评价此页
nn.Transformer 与嵌套张量和 torch.compile()"

通过使用嵌套张量和 torch.compile() 替换 nn.Transformer 来加速 PyTorch Transformer#

作者: Mikayla Gawarecki

您将学到什么
  • 了解 PyTorch 提供的低级构建块,以构建自定义 Transformer 层(嵌套张量、scaled_dot_product_attentiontorch.compile()FlexAttention

  • 以 MultiHeadAttention 为例,了解上述功能如何改进内存使用和性能

  • 使用上述构建块探索高级定制

先决条件
  • PyTorch v.2.6.0 或更高版本

在过去的几年中,PyTorch 团队开发了各种低级特性,这些特性组合起来可以创建各种 Transformer 变体。这些特性包括:

  • 具有 torch.jagged 布局的嵌套张量(又称 NJT)

  • scaled_dot_product_attention

  • torch.compile()

  • FlexAttention

本教程将简要概述上述技术,并演示如何将它们组合起来,以生成具有改进用户体验的灵活且高性能的 Transformer 层。

可能会观察到 torch.nn 模块目前提供了各种与 Transformer 相关的层。特别是,它包括 TransformerEncoderLayerTransformerEncoderTransformerDecoderLayerTransformerDecoderTransformerMultiheadAttention。这系列层最初是根据 Attention is All You Need 论文实现的。本教程中讨论的组件在现有 nn 层之上提供了改进的用户体验、灵活性和性能。

这个教程适合我吗?#

如果您想了解 torch 库提供了哪些构建块来编写您自己的 Transformer 层以及最佳实践,那么您来对地方了。请继续阅读!

如果您正在寻找流行 Transformer 架构的开箱即用实现,请注意有许多开源库提供它们,包括:

如果您只对高性能注意力分数修改感兴趣,请查看 FlexAttention 博客,其中包含一个 掩码操场

介绍构建块#

首先,我们将简要介绍引言中提到的四项技术

嵌套张量推广了常规密集张量的形状,允许使用相同的张量 UX 表示不规则大小的数据。在 Transformer 的上下文中,我们可以将嵌套张量视为表示可变序列长度的工具。它们消除了对显式填充和掩码(例如 nn.MultiHeadAttention 中的 key_padding_mask)这种易出错做法的需求。

scaled_dot_product_attention\(\text{softmax}(\frac{QK^T}{\sqrt{E}} + B)V\) 的一个原语,它会分派到操作符的融合实现或回退实现。它在 eager 模式(即 PyTorch 的默认使用模式,其中操作在遇到时即时执行)下开箱即用,并且还与 torch.compile() 无缝集成。截至 2.6 版本,它还将原生提供分组查询注意力。

torch.compile() 是 2.0 版本引入的编译器,能够捕获 PyTorch 代码的图并对其执行各种优化,例如将一系列操作融合在一起。具有 torch.jagged 布局的嵌套张量和 scaled_dot_product_attention 可以与编译器无缝协作。在 Transformer 的上下文中,将编译器与嵌套张量和 SDPA 结合使用的价值在于,编译器可以消除在 eager 模式下看到的框架开销,并将 Transformer 中一系列操作(例如投影和激活)融合在一起。

FlexAttention 是一个原语,允许用户在 softmax 操作之前修改注意力分数。它推广了上述用于 scaled_dot_product_attention 的加法 B 项,允许任意计算。它需要编译器才能获得良好的性能。

上述构建块是“你所需要的一切”(截至 2024 年 10 月)#

本节的主要前提是,大多数 Transformer 变体都是 GPT 风格的,由嵌入层、位置编码、注意力块和前馈网络等层组成。如果我们要尝试对该领域中的差异进行分类,我们可能会得出以下结论:

  1. 层类型(激活函数如 SwiGLU 等,归一化函数如 RMSNorm 等,位置编码如正弦、旋转。)

  2. 层顺序,例如在哪里应用范数和位置编码。

  3. 注意力分数的修改,例如 ALiBi、相对位置偏差等。

在预编译器环境中,您可能会编写一个自定义 Transformer 并注意到它功能正常但速度很慢。为了解决这个问题,您可能会为特定的操作序列开发一个自定义融合内核。在编译器环境中,您可以简单地执行初始步骤,然后进行编译并从改进的性能中获益。

