注意
点击此处下载完整的示例代码
（Beta）使用缩放点积注意力 (SDPA) 实现高性能 Transformer¶
创建于：2023 年 3 月 15 日 | 最后更新：2024 年 10 月 09 日 | 最后验证：2024 年 11 月 05 日
作者： Driss Guessous
摘要¶
在本教程中，我们想要重点介绍一个新的 torch.nn.functional 函数，该函数对于实现 Transformer 架构非常有用。该函数名为 torch.nn.functional.scaled_dot_product_attention。有关该函数的详细描述，请参阅 PyTorch 文档。此函数已集成到 torch.nn.MultiheadAttention 和 torch.nn.TransformerEncoderLayer 中。
概述¶
从高层次来看，此 PyTorch 函数根据论文 Attention is all you need 中定义的公式，计算查询、键和值之间的缩放点积注意力 (SDPA)。虽然可以使用现有的 PyTorch 函数编写此函数，但融合实现可以比朴素实现提供更大的性能优势。
融合实现¶
对于 CUDA 张量输入，该函数将调度到以下实现之一
用 C++ 定义的 PyTorch 实现
注意
本教程需要 PyTorch 2.0.0 或更高版本。
import torch
import torch.nn as nn
import torch.nn.functional as F
device = "cuda" if torch.cuda.is_available() else "cpu"
# Example Usage:
query, key, value = torch.randn(2, 3, 8, device=device), torch.randn(2, 3, 8, device=device), torch.randn(2, 3, 8, device=device)
F.scaled_dot_product_attention(query, key, value)
tensor([[[-1.3321, -0.3489, 0.3015, -0.3912, 0.9867, 0.3137, -0.0691,
-1.2593],
[-1.0882, 0.2506, 0.6491, 0.1360, 0.5238, -0.2448, -0.0820,
-0.6171],
[-1.0012, 0.3990, 0.6441, -0.0277, 0.5325, -0.2564, -0.0607,
-0.6404]],
[[ 0.6091, 0.0708, 0.6188, 0.3252, -0.1598, 0.4197, -0.2335,
0.0630],
[ 0.5285, 0.3890, -0.2649, 0.3706, -0.3839, 0.1963, -0.6242,
0.2312],
[ 0.4048, 0.0762, 0.3777, 0.4689, -0.2978, 0.2754, -0.6429,
0.1037]]], device='cuda:0')
显式调度器控制¶
虽然该函数将隐式调度到三个实现之一，但用户也可以通过使用上下文管理器显式控制调度。此上下文管理器允许用户显式禁用某些实现。如果用户想要确保该函数确实为其特定输入使用了最快的实现，则可以使用上下文管理器来扫描测量性能。
# Lets define a helpful benchmarking function:
import torch.utils.benchmark as benchmark
def benchmark_torch_function_in_microseconds(f, *args, **kwargs):
t0 = benchmark.Timer(
stmt="f(*args, **kwargs)", globals={"args": args, "kwargs": kwargs, "f": f}
)
return t0.blocked_autorange().mean * 1e6
# Lets define the hyper-parameters of our input
batch_size = 32
max_sequence_len = 1024
num_heads = 32
embed_dimension = 32
dtype = torch.float16
query = torch.rand(batch_size, num_heads, max_sequence_len, embed_dimension, device=device, dtype=dtype)
key = torch.rand(batch_size, num_heads, max_sequence_len, embed_dimension, device=device, dtype=dtype)
value = torch.rand(batch_size, num_heads, max_sequence_len, embed_dimension, device=device, dtype=dtype)
print(f"The default implementation runs in {benchmark_torch_function_in_microseconds(F.scaled_dot_product_attention, query, key, value):.3f} microseconds")
# Lets explore the speed of each of the 3 implementations
from torch.nn.attention import SDPBackend, sdpa_kernel
with sdpa_kernel(SDPBackend.MATH):
math_time=benchmark_torch_function_in_microseconds(F.scaled_dot_product_attention, query, key, value)
print(f"The math implementation runs in {math_time:.3f} microseconds")
with sdpa_kernel(SDPBackend.FLASH_ATTENTION):
try:
flash_time=benchmark_torch_function_in_microseconds(F.scaled_dot_product_attention, query, key, value)
print(f"The flash attention implementation runs in {flash_time:.3f} microseconds")
except RuntimeError:
print("FlashAttention is not supported. See warnings for reasons.")
with sdpa_kernel(SDPBackend.EFFICIENT_ATTENTION):
try:
efficient_time=benchmark_torch_function_in_microseconds(F.scaled_dot_product_attention, query, key, value)
print(f"The memory efficient implementation runs in {efficient_time:.3f} microseconds")
except RuntimeError:
print("EfficientAttention is not supported. See warnings for reasons.")
