注意
跳转至页面底部 下载完整示例代码。
(测试版) 使用缩放点积注意力 (SDPA) 实现高性能 Transformer#
创建日期：2023年3月15日 | 最后更新：2024年10月9日 | 最后验证：2024年11月5日
作者： Driss Guessous
摘要#
在本教程中，我们希望介绍一个对实现 Transformer 架构非常有用的新 torch.nn.functional 函数。该函数名为 torch.nn.functional.scaled_dot_product_attention。关于该函数的详细描述，请参阅 PyTorch 文档。该函数已被集成到 torch.nn.MultiheadAttention 和 torch.nn.TransformerEncoderLayer 中。
概述#
从宏观层面看，这个 PyTorch 函数根据论文 《Attention is all you need》 中的定义，计算查询 (query)、键 (key) 和值 (value) 之间的缩放点积注意力 (SDPA)。虽然该函数可以使用 PyTorch 的现有功能实现，但融合 (fused) 实现版本相比普通的实现能带来巨大的性能优势。
融合实现 (Fused implementations)#
对于 CUDA 张量输入，该函数将分发到以下实现之一：
使用 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([[[ 0.5959, 0.1429, 1.5727, 0.3284, -0.6335, 0.5448, 0.3931,
-0.4219],
[ 0.7123, -0.4191, 1.6644, 0.2195, -0.3375, 0.3153, 0.2167,
-1.7448],
[ 0.8462, -0.6929, 1.8627, 0.0798, -0.2044, 0.2906, 0.1808,
-2.2236]],
[[ 0.8980, -0.7502, -0.4390, -0.7741, -0.5893, 0.0356, 0.2328,
0.1365],
[ 1.6219, -0.0187, -0.1638, -0.0094, 0.0052, 0.1599, 0.6293,
-0.0462],
[ 1.0669, -0.9946, 0.1146, -0.8969, -0.0444, -0.0584, -0.9421,
-0.1227]]], device='cuda:0')
显式调度器控制 (Explicit Dispatcher Control)#
尽管该函数会自动隐式分发到上述三种实现之一，但用户也可以通过使用上下文管理器显式控制分发。此上下文管理器允许用户明确禁用某些实现。如果用户想要确保函数确实为他们的特定输入使用了最快的实现，可以使用该上下文管理器通过测试性能来进行筛选。
# 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 2274.641 microseconds
The math implementation runs in 87478.186 microseconds
The flash attention implementation runs in 2273.650 microseconds
The memory efficient implementation runs in 4336.528 microseconds
硬件依赖性 (Hardware dependence)#
根据您运行上述代码单元的机器及可用硬件，结果可能会有所不同。- 如果您没有 GPU 并且是在 CPU 上运行，那么在使用 FP32 时，上下文管理器将不起作用，三次运行应该会返回相似的耗时。- 根据您的显卡支持的计算能力，flash attention 或 memory efficient 实现可能会执行失败。
因果自注意力 (Causal Self Attention)#
以下是一个多头因果自注意力模块的实现示例，灵感来自 Andrej Karpathy 的 NanoGPT 仓库。
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 (Dense) 张量支持#
SDPA 同时支持 NestedTensor 和密集张量输入。NestedTensors 处理输入为一系列不等长序列的批次的情况，而无需将每个序列填充 (pad) 到该批次中的最大长度。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.")
/var/lib/ci-user/.local/lib/python3.10/site-packages/torch/nested/__init__.py:254: 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 /__w/pytorch/pytorch/aten/src/ATen/NestedTensorImpl.cpp:177.)