多头注意力#

请记住,MultiheadAttention 接受查询、键和值,并由一个输入投影、一个 scaled_dot_product_attention 运算符和一个输出投影组成。我们在这里想展示的主要亮点是,当我们用嵌套张量替换填充/掩码输入时所带来的改进。这些改进体现在三个方面:

  • 用户体验 请记住,nn.MultiheadAttention 要求 querykeyvalue 是稠密 torch.Tensors。它还提供了一个 key_padding_mask,用于掩盖批次中由于不同序列长度而产生的 key 中的填充标记。由于 nn.MHA 中没有 query_padding_mask,用户必须注意适当地掩盖/切分输出以考虑查询序列长度。NestedTensor 干净地消除了这种容易出错的填充掩码的需求。

  • 内存 嵌套张量允许您清晰地表示批次中可变序列长度,而不是具体化一个带有 [B, S] 填充掩码的密集 [B, S, D] 张量(其中 B 是批次大小,S 是批次中的最大序列长度,D 是嵌入大小)。因此,输入和中间激活将使用更少的内存。

  • 性能 由于不具体化填充并跳过不必要的填充计算,性能和内存使用得到改善。

我们将通过在 嵌套张量教程 中构建 MultiheadAttention 层并将其与 nn.MultiheadAttention 层进行比较来演示上述内容。

import torch
import torch.nn as nn
import torch.nn.functional as F


class MultiHeadAttention(nn.Module):
    """
    Computes multi-head attention. Supports nested or padded tensors.

    Args:
        E_q (int): Size of embedding dim for query
        E_k (int): Size of embedding dim for key
        E_v (int): Size of embedding dim for value
        E_total (int): Total embedding dim of combined heads post input projection. Each head
            has dim E_total // nheads
        nheads (int): Number of heads
        dropout (float, optional): Dropout probability. Default: 0.0
        bias (bool, optional): Whether to add bias to input projection. Default: True
    """

    def __init__(
        self,
        E_q: int,
        E_k: int,
        E_v: int,
        E_total: int,
        nheads: int,
        dropout: float = 0.0,
        bias=True,
        device=None,
        dtype=None,
    ):
        factory_kwargs = {"device": device, "dtype": dtype}
        super().__init__()
        self.nheads = nheads
        self.dropout = dropout
        self._qkv_same_embed_dim = E_q == E_k and E_q == E_v
        if self._qkv_same_embed_dim:
            self.packed_proj = nn.Linear(E_q, E_total * 3, bias=bias, **factory_kwargs)
        else:
            self.q_proj = nn.Linear(E_q, E_total, bias=bias, **factory_kwargs)
            self.k_proj = nn.Linear(E_k, E_total, bias=bias, **factory_kwargs)
            self.v_proj = nn.Linear(E_v, E_total, bias=bias, **factory_kwargs)
        E_out = E_q
        self.out_proj = nn.Linear(E_total, E_out, bias=bias, **factory_kwargs)
        assert E_total % nheads == 0, "Embedding dim is not divisible by nheads"
        self.E_head = E_total // nheads
        self.bias = bias

    def forward(
        self,
        query: torch.Tensor,
        key: torch.Tensor,
        value: torch.Tensor,
        attn_mask=None,
        is_causal=False,
    ) -> torch.Tensor:
        """
        Forward pass; runs the following process:
            1. Apply input projection
            2. Split heads and prepare for SDPA
            3. Run SDPA
            4. Apply output projection

        Args:
            query (torch.Tensor): query of shape (``N``, ``L_q``, ``E_qk``)
            key (torch.Tensor): key of shape (``N``, ``L_kv``, ``E_qk``)
            value (torch.Tensor): value of shape (``N``, ``L_kv``, ``E_v``)
            attn_mask (torch.Tensor, optional): attention mask of shape (``N``, ``L_q``, ``L_kv``) to pass to SDPA. Default: None
            is_causal (bool, optional): Whether to apply causal mask. Default: False

