long-context

长上下文：扩展 Transformer 上下文窗口

何时使用本技能

在以下情况下使用长上下文技术：

• 处理长文档（32k、64k、128k+ token），使用 transformer 模型
• 扩展预训练模型的上下文窗口（LLaMA、Mistral 等）
• 实现高效的位置编码（RoPE、ALiBi）
• 训练具备长度外推能力的模型
• 部署能高效处理可变长度输入的模型
• 微调现有模型以支持更长上下文，且计算开销极小

关键技术：RoPE（旋转位置编码）、YaRN、ALiBi（带线性偏置的注意力）、位置插值

论文：RoFormer（arXiv 2104.09864）、YaRN（arXiv 2309.00071）、ALiBi（arXiv 2108.12409）、位置插值（arXiv 2306.15595）

安装

# HuggingFace Transformers (includes RoPE, YaRN support)
pip install transformers torch

# For custom implementations
pip install einops  # Tensor operations
pip install rotary-embedding-torch  # Standalone RoPE

# Optional: FlashAttention for efficiency
pip install flash-attn --no-build-isolation

快速开始

RoPE（旋转位置编码）

import torch
import torch.nn as nn

class RotaryEmbedding(nn.Module):
    """Rotary Position Embeddings (RoPE)."""

    def __init__(self, dim, max_seq_len=8192, base=10000):
        super().__init__()
        # Compute inverse frequencies
        inv_freq = 1.0 / (base ** (torch.arange(0, dim, 2).float() / dim))
        self.register_buffer("inv_freq", inv_freq)
        self.max_seq_len = max_seq_len

    def forward(self, seq_len, device):
        # Position indices
        t = torch.arange(seq_len, device=device).type_as(self.inv_freq)

        # Compute frequencies
        freqs = torch.outer(t, self.inv_freq)  # (seq_len, dim/2)

        # Compute sin and cos
        emb = torch.cat((freqs, freqs), dim=-1)  # (seq_len, dim)
        return emb.cos(), emb.sin()

def rotate_half(x):
    """Rotate half the hidden dimensions."""
    x1, x2 = x.chunk(2, dim=-1)
    return torch.cat((-x2, x1), dim=-1)

def apply_rotary_pos_emb(q, k, cos, sin):
    """Apply rotary embeddings to queries and keys."""
    # q, k shape: (batch, heads, seq_len, dim)
    q_embed = (q * cos) + (rotate_half(q) * sin)
    k_embed = (k * cos) + (rotate_half(k) * sin)
    return q_embed, k_embed

# Usage
rope = RotaryEmbedding(dim=64, max_seq_len=8192)
cos, sin = rope(seq_len=2048, device='cuda')

# In attention layer
q_rotated, k_rotated = apply_rotary_pos_emb(query, key, cos, sin)

ALiBi（带线性偏置的注意力）

def get_alibi_slopes(num_heads):
    """Get ALiBi slope values for each attention head."""
    def get_slopes_power_of_2(n):
        start = 2 ** (-(2 ** -(math.log2(n) - 3)))
        ratio = start
        return [start * (ratio ** i) for i in range(n)]

    if math.log2(num_heads).is_integer():
        return get_slopes_power_of_2(num_heads)
    else:
        # Closest power of 2
        closest_power = 2 ** math.floor(math.log2(num_heads))
        slopes = get_slopes_power_of_2(closest_power)
        # Add extra slopes
        extra = get_slopes_power_of_2(2 * closest_power)
        slopes.extend(extra[0::2][:num_heads - closest_power])
        return slopes

def create_alibi_bias(seq_len, num_heads):
    """Create ALiBi attention bias."""
    # Distance matrix
    context_position = torch.arange(seq_len)
    memory_position = torch.arange(seq_len)
    relative_position = memory_position[None, :] - context_position[:, None]