return _nested.nested_tensor(
Random NT runs in 597.888 microseconds
Random Dense runs in 951.113 microseconds
在 torch.compile 中使用 SDPA#
随着 PyTorch 2.0 的发布，引入了一个名为 torch.compile() 的新功能，它相比 Eager 模式可以提供显著的性能提升。缩放点积注意力与 torch.compile() 完全兼容。为了证明这一点，让我们使用 torch.compile() 编译 CausalSelfAttention 模块，并观察所带来的性能提升。
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 424.933 microseconds
The compiled module runs in 547.996 microseconds
确切的执行时间取决于机器，但我的测试结果是：非编译模块运行耗时 166.616 微秒，编译后的模块运行耗时 166.726 微秒。这与我们的预期不符。让我们深入挖掘一下。PyTorch 提供了一个出色的内置性能分析器 (profiler)，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 20.54% 2.744ms 75.79% 10.126ms 10.126ms 0.000us 0.00% 10.851ms 10.851ms 1
Non-Compilied Causal Attention 0.00% 0.000us 0.00% 0.000us 0.000us 10.720ms 100.87% 10.720ms 10.720ms 1
aten::linear 0.98% 130.603us 32.46% 4.337ms 86.732us 0.000us 0.00% 8.022ms 160.440us 50
aten::matmul 1.73% 231.285us 29.26% 3.909ms 78.188us 0.000us 0.00% 8.022ms 160.440us 50
aten::mm 9.36% 1.251ms 25.80% 3.447ms 68.948us 7.799ms 73.38% 8.022ms 160.440us 50
ampere_fp16_s1688gemm_fp16_128x128_ldg8_f2f_tn 0.00% 0.000us 0.00% 0.000us 0.000us 5.584ms 52.54% 5.584ms 223.356us 25
aten::scaled_dot_product_attention 1.72% 229.312us 15.01% 2.006ms 80.225us 0.000us 0.00% 2.829ms 113.155us 25
aten::_scaled_dot_product_flash_attention 2.22% 297.159us 13.30% 1.776ms 71.052us 0.000us 0.00% 2.829ms 113.155us 25
aten::_flash_attention_forward 2.66% 355.858us 9.90% 1.323ms 52.907us 2.829ms 26.62% 2.829ms 113.155us 25
void pytorch_flash::flash_fwd_kernel<Flash_fwd_kerne... 0.00% 0.000us 0.00% 0.000us 0.000us 2.829ms 26.62% 2.829ms 113.155us 25
------------------------------------------------------- ------------ ------------ ------------ ------------ ------------ ------------ ------------ ------------ ------------ ------------
Self CPU time total: 13.360ms
Self CUDA time total: 10.628ms
------------------------------------------------------- ------------ ------------ ------------ ------------ ------------ ------------ ------------ ------------ ------------ ------------
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 10.749ms 101.26% 10.749ms 10.749ms 1
Compiled Causal Attention 10.30% 1.330ms 87.67% 11.317ms 11.317ms 0.000us 0.00% 10.615ms 10.615ms 1
Torch-Compiled Region: 0/0 10.22% 1.319ms 74.71% 9.644ms 385.777us 0.000us 0.00% 10.615ms 424.603us 25
CompiledFunction 8.52% 1.099ms 62.42% 8.058ms 322.322us 0.000us 0.00% 10.615ms 424.603us 25
## Call CompiledFxGraph fut22jywhsl3h6u7mliip2eqy2i7... 14.37% 1.855ms 53.91% 6.959ms 278.348us 0.000us 0.00% 10.615ms 424.603us 25
aten::mm 7.95% 1.026ms 12.59% 1.625ms 32.502us 7.800ms 73.48% 7.800ms 155.995us 50
ampere_fp16_s1688gemm_fp16_128x128_ldg8_f2f_tn 0.00% 0.000us 0.00% 0.000us 0.000us 5.583ms 52.60% 5.583ms 223.322us 25
aten::_scaled_dot_product_flash_attention 2.11% 271.858us 13.66% 1.764ms 70.544us 0.000us 0.00% 2.815ms 112.613us 25
aten::_flash_attention_forward 2.84% 366.063us 10.01% 1.292ms 51.696us 2.815ms 26.52% 2.815ms 112.613us 25
void pytorch_flash::flash_fwd_kernel<Flash_fwd_kerne... 0.00% 0.000us 0.00% 0.000us 0.000us 2.815ms 26.52% 2.815ms 112.613us 25
------------------------------------------------------- ------------ ------------ ------------ ------------ ------------ ------------ ------------ ------------ ------------ ------------
Self CPU time total: 12.909ms
Self CUDA time total: 10.615ms
前面的代码片段生成了一份报告，列出了在编译和非编译模块中消耗 GPU 执行时间最多的前 10 个 PyTorch 函数。分析显示，两个模块在 GPU 上花费的大部分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 就是这样），那么 PyTorch 的开销就可以被隐藏。
实际上，您的模块通常不会仅由一个单一的 CausalSelfAttention 块组成。在使用 Andrej Karpathy 的 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 缩短到了 3273.17ms！这是在 NanoGPT 训练 Shakespeare 数据集的 commit ae3a8d5 上完成的。
将 SDPA 与 attn_bias 子类配合使用#
# 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 上下文管理器来确保在 GPU 上使用特定的实现。此外，我们构建了一个简单的 CausalSelfAttention 模块，它可以与 NestedTensor 一起工作并可被 torch 编译。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。
脚本总运行时间： (0 分钟 7.309 秒)