        Returns:
            attn_output (torch.Tensor): output of shape (N, L_t, E_q)
        """
        # Step 1. Apply input projection
        if self._qkv_same_embed_dim:
            if query is key and key is value:
                result = self.packed_proj(query)
                query, key, value = torch.chunk(result, 3, dim=-1)
            else:
                q_weight, k_weight, v_weight = torch.chunk(
                    self.packed_proj.weight, 3, dim=0
                )
                if self.bias:
                    q_bias, k_bias, v_bias = torch.chunk(
                        self.packed_proj.bias, 3, dim=0
                    )
                else:
                    q_bias, k_bias, v_bias = None, None, None
                query, key, value = (
                    F.linear(query, q_weight, q_bias),
                    F.linear(key, k_weight, k_bias),
                    F.linear(value, v_weight, v_bias),
                )

        else:
            query = self.q_proj(query)
            key = self.k_proj(key)
            value = self.v_proj(value)

        # Step 2. Split heads and prepare for SDPA
        # reshape query, key, value to separate by head
        # (N, L_t, E_total) -> (N, L_t, nheads, E_head) -> (N, nheads, L_t, E_head)
        query = query.unflatten(-1, [self.nheads, self.E_head]).transpose(1, 2)
        # (N, L_s, E_total) -> (N, L_s, nheads, E_head) -> (N, nheads, L_s, E_head)
        key = key.unflatten(-1, [self.nheads, self.E_head]).transpose(1, 2)
        # (N, L_s, E_total) -> (N, L_s, nheads, E_head) -> (N, nheads, L_s, E_head)
        value = value.unflatten(-1, [self.nheads, self.E_head]).transpose(1, 2)

        # Step 3. Run SDPA
        # (N, nheads, L_t, E_head)
        attn_output = F.scaled_dot_product_attention(
            query, key, value, dropout_p=self.dropout, is_causal=is_causal
        )
        # (N, nheads, L_t, E_head) -> (N, L_t, nheads, E_head) -> (N, L_t, E_total)
        attn_output = attn_output.transpose(1, 2).flatten(-2)

        # Step 4. Apply output projection
        # (N, L_t, E_total) -> (N, L_t, E_out)
        attn_output = self.out_proj(attn_output)

        return attn_output

实用工具#

在本节中,我们提供了一个实用工具,用于使用 Zipf 分布生成半真实数据以获取句子长度。这用于生成嵌套的查询、键和值张量。我们还包含了一个基准测试实用工具。

import numpy as np


def zipf_sentence_lengths(alpha: float, batch_size: int) -> torch.Tensor:
    # generate fake corpus by unigram Zipf distribution
    # from wikitext-2 corpus, we get rank "." = 3, "!" = 386, "?" = 858
    sentence_lengths = np.empty(batch_size, dtype=int)
    for ibatch in range(batch_size):
        sentence_lengths[ibatch] = 1
        word = np.random.zipf(alpha)
        while word != 3 and word != 386 and word != 858:
            sentence_lengths[ibatch] += 1
            word = np.random.zipf(alpha)
    return torch.tensor(sentence_lengths)


# Generate a batch of semi-realistic data using Zipf distribution for sentence lengths
# in the form of nested tensors with the jagged layout.
def gen_batch(N, E_q, E_k, E_v, device, dtype=torch.float32, query_seq_len_1=False):
    # generate semi-realistic data using Zipf distribution for sentence lengths
    sentence_lengths = zipf_sentence_lengths(alpha=1.2, batch_size=N)

    # Note: the torch.jagged layout is a nested tensor layout that supports a single ragged
    # dimension and works with torch.compile. The batch items each have shape (B, S*, D)
    # where B = batch size, S* = ragged sequence length, and D = embedding dimension.
    if query_seq_len_1:
        query = torch.nested.nested_tensor(
            [torch.randn(1, E_q, dtype=dtype, device=device) for l in sentence_lengths],
            layout=torch.jagged,
        )
    else:
        query = torch.nested.nested_tensor(
            [
                torch.randn(l.item(), E_q, dtype=dtype, device=device)
                for l in sentence_lengths
            ],
            layout=torch.jagged,
        )

    key = torch.nested.nested_tensor(
        [
            torch.randn(s.item(), E_k, dtype=dtype, device=device)
            for s in sentence_lengths
        ],
        layout=torch.jagged,
    )

    value = torch.nested.nested_tensor(
        [
            torch.randn(s.item(), E_v, dtype=dtype, device=device)
            for s in sentence_lengths
        ],
        layout=torch.jagged,
    )

    return query, key, value, sentence_lengths


import math
import timeit


def benchmark(func, *args, **kwargs):
    torch.cuda.synchronize()
    torch.cuda.reset_peak_memory_stats()
    begin = timeit.default_timer()
    output = func(*args, **kwargs)
    torch.cuda.synchronize()
    end = timeit.default_timer()
    return output, (end - begin), torch.cuda.max_memory_allocated()