The default implementation runs in 2326.842 microseconds
The math implementation runs in 87127.629 microseconds
The flash attention implementation runs in 2332.811 microseconds
The memory efficient implementation runs in 4343.845 microseconds
硬件依赖性¶
根据您运行上述单元格的机器以及可用的硬件，您的结果可能会有所不同。- 如果您没有 GPU 并且在 CPU 上以 FP32 运行，则上下文管理器将不起作用，并且所有三次运行应返回相似的计时。- 根据您的显卡支持的计算能力，Flash Attention 或内存高效注意力可能会失败。
因果自注意力¶
以下是受 Andrej Karpathy NanoGPT 仓库启发的，多头因果自注意力模块的示例实现。
class CausalSelfAttention(nn.Module):
def __init__(self, num_heads: int, embed_dimension: int, bias: bool=False, is_causal: bool=False, dropout:float=0.0):
super().__init__()
assert embed_dimension % num_heads == 0
# key, query, value projections for all heads, but in a batch
self.c_attn = nn.Linear(embed_dimension, 3 * embed_dimension, bias=bias)
# output projection
self.c_proj = nn.Linear(embed_dimension, embed_dimension, bias=bias)
# regularization
self.dropout = dropout
self.resid_dropout = nn.Dropout(dropout)
self.num_heads = num_heads
self.embed_dimension = embed_dimension
# Perform causal masking
self.is_causal = is_causal
def forward(self, x):
# calculate query, key, values for all heads in batch and move head forward to be the batch dim
query_projected = self.c_attn(x)
batch_size = query_projected.size(0)
embed_dim = query_projected.size(2)
head_dim = embed_dim // (self.num_heads * 3)
query, key, value = query_projected.chunk(3, -1)
query = query.view(batch_size, -1, self.num_heads, head_dim).transpose(1, 2)
key = key.view(batch_size, -1, self.num_heads, head_dim).transpose(1, 2)
value = value.view(batch_size, -1, self.num_heads, head_dim).transpose(1, 2)
if self.training:
dropout = self.dropout
is_causal = self.is_causal
else:
dropout = 0.0
is_causal = False
y = F.scaled_dot_product_attention(query, key, value, attn_mask=None, dropout_p=dropout, is_causal=is_causal)
y = y.transpose(1, 2).view(batch_size, -1, self.num_heads * head_dim)
y = self.resid_dropout(self.c_proj(y))
return y
num_heads = 8
heads_per_dim = 64
embed_dimension = num_heads * heads_per_dim
dtype = torch.float16
model = CausalSelfAttention(num_heads=num_heads, embed_dimension=embed_dimension, bias=False, is_causal=True, dropout=0.1).to("cuda").to(dtype).eval()
print(model)
CausalSelfAttention(
(c_attn): Linear(in_features=512, out_features=1536, bias=False)
(c_proj): Linear(in_features=512, out_features=512, bias=False)
(resid_dropout): Dropout(p=0.1, inplace=False)
)
NestedTensor 和密集张量支持¶
SDPA 同时支持 NestedTensor 和密集张量输入。NestedTensors 处理输入是一批可变长度序列的情况，而无需将每个序列填充到批次中的最大长度。有关 NestedTensors 的更多信息，请参阅 torch.nested 和 NestedTensors 教程。
import random
def generate_rand_batch(
batch_size,
max_sequence_len,
embed_dimension,
pad_percentage=None,
dtype=torch.float16,
device="cuda",
):
if not pad_percentage:
return (
torch.randn(
batch_size,
max_sequence_len,
embed_dimension,
dtype=dtype,
device=device,
),
None,
)
# Random sequence lengths
seq_len_list = [
int(max_sequence_len * (1 - random.gauss(pad_percentage, 0.01)))
for _ in range(batch_size)
]
# Make random entry in the batch have max sequence length
seq_len_list[random.randint(0, batch_size - 1)] = max_sequence_len
return (
torch.nested.nested_tensor(
[
torch.randn(seq_len, embed_dimension,
dtype=dtype, device=device)
for seq_len in seq_len_list
]
),
seq_len_list,
)
random_nt, _ = generate_rand_batch(32, 512, embed_dimension, pad_percentage=0.5, dtype=dtype, device=device)
random_dense, _ = generate_rand_batch(32, 512, embed_dimension, pad_percentage=None, dtype=dtype, device=device)
# Currently the fused implementations don't support ``NestedTensor`` for training
model.eval()
with sdpa_kernel(SDPBackend.FLASH_ATTENTION):
try:
print(f"Random NT runs in {benchmark_torch_function_in_microseconds(model, random_nt):.3f} microseconds")
print(f"Random Dense runs in {benchmark_torch_function_in_microseconds(model, random_dense):.3f} microseconds")
except RuntimeError:
print("FlashAttention is not supported. See warnings for reasons.")