    # Get slopes
    slopes = torch.tensor(get_alibi_slopes(num_heads))

    # Apply slopes to distances
    alibi = slopes[:, None, None] * relative_position[None, :, :]
    return alibi  # (num_heads, seq_len, seq_len)

# Usage in attention
num_heads = 8
seq_len = 2048
alibi_bias = create_alibi_bias(seq_len, num_heads).to('cuda')

# Add bias to attention scores
# attn_scores shape: (batch, num_heads, seq_len, seq_len)
attn_scores = attn_scores + alibi_bias
attn_weights = torch.softmax(attn_scores, dim=-1)

面向 LLaMA 的位置插值

from transformers import LlamaForCausalLM, LlamaTokenizer

# Original context: 2048 tokens
model = LlamaForCausalLM.from_pretrained("meta-llama/Llama-2-7b-hf")

# Extend to 32k with position interpolation
# Modify RoPE base frequency
model.config.rope_scaling = {
    "type": "linear",
    "factor": 16.0  # 2048 * 16 = 32768
}

# Or use dynamic scaling
model.config.rope_scaling = {
    "type": "dynamic",
    "factor": 16.0
}

# Fine-tune with long documents (minimal steps needed)
# Position interpolation works out-of-the-box after this config change

核心概念

1. RoPE（旋转位置编码）

工作原理：

• 通过旋转矩阵编码绝对位置
• 在注意力中提供相对位置依赖
• 支持长度外推

数学表达式：

q_m = (W_q * x_m) * e^(imθ)
k_n = (W_k * x_n) * e^(inθ)

where θ_j = base^(-2j/d) for j ∈ [0, d/2)

优势：

• token 间依赖随距离衰减
• 兼容线性注意力
• 比绝对位置编码具有更好的外推能力

2. YaRN（Yet another RoPE extensioN）

关键创新：

• NTK 感知插值（神经正切核）
• 注意力温度缩放
• 高效的上下文扩展（相比基线减少 10× token）

参数：

# YaRN configuration
yarn_config = {
    "scale": 16,                    # Extension factor
    "original_max_position": 2048,  # Base context
    "extrapolation_factor": 1.0,    # NTK parameter
    "attn_factor": 1.0,             # Attention scaling
    "beta_fast": 32,                # High-frequency scale
    "beta_slow": 1,                 # Low-frequency scale
}

性能：

• 将 LLaMA 扩展至 128k token
• 训练步数比基线减少 2.5×
• 最先进的上下文窗口扩展

3. ALiBi（带线性偏置的注意力）

核心思想：

• 不向 token 添加位置嵌入
• 直接对注意力分数施加距离惩罚
• 偏置与 key-query 距离成正比

公式：

attention_bias[i, j] = -m * |i - j|

where m = slope for each attention head

优势：

• 训练速度比正弦嵌入快 11%
• 内存占用减少 11%
• 强大的长度外推能力（训练 1k，测试 2k+）
• 对近期 token 的归纳偏置

4. 位置插值

技术：

• 线性缩小位置索引
• 在训练范围内插值（而非超出范围外推）
• 只需极少的微调

公式：

# Original: position indices [0, 1, 2, ..., L]
# Extended: position indices [0, 0.5, 1.0, ..., L/2]
# (for 2× extension)

scaled_position[i] = i / extension_factor

结果：

• LLaMA 7B-65B 扩展至 32k token
• 1000 步微调即可
• 稳定性比外推提升 600×

方法对比

方法	最大上下文	所需训练	内存	外推能力	最适用于
RoPE	8k-32k	完整预训练	中等	良好	新模型
YaRN	32k-128k	极少（高效 10×）	中等	优秀	扩展现有模型
ALiBi	无限制	完整预训练	低（-11%）	优秀	从零训练
位置插值	32k+	极少（1k 步）	中等	差（设计使然）	快速扩展