我们现在将演示在 MultiheadAttention 层 + 编译中为自注意力使用嵌套张量所带来的性能改进。我们将此与传统的使用填充和掩码的 nn.MultiheadAttention + 编译进行比较。

N, E_q, E_k, E_v, E_total = 512, 512, 512, 512, 512
E_out = E_q
d_model = E_q
nheads = 8
dropout = 0.0
bias = True
device = "cuda"
torch.manual_seed(6)
query, key, value, sentence_lengths = gen_batch(N, E_q, E_k, E_v, device)
S = sentence_lengths.max().item()
print(
    f"Total sequence length in nested query {sentence_lengths.sum().item()}, max sequence length {S}"
)
padded_query, padded_key, padded_value = (
    t.to_padded_tensor(0.0) for t in (query, key, value)
)

torch.manual_seed(6)
mha_layer = MultiHeadAttention(
    E_q, E_k, E_v, E_total, nheads, dropout=dropout, bias=bias, device="cuda"
)
torch.manual_seed(6)
vanilla_mha_layer = nn.MultiheadAttention(
    E_q, nheads, dropout=dropout, batch_first=True, bias=bias, device="cuda"
)

# ``nn.MultiheadAttention`` uses a non conventional initialization for layers, so do this for exact parity :(
mha_layer.out_proj.weight = nn.Parameter(
    vanilla_mha_layer.out_proj.weight.clone().detach()
)
mha_layer.packed_proj.weight = nn.Parameter(
    vanilla_mha_layer.in_proj_weight.clone().detach()
)
mha_layer.out_proj.bias = nn.Parameter(vanilla_mha_layer.out_proj.bias.clone().detach())
mha_layer.packed_proj.bias = nn.Parameter(
    vanilla_mha_layer.in_proj_bias.clone().detach()
)

new_mha_layer = torch.compile(mha_layer)
# warmup compile
nested_result_warmup = new_mha_layer(query, query, query, is_causal=True)

# benchmark
nested_result, nested_time, nested_peak_memory = benchmark(
    new_mha_layer, query, query, query, is_causal=True
)
padded_nested_result = nested_result.to_padded_tensor(0.0)

# For the vanilla ``nn.MultiheadAttention``, we need to construct the ``key_padding_mask``
# Further, ``nn.MultiheadAttention`` forces one to materialize the ``attn_mask`` even if using ``is_causal``
src_key_padding_mask = torch.where(padded_query == 0.0, -math.inf, 0)[:, :, 0]
attn_mask = torch.empty((N, S, S), device=device).fill_(float("-inf"))
for i, s in enumerate(sentence_lengths):
    attn_mask[i, :s, :s] = nn.Transformer.generate_square_subsequent_mask(s)
attn_mask = attn_mask.unsqueeze(1).expand(N, nheads, S, S).reshape(N * nheads, S, S)

vanilla_mha_layer = torch.compile(vanilla_mha_layer)
# warmup compile
warmup_vanilla_result = vanilla_mha_layer(
    padded_query,
    padded_query,
    padded_query,
    attn_mask=attn_mask,
    key_padding_mask=src_key_padding_mask,
    need_weights=False,
    is_causal=True,
)

# benchmark
(padded_result, _), padded_time, padded_peak_memory = benchmark(
    vanilla_mha_layer,
    padded_query,
    padded_query,
    padded_query,
    key_padding_mask=src_key_padding_mask,
    need_weights=False,
    attn_mask=attn_mask,
    is_causal=True,
)

print(f"{padded_time=:.5f}, padded_peak_memory={padded_peak_memory/1e9:.2f} GB")
print(f"{nested_time=:.5f}, nested_peak_memory={nested_peak_memory/1e9:.2f} GB")
print(
    "Max difference between vanilla and nested result",
    (padded_result - padded_nested_result).abs().max().item(),
)
print(f"Nested speedup: {(padded_time/nested_time):.2f}")
print(
    f"Nested peak memory reduction {((padded_peak_memory - nested_peak_memory)/1e9):.2f} GB"
)
Total sequence length in nested query 10108, max sequence length 114
/usr/local/lib/python3.10/dist-packages/torch/_inductor/compile_fx.py:282: UserWarning:

