n or (num_blocks, page_block_size, nheads_k, headdim) if there's a block_table (i.e. paged KV cache)\n page_block_size must be a multiple of 256.\n v_cache: (batch_size_cache, seqlen_cache, nheads_k, headdim) if there's no block_table,\n or (num_blocks, page_block_size, nheads_k, headdim) if there's a block_table (i.e. paged KV cache)\n k [optional]: (batch_size, seqlen_new, nheads_k, headdim). If not None, we concatenate\n k with k_cache, starting at the indices specified by cache_seqlens.\n v [optional]: (batch_size, seqlen_new, nheads_k, headdim). Similar to k.\n rotary_cos [optional]: (seqlen_ro, rotary_dim / 2). If not None, we apply rotary embedding\n to k and q. Only applicable if k and v are passed in. rotary_dim must be divisible by 16.\n rotary_sin [optional]: (seqlen_ro, rotary_dim / 2). Similar to rotary_cos.\n cache_seqlens: int, or (batch_size,), dtype torch.int32. The sequence lengths of the\n KV cache.\n block_table [optional]: (batch_size, max_num_blocks_per_seq), dtype torch.int32.\n cache_batch_idx: (batch_size,), dtype torch.int32. The indices used to index into the KV cache.\n If None, we assume that the batch indices are [0, 1, 2, ..., batch_size - 1].\n If the indices are not distinct, and k and v are provided, the values updated in the cache\n might come from any of the duplicate indices.\n softmax_scale: float. The scaling of QK^T before applying softmax.\n Default to 1 / sqrt(headdim).\n causal: bool. Whether to apply causal attention mask (e.g., for auto-regressive modeling).\n window_size: (left, right). If not (-1, -1), implements sliding window local attention.\n rotary_interleaved: bool. Only applicable if rotary_cos and rotary_sin are passed in.\n If True, rotary embedding will combine dimensions 0 & 1, 2 & 3, etc. If False,\n rotary embedding will combine dimensions 0 & rotary_dim / 2, 1 & rotary_dim / 2 + 1\n (i.e. GPT-NeoX style).\n alibi_slopes: (nheads,) or (batch_size, nheads), fp32. A bias of\n (-alibi_slope * |i + seqlen_k - seqlen_q - j|)\n is added to the attention score of query i and key j.\n\n Return:\n out: (batch_size, seqlen, nheads, headdim).\n \"\"\"\n```\n\nTo see how these functions are used in a multi-head attention layer (which\nincludes QKV projection, output projection), see the MHA [implementation](https://github.com/Dao-AILab/flash-attention/blob/main/flash_attn/modules/mha.py).\n\n## Changelog\n\n### 2.0: Complete rewrite, 2x faster\nUpgrading from FlashAttention (1.x) to FlashAttention-2\n\nThese functions have been renamed:\n- `flash_attn_unpadded_func` -> `flash_attn_varlen_func`\n- `flash_attn_unpadded_qkvpacked_func` -> `flash_attn_varlen_qkvpacked_func`\n- `flash_attn_unpadded_kvpacked_func` -> `flash_attn_varlen_kvpacked_func`\n\nIf the inputs have the same sequence lengths in the same batch, it is simpler\nand faster to use these functions:\n```python\nflash_attn_qkvpacked_func(qkv, dropout_p=0.0, softmax_scale=None, causal=False)\n```\n```python\nflash_attn_func(q, k, v, dropout_p=0.0, softmax_scale=None, causal=False)\n```\n### 2.1: Change behavior of causal flag\n\nIf seqlen_q != seqlen_k and causal=True, the causal mask is aligned to the\nbottom right corner of the attention matrix, instead of the top-left corner.\n\nFor example, if seqlen_q = 2 and seqlen_k = 5, the causal mask (1 = keep, 0 =\nmasked out) is: \nv2.0: \n 1 0 0 0 0 \n 1 1 0 0 0 \nv2.1: \n 1 1 1 1 0 \n 1 1 1 1 1 \n\nIf seqlen_q = 5 and seqlen_k = 2, the causal mask is: \nv2.0: \n 1 0 \n 1 1 \n 1 1 \n 1 1 \n 1 1 \nv2.1: \n 0 0 \n 0 0 \n 0 0 \n 1 0 \n 1 1 \nIf the row of the mask is all zero, the output will be zero.\n\n### 2.2: Optimize for inference\n\nOptimize for inference (iterative decoding) when query has very small sequence\nlength (e.g., query sequence length = 1). The bottleneck here is to load KV\ncache as fast as possible, and we split the loading across different thread\nblocks, with a separate kernel to combine results.\n\nSee the function `flash_attn_with_kvcache` with more features for inference\n(perform rotary embedding, updating KV cache inplace).\n\nThanks to the xformers team, and in particular Daniel Haziza, for this\ncollaboration.\n\n### 2.3: Local (i.e., sliding window) attention\n\nImplement sliding window attention (i.e., local attention). Thanks to [Mistral\nAI](https://mistral.ai/) and in particular Timoth\u00e9e Lacroix for this\ncontribution. Sliding window was used in the [Mistral 7B](https://mistral.ai/news/announcing-mistral-7b/) model.\n\n### 2.4: ALiBi (attention with linear bias), deterministic backward pass.\n\nImplement ALiBi (Press et al., 2021). Thanks to Sanghun Cho from Kakao Brain for this contribution.\n\nImplement deterministic backward pass. Thanks to engineers from [Meituan](www.meituan.com) for this contribution.\n\n### 2.5: Paged KV cache.\n\nSupport paged KV cache (i.e., [PagedAttention](https://arxiv.org/abs/2309.06180)).\nThanks to @beginlner for this contribution.