/usr/local/lib/python3.10/dist-packages/torch/nested/__init__.py:228: UserWarning:
The PyTorch API of nested tensors is in prototype stage and will change in the near future. We recommend specifying layout=torch.jagged when constructing a nested tensor, as this layout receives active development, has better operator coverage, and works with torch.compile. (Triggered internally at /pytorch/aten/src/ATen/NestedTensorImpl.cpp:178.)
Random NT runs in 563.476 microseconds
Random Dense runs in 947.731 microseconds
将 SDPA 与 torch.compile 一起使用¶
随着 PyTorch 2.0 的发布，引入了一项名为 torch.compile() 的新功能，它可以比 eager 模式提供显着的性能改进。缩放点积注意力与 torch.compile() 完全可组合。为了演示这一点，让我们使用 torch.compile() 编译 CausalSelfAttention 模块，并观察由此产生的性能改进。
batch_size = 32
max_sequence_len = 256
x = torch.rand(batch_size, max_sequence_len,
embed_dimension, device=device, dtype=dtype)
print(
f"The non compiled module runs in {benchmark_torch_function_in_microseconds(model, x):.3f} microseconds")
compiled_model = torch.compile(model)
# Let's compile it
compiled_model(x)
print(
f"The compiled module runs in {benchmark_torch_function_in_microseconds(compiled_model, x):.3f} microseconds")
The non compiled module runs in 415.033 microseconds
The compiled module runs in 515.460 microseconds
确切的执行时间取决于机器，但我的结果是：未编译的模块运行时间为 166.616 微秒，编译后的模块运行时间为 166.726 微秒。 这不是我们所期望的。 让我们深入研究一下。 PyTorch 附带了一个出色的内置性能分析器，您可以使用它来检查代码的性能特征。
from torch.profiler import profile, record_function, ProfilerActivity
activities = [ProfilerActivity.CPU]
if device == 'cuda':
activities.append(ProfilerActivity.CUDA)
with profile(activities=activities, record_shapes=False) as prof:
with record_function(" Non-Compilied Causal Attention"):
for _ in range(25):
model(x)
print(prof.key_averages().table(sort_by="cuda_time_total", row_limit=10))
with profile(activities=activities, record_shapes=False) as prof:
with record_function("Compiled Causal Attention"):
for _ in range(25):
compiled_model(x)
print(prof.key_averages().table(sort_by="cuda_time_total", row_limit=10))