TensorFloat32 tensor cores for float32 matrix multiplication available but not enabled. Consider setting `torch.set_float32_matmul_precision('high')` for better performance.

padded_time=0.01466, padded_peak_memory=3.23 GB
nested_time=0.00239, nested_peak_memory=0.66 GB
Max difference between vanilla and nested result 0.0
Nested speedup: 6.13
Nested peak memory reduction 2.58 GB

作为参考,以下是 A100 上的一些示例输出

padded_time=0.03454, padded_peak_memory=4.14 GB
nested_time=0.00612, nested_peak_memory=0.76 GB
Max difference between vanilla and nested result 0.0
Nested speedup: 5.65
Nested peak memory reduction 3.39 GB

我们也可以看到反向传播的结果

for i, entry_length in enumerate(sentence_lengths):
    # padding-specific step: remove output projection bias from padded entries for fair comparison
    padded_result[i, entry_length:, :] = 0.0

_, padded_bw_time, padded_bw_peak_mem = benchmark(
    lambda: padded_result.sum().backward()
)
_, nested_bw_time, nested_bw_peak_mem = benchmark(
    lambda: padded_nested_result.sum().backward()
)

print(f"{padded_bw_time=:.5f}, padded_bw_peak_mem={padded_bw_peak_mem/1e9:.2f} GB")
print(f"{nested_bw_time=:.5f}, nested_bw_peak_mem={nested_bw_peak_mem/1e9:.2f} GB")
print(f"Nested backward speedup: {(padded_bw_time/nested_bw_time):.2f}")
print(
    f"Nested backward peak memory reduction {((padded_bw_peak_mem - nested_bw_peak_mem)/1e9):.2f} GB"
)

print(
    "Difference in out_proj.weight.grad",
    (mha_layer.out_proj.weight.grad - vanilla_mha_layer.out_proj.weight.grad)
    .abs()
    .max()
    .item(),
)
print(
    "Difference in packed_proj.weight.grad",
    (mha_layer.packed_proj.weight.grad - vanilla_mha_layer.in_proj_weight.grad)
    .abs()
    .max()
    .item(),
)
print(
    "Difference in out_proj.bias.grad",
    (mha_layer.out_proj.bias.grad - vanilla_mha_layer.out_proj.bias.grad)
    .abs()
    .max()
    .item(),
)
print(
    "Difference in packed_proj.bias.grad",
    (mha_layer.packed_proj.bias.grad - vanilla_mha_layer.in_proj_bias.grad)
    .abs()
    .max()
    .item(),
)
padded_bw_time=1.54396, padded_bw_peak_mem=3.98 GB
nested_bw_time=0.06611, nested_bw_peak_mem=2.54 GB
Nested backward speedup: 23.35
Nested backward peak memory reduction 1.44 GB
Difference in out_proj.weight.grad 0.0002899169921875
Difference in packed_proj.weight.grad 0.001434326171875
Difference in out_proj.bias.grad 0.0
Difference in packed_proj.bias.grad 0.001953125

A100上的样本输出

padded_bw_time=2.09337, padded_bw_peak_mem=5.10 GB
nested_bw_time=0.01452, nested_bw_peak_mem=3.24 GB
Nested backward speedup: 144.13
Nested backward peak memory reduction 1.86 GB
Difference in out_proj.weight.grad 0.000244140625
Difference in packed_proj.weight.grad 0.001556396484375
Difference in out_proj.bias.grad 0.0
Difference in packed_proj.bias.grad 0.001953125

GPT 风格层#

一个基本的 GPT 风格 Transformer 层由一个因果自注意力层和一个带有跳跃连接的前馈网络(FFN)组成。使用上述 MultiheadAttention 层实现这一点相当简单,并且与带有 is_causal=Truenn.TransformerEncoderLayer 具有等效的结果。

为了简洁起见,我们在此教程中省略了实现其余 nn 层的示例,但您可以在此处查看。

更进一步#

到目前为止,我们已经演示了如何实现一个遵循传统 nn.MultiheadAttention 的高性能 MultiheadAttention 层。回顾我们对 Transformer 架构修改的分类,请记住我们将修改分为层类型、层顺序和注意力分数修改。我们相信更改层类型和层顺序(例如将 LayerNorm 替换为 RMSNorm)是相当简单的。

在本节中,我们将讨论使用上述构建块的各种功能,包括以下内容:

