long-context

Long Context: Extending Transformer Context Windows

When to Use This Skill

Use Long Context techniques when you need to:

• Process long documents (32k, 64k, 128k+ tokens) with transformer models
• Extend context windows of pre-trained models (LLaMA, Mistral, etc.)
• Implement efficient positional encodings (RoPE, ALiBi)
• Train models with length extrapolation capabilities
• Deploy models that handle variable-length inputs efficiently
• Fine-tune existing models for longer contexts with minimal compute

Key Techniques: RoPE (Rotary Position Embeddings), YaRN, ALiBi (Attention with Linear Biases), Position Interpolation

Papers: RoFormer (arXiv 2104.09864), YaRN (arXiv 2309.00071), ALiBi (arXiv 2108.12409), Position Interpolation (arXiv 2306.15595)

Installation

# 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

Quick Start

RoPE (Rotary Position Embeddings)

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)