# For even more insights, you can export the trace and use ``chrome://tracing`` to view the results
#
# .. code-block:: python
#
# prof.export_chrome_trace("compiled_causal_attention_trace.json").
------------------------------------------------------- ------------ ------------ ------------ ------------ ------------ ------------ ------------ ------------ ------------ ------------
Name Self CPU % Self CPU CPU total % CPU total CPU time avg Self CUDA Self CUDA % CUDA total CUDA time avg # of Calls
------------------------------------------------------- ------------ ------------ ------------ ------------ ------------ ------------ ------------ ------------ ------------ ------------
Non-Compilied Causal Attention 0.00% 0.000us 0.00% 0.000us 0.000us 10.559ms 101.70% 10.559ms 10.559ms 1
Non-Compilied Causal Attention 20.46% 2.262ms 75.78% 8.377ms 8.377ms 0.000us 0.00% 10.383ms 10.383ms 1
aten::linear 1.14% 126.082us 28.04% 3.100ms 61.993us 0.000us 0.00% 7.781ms 155.623us 50
aten::matmul 2.25% 248.826us 24.10% 2.664ms 53.285us 0.000us 0.00% 7.781ms 155.623us 50
aten::mm 15.06% 1.665ms 19.47% 2.152ms 43.047us 7.781ms 74.94% 7.781ms 155.623us 50
ampere_fp16_s1688gemm_fp16_128x128_ldg8_f2f_tn 0.00% 0.000us 0.00% 0.000us 0.000us 5.572ms 53.67% 5.572ms 222.899us 25
aten::scaled_dot_product_attention 1.96% 216.958us 18.01% 1.990ms 79.616us 0.000us 0.00% 2.601ms 104.057us 25
aten::_scaled_dot_product_flash_attention 3.02% 334.253us 16.04% 1.773ms 70.938us 0.000us 0.00% 2.601ms 104.057us 25
aten::_flash_attention_forward 3.74% 413.039us 11.22% 1.241ms 49.633us 2.601ms 25.06% 2.601ms 104.057us 25
void pytorch_flash::flash_fwd_kernel<pytorch_flash::... 0.00% 0.000us 0.00% 0.000us 0.000us 2.601ms 25.06% 2.601ms 104.057us 25
------------------------------------------------------- ------------ ------------ ------------ ------------ ------------ ------------ ------------ ------------ ------------ ------------
Self CPU time total: 11.055ms
Self CUDA time total: 10.383ms
------------------------------------------------------- ------------ ------------ ------------ ------------ ------------ ------------ ------------ ------------ ------------ ------------
Name Self CPU % Self CPU CPU total % CPU total CPU time avg Self CUDA Self CUDA % CUDA total CUDA time avg # of Calls
------------------------------------------------------- ------------ ------------ ------------ ------------ ------------ ------------ ------------ ------------ ------------ ------------
Compiled Causal Attention 0.00% 0.000us 0.00% 0.000us 0.000us 10.463ms 100.71% 10.463ms 10.463ms 1
Compiled Causal Attention 9.15% 1.015ms 74.41% 8.258ms 8.258ms 0.000us 0.00% 10.389ms 10.389ms 1
Torch-Compiled Region: 2/0 8.42% 934.690us 63.27% 7.022ms 280.881us 0.000us 0.00% 10.389ms 415.559us 25
CompiledFunction 25.95% 2.880ms 54.85% 6.087ms 243.494us 0.000us 0.00% 10.389ms 415.559us 25
aten::mm 9.55% 1.059ms 14.12% 1.568ms 31.353us 7.784ms 74.92% 7.784ms 155.676us 50
ampere_fp16_s1688gemm_fp16_128x128_ldg8_f2f_tn 0.00% 0.000us 0.00% 0.000us 0.000us 5.576ms 53.67% 5.576ms 223.034us 25
aten::_scaled_dot_product_flash_attention 2.17% 240.384us 14.77% 1.640ms 65.593us 0.000us 0.00% 2.605ms 104.208us 25
aten::_flash_attention_forward 3.87% 429.608us 10.87% 1.206ms 48.248us 2.605ms 25.08% 2.605ms 104.208us 25
void pytorch_flash::flash_fwd_kernel<pytorch_flash::... 0.00% 0.000us 0.00% 0.000us 0.000us 2.605ms 25.08% 2.605ms 104.208us 25
ampere_fp16_s1688gemm_fp16_128x128_ldg8_f2f_stages_3... 0.00% 0.000us 0.00% 0.000us 0.000us 2.208ms 21.25% 2.208ms 88.318us 25
------------------------------------------------------- ------------ ------------ ------------ ------------ ------------ ------------ ------------ ------------ ------------ ------------
Self CPU time total: 11.099ms
Self CUDA time total: 10.389ms
之前的代码片段生成了一份报告，列出了编译模块和未编译模块中消耗 GPU 执行时间最多的前 10 个 PyTorch 函数。分析表明，对于这两个模块，大部分 GPU 时间都集中在同一组函数上。这里的原因是 torch.compile 非常擅长消除与 PyTorch 相关的框架开销。如果您的模型正在启动大型、高效的 CUDA 内核（在本例中 CausalSelfAttention 就是这种情况），那么 PyTorch 的开销可能会被隐藏。
实际上，您的模块通常不只包含一个 CausalSelfAttention 块。当使用 Andrej Karpathy NanoGPT 仓库进行实验时，编译模块将每个训练步骤的时间从 6090.49ms 缩短到 3273.17ms！ 这是在 NanoGPT 训练 Shakespeare 数据集的 commit：ae3a8d5 上完成的。