  • 交叉注意力

  • 完全掩码的行不再导致 NaN

  • 修改注意力分数:带有 FlexAttention 和 NJT 的 ALiBi

  • 打包投影

交叉注意力#

交叉注意力是一种注意力形式,其中查询和键/值张量来自不同的序列。

一个例子是 nn.TransformerDecoderLayer,其中查询来自解码器,键/值来自编码器。

上述 MultiheadAttention 层可以很好地推广到此情况,查询和键/值都使用嵌套张量。

query, _, _, q_len = gen_batch(N, E_q, E_k, E_v, device)
_, key, value, kv_len = gen_batch(N, E_q, E_k, E_v, device)

print(
    f"Total sequence length in nested query {q_len.sum().item()}, max sequence length {q_len.max().item()}"
)
print(
    f"Total sequence length in nested key/value {kv_len.sum().item()}, max sequence length {kv_len.max().item()}"
)
out = new_mha_layer(query, key, value, is_causal=False)
Total sequence length in nested query 10753, max sequence length 141
Total sequence length in nested key/value 10332, max sequence length 125

如上所述,我们可以将其与香草编译的 nn.MultiheadAttention 进行比较。

torch.manual_seed(6)
query, _, _, q_len = gen_batch(N, E_q, E_k, E_v, device)
_, key, value, kv_len = gen_batch(N, E_q, E_k, E_v, device)
padded_query, padded_key, padded_value = (
    t.to_padded_tensor(0.0) for t in (query, key, value)
)

key_padding_mask = torch.where(padded_key == 0.0, -math.inf, 0)[:, :, 0]

# warmup compile
warmup_nested_result = new_mha_layer(query, key, value, is_causal=False)
warmup_vanilla_result = vanilla_mha_layer(
    padded_query,
    padded_key,
    padded_value,
    key_padding_mask=key_padding_mask,
    need_weights=False,
    is_causal=False,
)

nested_result, nested_time, nested_peak_memory = benchmark(
    new_mha_layer, query, key, value, is_causal=False
)
(padded_result, _), padded_time, padded_peak_memory = benchmark(
    vanilla_mha_layer,
    padded_query,
    padded_key,
    padded_value,
    key_padding_mask=key_padding_mask,
    need_weights=False,
    is_causal=False,
)
padded_nested_result = nested_result.to_padded_tensor(0.0)
for i, entry_length in enumerate(q_len):
    # padding-specific step: remove output projection bias from padded entries for fair comparison
    padded_result[i, entry_length:, :] = 0.0

print(
    "Max difference between vanilla and nested result",
    (padded_result - padded_nested_result).abs().max().item(),
)
print(f"Nested speedup: {(padded_time/nested_time):.2f}")
print(
    f"Nested peak memory reduction {((padded_peak_memory - nested_peak_memory)/1e9):.2f} GB"
)
Max difference between vanilla and nested result 0.0
Nested speedup: 5.15
Nested peak memory reduction 1.23 GB

A100上的样本输出

Max difference between vanilla and nested result 0.0
Nested speedup: 4.01
Nested peak memory reduction 1.40 GB

完全遮罩的行不再导致 NaN#

nn.MultiheadAttentionscaled_dot_product_attention 长期以来一直存在一个问题:如果一行被完全遮罩,注意力层的输出将是 NaN。参见此问题。这是因为空集的 softmax 是未定义的。

感谢 此 PR,现在不再是这种情况。相反,scaled_dot_product_attention 中完全遮罩的行对应的输出将是 0。对于 nn.MHA 不采用“快速路径”的情况,这也将适用。

强烈建议使用带有 NJT 的自定义 MHA 层,而不是 nn.MultiheadAttention 中现有的“快速路径”,因为 NJT 适当建模不规则性的能力使得正确表达空序列成为可能。