实现模式

HuggingFace Transformers 集成

from transformers import AutoModelForCausalLM, AutoConfig

# RoPE with YaRN scaling
config = AutoConfig.from_pretrained("mistralai/Mistral-7B-v0.1")
config.rope_scaling = {
    "type": "yarn",
    "factor": 8.0,
    "original_max_position_embeddings": 8192,
    "attention_factor": 1.0
}

model = AutoModelForCausalLM.from_config(config)

# Position interpolation (simpler)
config.rope_scaling = {
    "type": "linear",
    "factor": 4.0
}

# Dynamic scaling (adjusts based on input length)
config.rope_scaling = {
    "type": "dynamic",
    "factor": 8.0
}

自定义 RoPE 实现

class LongContextAttention(nn.Module):
    """Multi-head attention with RoPE."""

    def __init__(self, hidden_size, num_heads, max_seq_len=32768):
        super().__init__()
        self.num_heads = num_heads
        self.head_dim = hidden_size // num_heads

        # Q, K, V projections
        self.q_proj = nn.Linear(hidden_size, hidden_size)
        self.k_proj = nn.Linear(hidden_size, hidden_size)
        self.v_proj = nn.Linear(hidden_size, hidden_size)
        self.o_proj = nn.Linear(hidden_size, hidden_size)

        # RoPE
        self.rotary_emb = RotaryEmbedding(
            dim=self.head_dim,
            max_seq_len=max_seq_len
        )

    def forward(self, hidden_states):
        batch_size, seq_len, _ = hidden_states.shape

        # Project to Q, K, V
        q = self.q_proj(hidden_states)
        k = self.k_proj(hidden_states)
        v = self.v_proj(hidden_states)

        # Reshape for multi-head
        q = q.view(batch_size, seq_len, self.num_heads, self.head_dim).transpose(1, 2)
        k = k.view(batch_size, seq_len, self.num_heads, self.head_dim).transpose(1, 2)
        v = v.view(batch_size, seq_len, self.num_heads, self.head_dim).transpose(1, 2)

        # Apply RoPE
        cos, sin = self.rotary_emb(seq_len, device=hidden_states.device)
        q, k = apply_rotary_pos_emb(q, k, cos, sin)

        # Standard attention
        attn_output = F.scaled_dot_product_attention(q, k, v)

        # Reshape and project
        attn_output = attn_output.transpose(1, 2).contiguous()
        attn_output = attn_output.view(batch_size, seq_len, -1)
        output = self.o_proj(attn_output)

        return output

长上下文微调

最小化微调（位置插值）

from transformers import Trainer, TrainingArguments

# Extend model config
model.config.max_position_embeddings = 32768
model.config.rope_scaling = {"type": "linear", "factor": 16.0}

# Training args (minimal steps needed)
training_args = TrainingArguments(
    output_dir="./llama-32k",
    num_train_epochs=1,
    max_steps=1000,           # Only 1000 steps!
    per_device_train_batch_size=1,
    gradient_accumulation_steps=16,
    learning_rate=2e-5,
    warmup_steps=100,
    logging_steps=10,
    save_steps=500,
)

# Train on long documents
trainer = Trainer(
    model=model,
    args=training_args,
    train_dataset=long_document_dataset,  # 32k token sequences
)

trainer.train()

YaRN 微调

# Clone YaRN implementation
git clone https://github.com/jquesnelle/yarn
cd yarn

# Fine-tune LLaMA with YaRN
python scripts/train.py \
    --model meta-llama/Llama-2-7b-hf \
    --scale 16 \
    --rope_theta 10000 \
    --max_length 32768 \
    --batch_size 1 \
    --gradient_accumulation 16 \
    --steps 400 \
    --learning_rate 2e-5