将 SDPA 与 attn_bias 子类一起使用¶
# As of PyTorch 2.3, we have added a new submodule that contains tensor subclasses.
# Designed to be used with ``torch.nn.functional.scaled_dot_product_attention``.
# The module is named ``torch.nn.attention.bias`` and contains the following two
# utilities for generating causal attention variants:
#
# - ``torch.nn.attention.bias.causal_upper_left``
# - ``torch.nn.attention.bias.causal_lower_right``
#
# .. note::
# The current argument ``is_causal`` in ``torch.nn.functional.scaled_dot_product_attention``
# is the same as using ``torch.nn.attention.bias.causal_upper_left``.
#
from torch.nn.attention.bias import causal_lower_right, causal_upper_left
batch_size = 32
sequence_length_q = 2
sequence_length_kv = 10
num_heads = 16
embed_dimension = 32
dtype = torch.float16
query = torch.rand(batch_size, num_heads, sequence_length_q, embed_dimension, device=device, dtype=dtype)
key = torch.rand(batch_size, num_heads, sequence_length_kv, embed_dimension, device=device, dtype=dtype)
value = torch.rand(batch_size, num_heads, sequence_length_kv, embed_dimension, device=device, dtype=dtype)
upper_left_bias = causal_upper_left(sequence_length_q, sequence_length_kv)
lower_right_bias = causal_lower_right(sequence_length_q, sequence_length_kv)
print(type(upper_left_bias))
print(type(lower_right_bias))
assert type(upper_left_bias) == type(lower_right_bias)
assert issubclass(type(upper_left_bias), torch.Tensor)
# As you can see from the previous output, are the same type ``torch.nn.attention.bias.CausalBias``
# and subclass ``torch.Tensor``
# Lets see what these tensors look like
print(upper_left_bias)
print(lower_right_bias)
# Upper Left Bias aligns the causal attention mask to the upper left corner of the attention scores matrix.
# This only has an impact when the attention scores matrix is not square, which is common for decoding use cases.
# Another way of thinking about this concept is that when you use upper left bias,
# the 0th token in the query is aligned to the 0th token in the key, while for lower right bias,
# Assuming the attention score matrix is two dimensional, ``attn_score[0][0]`` is the attention score
# between the 0th token in the query and the 0th token in the key.
# For lower right bias, the sequence of q is aligned so that the last token in q is aligned to the last token in k
# (for example, ``attn_score[-1][-1])`` is all True since the last token in q is at the same position as the last token in k
# even if the sequence length of q and k are different.
# These objects are intended to be used with sdpa
out_upper_left = F.scaled_dot_product_attention(query, key, value, upper_left_bias)
out_lower_right = F.scaled_dot_product_attention(query, key, value, lower_right_bias)
out_is_causal = F.scaled_dot_product_attention(query, key, value, is_causal=True)
assert torch.allclose(out_upper_left, out_is_causal)
assert not torch.allclose(out_upper_left, out_lower_right)
# These attention biases should also be compatible with torch.compile
compiled_sdpa = torch.compile(F.scaled_dot_product_attention, fullgraph=True)
out_upper_left = compiled_sdpa(query, key, value, upper_left_bias)
<class 'torch.nn.attention.bias.CausalBias'>
<class 'torch.nn.attention.bias.CausalBias'>
tensor([[ True, False, False, False, False, False, False, False, False, False],
[ True, True, False, False, False, False, False, False, False, False]])
tensor([[ True, True, True, True, True, True, True, True, True, False],
[ True, True, True, True, True, True, True, True, True, True]])
结论¶
在本教程中，我们演示了 torch.nn.functional.scaled_dot_product_attention 的基本用法。我们展示了如何使用 sdpa_kernel 上下文管理器来断言在 GPU 上使用了特定的实现。此外，我们构建了一个简单的 CausalSelfAttention 模块，该模块与 NestedTensor 配合使用并且是 torch 可编译的。在此过程中，我们展示了如何使用性能分析工具来探索用户定义模块的性能特征。
脚本的总运行时间： ( 0 分钟 7.314 秒)