\n\n## Performance\n\nWe present expected speedup (combined forward + backward pass) and memory savings from using FlashAttention against PyTorch standard attention, depending on sequence length, on different GPUs (speedup depends on memory bandwidth - we see more speedup on slower GPU memory).\n\nWe currently have benchmarks for these GPUs:\n* [A100](#a100)\n* [H100](#h100)\n<!-- * [RTX 3090](#rtx-3090) -->\n<!-- * [T4](#t4) -->\n\n### A100\n\nWe display FlashAttention speedup using these parameters:\n* Head dimension 64 or 128, hidden dimension 2048 (i.e. either 32 or 16 heads).\n* Sequence length 512, 1k, 2k, 4k, 8k, 16k.\n* Batch size set to 16k / seqlen.\n\n#### Speedup\n\n\n\n#### Memory\n\n\n\nWe show memory savings in this graph (note that memory footprint is the same no matter if you use dropout or masking).\nMemory savings are proportional to sequence length -- since standard attention has memory quadratic in sequence length, whereas FlashAttention has memory linear in sequence length.\nWe see 10X memory savings at sequence length 2K, and 20X at 4K.\nAs a result, FlashAttention can scale to much longer sequence lengths.\n\n### H100\n\n\n\n## Full model code and training script\n\nWe have released the full GPT model\n[implementation](https://github.com/Dao-AILab/flash-attention/blob/main/flash_attn/models/gpt.py).\nWe also provide optimized implementations of other layers (e.g., MLP, LayerNorm,\ncross-entropy loss, rotary embedding). Overall this speeds up training by 3-5x\ncompared to the baseline implementation from Huggingface, reaching up to 225\nTFLOPs/sec per A100, equivalent to 72% model FLOPs utilization (we don't need\nany activation checkpointing).\n\nWe also include a training\n[script](https://github.com/Dao-AILab/flash-attention/tree/main/training) to\ntrain GPT2 on Openwebtext and GPT3 on The Pile.\n\n## Triton implementation of FlashAttention\n\nPhil Tillet (OpenAI) has an experimental implementation of FlashAttention in Triton:\nhttps://github.com/openai/triton/blob/master/python/tutorials/06-fused-attention.py\n\nAs Triton is a higher-level language than CUDA, it might be easier to understand\nand experiment with. The notations in the Triton implementation are also closer\nto what's used in our paper.\n\nWe also have an experimental implementation in Triton that support attention\nbias (e.g. ALiBi):\nhttps://github.com/Dao-AILab/flash-attention/blob/main/flash_attn/flash_attn_triton.py\n\n\n## Tests\nWe test that FlashAttention produces the same output and gradient as a reference\nimplementation, up to some numerical tolerance. In particular, we check that the\nmaximum numerical error of FlashAttention is at most twice the numerical error\nof a baseline implementation in Pytorch (for different head dimensions, input\ndtype, sequence length, causal / non-causal).\n\nTo run the tests:\n```sh\npytest -q -s tests/test_flash_attn.py\n```\n## When you encounter issues\n\nThis new release of FlashAttention-2 has been tested on several GPT-style\nmodels, mostly on A100 GPUs.\n\nIf you encounter bugs, please open a GitHub Issue!\n\n## Citation\nIf you use this codebase, or otherwise found our work valuable, please cite:\n```\n@inproceedings{dao2022flashattention,\n title={Flash{A}ttention: Fast and Memory-Efficient Exact Attention with {IO}-Awareness},\n author={Dao, Tri and Fu, Daniel Y. and Ermon, Stefano and Rudra, Atri and R{\\'e}, Christopher},\n booktitle={Advances in Neural Information Processing Systems},\n year={2022}\n}\n@article{dao2023flashattention2,\n title={Flash{A}ttention-2: Faster Attention with Better Parallelism and Work Partitioning},\n author={Dao, Tri},\n year={2023}\n}\n```\n",
"bugtrack_url": null,
"license": null,
"summary": "Flash Attention: Fast and Memory-Efficient Exact Attention",
"version": "2.5.8",
"project_urls": {
"Homepage": "https://github.com/Dao-AILab/flash-attention"
},
"split_keywords": [],
"urls": [
{
"comment_text": "",
"digests": {
"blake2b_256": "729406f618bb338ec7203b48ac542e73087362b7750f9c568b13d213a3f181bb",
"md5": "cf764329c672c3df0e4e228042ec479d",
"sha256": "2e5b2bcff6d5cff40d494af91ecd1eb3c5b4520a6ce7a0a8b1f9c1ed129fb402"
},
"downloads": -1,
"filename": "flash_attn-2.5.8.tar.gz",
"has_sig": false,
"md5_digest": "cf764329c672c3df0e4e228042ec479d",
"packagetype": "sdist",
"python_version": "source",
"requires_python": ">=3.7",
"size": 2474099,
"upload_time": "2024-04-26T23:45:51",
"upload_time_iso_8601": "2024-04-26T23:45:51.981346Z",
"url": "https://files.pythonhosted.org/packages/72/94/06f618bb338ec7203b48ac542e73087362b7750f9c568b13d213a3f181bb/flash_attn-2.5.8.tar.gz",
"yanked": false,
"yanked_reason": null
}
],
"upload_time": "2024-04-26 23:45:51",
"github": true,
"gitlab": false,
"bitbucket": false,
"codeberg": false,
"github_user": "Dao-AILab",
"github_project": "flash-attention",
"travis_ci": false,
"coveralls": false,
"github_actions": true,
"lcname": "flash-attn"
}
JSON