FlexAttention + NJT#

NJT 也与 FlexAttention 模块结合使用。这是 MultiheadAttention 层的一种泛化,允许对注意力分数进行任意修改。下面的示例采用了 注意力操场 中实现 ALiBialibi_mod,并将其与嵌套输入张量一起使用。

from torch.nn.attention.flex_attention import flex_attention


def generate_alibi_bias(H: int):
    """Returns an alibi bias score_mod given the number of heads H
    Args:
        H: number of heads
    Returns:
        alibi_bias: alibi bias score_mod
    """

    def alibi_mod(score, b, h, q_idx, kv_idx):
        scale = torch.exp2(-((h + 1) * 8.0 / H))
        bias = (q_idx - kv_idx) * scale
        return score + bias

    return alibi_mod


query, key, value, _ = gen_batch(N, E_q, E_k, E_v, device)
n_heads, D = 8, E_q // 8
alibi_score_mod = generate_alibi_bias(n_heads)
query = query.unflatten(-1, [n_heads, D]).transpose(1, 2).detach().requires_grad_()
key = key.unflatten(-1, [n_heads, D]).transpose(1, 2).detach().requires_grad_()
value = value.unflatten(-1, [n_heads, D]).transpose(1, 2).detach().requires_grad_()
out_flex2 = flex_attention(query, key, value, score_mod=alibi_score_mod)

此外,还可以通过 create_nested_block_mask 函数将 FlexAttentionblock_mask 实用程序与 NJT 结合使用。这对于利用掩码的稀疏性来加速注意力计算非常有用。具体来说,该函数为 NJT 中所有可变长度序列组合成一个“堆叠序列”创建稀疏块掩码,同时正确掩盖序列间的注意力。在下面的示例中,我们展示了如何使用此实用程序创建因果块掩码。

from torch.nn.attention.flex_attention import create_nested_block_mask


def causal_mask(b, h, q_idx, kv_idx):
    return q_idx >= kv_idx


query, key, value, _ = gen_batch(N, E_q, E_k, E_v, device)
block_mask = create_nested_block_mask(causal_mask, 1, 1, query, _compile=True)
query = query.unflatten(-1, [n_heads, D]).transpose(1, 2).detach().requires_grad_()
key = key.unflatten(-1, [n_heads, D]).transpose(1, 2).detach().requires_grad_()
value = value.unflatten(-1, [n_heads, D]).transpose(1, 2).detach().requires_grad_()
out_flex = flex_attention(query, key, value, block_mask=block_mask)

打包投影#

打包投影是一种技术,它利用了当投影(矩阵乘法)的输入相同时(自注意力),我们可以将投影权重和偏差打包成单个张量的事实。当单个投影是内存密集型而非计算密集型时,它特别有用。我们将在此处演示两个示例:

  • MultiheadAttention 的输入投影

  • Transformer 层前馈网络中的 SwiGLU 激活

MultiheadAttention 的输入投影#

当进行自注意力时,querykeyvalue 是相同的张量。这些张量中的每一个都通过 Linear(E_q, E_total) 层进行投影。相反,我们可以将其打包成一个层,这也是我们在上面的 MultiheadAttention 层中所做的。

让我们比较一下打包投影与常规方法的性能

class InputProjection(nn.Module):
    def __init__(self, E_q, E_total, bias=False, device=None, dtype=None):
        factory_kwargs = {"device": device, "dtype": dtype}
        super().__init__()
        self.q_proj = nn.Linear(E_q, E_total, bias=bias, **factory_kwargs)
        self.k_proj = nn.Linear(E_q, E_total, bias=bias, **factory_kwargs)
        self.v_proj = nn.Linear(E_q, E_total, bias=bias, **factory_kwargs)

    def forward(self, x):
        return self.q_proj(x), self.k_proj(x), self.v_proj(x)


class PackedInputProjection(nn.Module):
    def __init__(self, E_q, E_total, bias=False, device=None, dtype=None):
        factory_kwargs = {"device": device, "dtype": dtype}
        super().__init__()
        self.packed_proj = nn.Linear(E_q, E_total * 3, bias=bias, **factory_kwargs)

    def forward(self, query):
        return torch.chunk(self.packed_proj(query), 3, dim=-1)


B, D, dtype = 256, 8192, torch.bfloat16

torch.set_float32_matmul_precision("high")
in_proj = torch.compile(InputProjection(D, D, device="cuda", dtype=torch.bfloat16))
packed_in_proj = torch.compile(
    PackedInputProjection(D, D, device="cuda", dtype=torch.bfloat16)
)

q, _, _, sequence_lengths = gen_batch(B, D, D, D, device="cuda", dtype=torch.bfloat16)