最佳实践

1. 选择合适的方法

# For NEW models (training from scratch)
use_method = "ALiBi"  # Best extrapolation, lowest memory

# For EXTENDING existing RoPE models
use_method = "YaRN"  # Most efficient extension (10× less data)

# For QUICK extension with minimal compute
use_method = "Position Interpolation"  # 1000 steps

# For MODERATE extension with good efficiency
use_method = "Linear RoPE Scaling"  # Built-in, simple

2. 缩放因子选择

# Conservative (safer, better quality)
scaling_factor = 2.0  # 8k → 16k

# Moderate (good balance)
scaling_factor = 4.0  # 8k → 32k

# Aggressive (requires more fine-tuning)
scaling_factor = 8.0  # 8k → 64k
scaling_factor = 16.0  # 8k → 128k

# Rule: Larger factors need more fine-tuning steps
steps_needed = 100 * scaling_factor  # Rough estimate

3. 微调数据

# ✅ Good: Long documents matching target length
train_data = [
    {"text": long_doc_32k_tokens},  # Full 32k
    {"text": long_doc_24k_tokens},  # Varied lengths
    {"text": long_doc_16k_tokens},
]

# ❌ Bad: Short documents (won't learn long context)
train_data = [
    {"text": short_doc_2k_tokens},
]

# Use datasets like:
# - PG-19 (books, long texts)
# - arXiv papers
# - Long-form conversations
# - GitHub repositories (concatenated files)

4. 避免常见陷阱

# ❌ Bad: Applying position interpolation without fine-tuning
model.config.rope_scaling = {"type": "linear", "factor": 16.0}
# Model will perform poorly without fine-tuning!

# ✅ Good: Fine-tune after scaling
model.config.rope_scaling = {"type": "linear", "factor": 16.0}
fine_tune(model, long_documents, steps=1000)

# ❌ Bad: Too aggressive scaling without data
scale_to_1M_tokens()  # Won't work without massive fine-tuning

# ✅ Good: Incremental scaling
# 8k → 16k → 32k → 64k (fine-tune at each step)

生产部署

长上下文推理

from transformers import AutoModelForCausalLM, AutoTokenizer

# Load long-context model
model = AutoModelForCausalLM.from_pretrained(
    "togethercomputer/LLaMA-2-7B-32K",  # 32k context
    torch_dtype=torch.float16,
    device_map="auto"
)
tokenizer = AutoTokenizer.from_pretrained("togethercomputer/LLaMA-2-7B-32K")

# Process long document
long_text = "..." * 30000  # 30k tokens
inputs = tokenizer(long_text, return_tensors="pt", truncation=False).to('cuda')

# Generate
outputs = model.generate(
    **inputs,
    max_new_tokens=512,
    temperature=0.7,
)

response = tokenizer.decode(outputs[0], skip_special_tokens=True)

内存优化

# Use gradient checkpointing for fine-tuning
model.gradient_checkpointing_enable()

# Use Flash Attention 2
model = AutoModelForCausalLM.from_pretrained(
    "meta-llama/Llama-2-7b-hf",
    attn_implementation="flash_attention_2",  # 2-3× faster
    torch_dtype=torch.float16
)

# Use paged attention (vLLM)
from vllm import LLM

llm = LLM(
    model="togethercomputer/LLaMA-2-7B-32K",
    max_model_len=32768,  # 32k context
    gpu_memory_utilization=0.9
)

资源

• RoPE 论文：https://arxiv.org/abs/2104.09864 （RoFormer）
• YaRN 论文：https://arxiv.org/abs/2309.00071
• ALiBi 论文：https://arxiv.org/abs/2108.12409 （Train Short, Test Long）
• 位置插值：https://arxiv.org/abs/2306.15595
• HuggingFace RoPE 工具：https://github.com/huggingface/transformers/blob/main/src/transformers/modeling_rope_utils.py
• YaRN 实现：https://github.com/jquesnelle/yarn
• Together AI 博客：https://www.together.ai/blog/llama-2-7b-32k

另请参阅

• references/rope.md - RoPE 实现与理论的详细说明
• references/extension_methods.md - YaRN、ALiBi、位置插值的对比
• references/fine_tuning.md - 上下文扩展的完整微调指南