# warmup
in_proj(q)
packed_in_proj(q)

# benchmark
(q_out, k_out, v_out), time, _ = benchmark(in_proj, q)
(q_out, k_out, v_out), time_packed, _ = benchmark(packed_in_proj, q)
# On my A100 prints 1.05x speedup
print(
    f"InputProjection: {time:5f} s, PackedInputProjection: {time_packed:5f} s, speedup: {time/time_packed:.2f}x"
)
InputProjection: 0.030330 s, PackedInputProjection: 0.030233 s, speedup: 1.00x

Transformer 层中的 SwiGLU 前馈网络#

Swish-Gated Linear Unit (SwiGLU) 是一种非线性激活函数,在 Transformer 层(例如 Llama)的前馈网络中越来越受欢迎。具有 SwiGLU 激活的前馈网络定义为:

class SwiGLUFFN(nn.Module):
    def __init__(
        self,
        dim,
        hidden_dim,
        multiple_of,
        ffn_dim_multiplier=None,
        device=None,
        dtype=None,
    ):
        factory_kwargs = {"device": device, "dtype": dtype}
        super().__init__()
        hidden_dim = int(2 * hidden_dim / 3)
        # custom dim factor multiplier
        if ffn_dim_multiplier is not None:
            hidden_dim = int(ffn_dim_multiplier * hidden_dim)
        hidden_dim = multiple_of * ((hidden_dim + multiple_of - 1) // multiple_of)

        self.w1 = nn.Linear(dim, hidden_dim, bias=False, **factory_kwargs)
        self.w2 = nn.Linear(hidden_dim, dim, bias=False, **factory_kwargs)
        self.w3 = nn.Linear(dim, hidden_dim, bias=False, **factory_kwargs)

    def forward(self, x):
        return self.w2(F.silu(self.w1(x)) * self.w3(x))

另一种实现方式使用打包投影,如下所示:

class PackedSwiGLUFFN(nn.Module):
    def __init__(
        self,
        dim,
        hidden_dim,
        multiple_of,
        ffn_dim_multiplier=None,
        device=None,
        dtype=None,
    ):
        factory_kwargs = {"device": device, "dtype": dtype}
        super().__init__()
        hidden_dim = int(2 * hidden_dim / 3)
        # custom dim factor multiplier
        if ffn_dim_multiplier is not None:
            hidden_dim = int(ffn_dim_multiplier * hidden_dim)
        hidden_dim = multiple_of * ((hidden_dim + multiple_of - 1) // multiple_of)

        self.w13 = nn.Linear(dim, 2 * hidden_dim, bias=False, **factory_kwargs)
        self.w2 = nn.Linear(hidden_dim, dim, bias=False, **factory_kwargs)

    def forward(self, x):
        x1, x3 = torch.chunk(self.w13(x), 2, dim=-1)
        return self.w2(F.silu(x1) * x3)

我们可以像下面这样比较两种实现的性能。根据您的硬件,您可能会看到不同的结果。在 A100 上,D=128 时我看到了 1.12 倍的加速。

D = 128

swigluffn = torch.compile(SwiGLUFFN(D, D * 4, 256, device="cuda", dtype=torch.bfloat16))
packed_swigluffn = torch.compile(
    PackedSwiGLUFFN(D, D * 4, 256, device="cuda", dtype=torch.bfloat16)
)

q, _, _, sentence_lengths = gen_batch(D, D, D, D, device="cuda", dtype=torch.bfloat16)

# warmup
swigluffn(q)
packed_swigluffn(q)

# benchmark
_, time, _ = benchmark(swigluffn, q)
_, time_packed, _ = benchmark(packed_swigluffn, q)
# On my A100 prints 1.08x speedup
print(
    f"SwiGLUFFN: {time} s, PackedSwiGLUFFN: {time_packed} s, speedup: {time/time_packed:.2f}x"
)
SwiGLUFFN: 0.0009382299999742827 s, PackedSwiGLUFFN: 0.0008510179998211242 s, speedup: 1.10x

扩展示例#

我们打算更新本教程,以演示如何使用各种高性能构建块(如 KV 缓存、分组查询注意力等)的更多示例。此外,还有一些使用各种高性能构建块实现各种 Transformer 架构的优秀示例。一些示例包括:

结论#

在本教程中,我们介绍了 PyTorch 提供的用于编写 Transformer 层的低级构建块,并演示了如何组合它们。我们希望本教程能让读者了解 PyTorch 用户可以轻松实现灵活且高性能的 Transformer 层。

脚本总运行时间: (1 分钟 3.545 秒)