2nd/10_Wiki/Topics/AI_and_ML/Sequence-to-Sequence-Models.md

---
id: wiki-2026-0508-sequence-to-sequence-models
title: Sequence to Sequence Models
category: 10_Wiki/Topics
status: verified
canonical_id: self
aliases: [seq2seq, Encoder-Decoder, Sequence Modeling, Sequence-to-Sequence]
duplicate_of: none
source_trust_level: A
confidence_score: 0.9
verification_status: applied
tags: [architecture, nlp, transformer, encoder-decoder]
raw_sources: []
last_reinforced: 2026-05-10
github_commit: pending
tech_stack:
  language: python
  framework: PyTorch / Transformers
---

# Sequence to Sequence Models

## 매 한 줄
> **"매 input sequence → output sequence — 매 길이 다른 변환"**. 매 Sutskever (2014) RNN encoder-decoder → Bahdanau (2015) attention → Vaswani (2017) Transformer 의 진화. 매 2026: 거의 모든 generative LLM (GPT, Claude, Gemini) 이 매 decoder-only seq2seq, 매 T5/BART 같은 encoder-decoder 는 specific task (번역, summarization fine-tune) 에 잔존.

## 매 핵심

### 매 Architecture family
- **RNN encoder-decoder** (2014): 매 historical, vanishing gradient, no attention.
- **Attention seq2seq** (2015): 매 alignment 학습 — 번역 quality 점프.
- **Transformer encoder-decoder** (2017): 매 self-attention, parallelizable. T5, BART, mT5.
- **Decoder-only** (2018+): GPT family. 매 LLM 의 dominant pattern.
- **Encoder-only** (BERT): classification/embedding, generation 아님.

### 매 핵심 컴포넌트
- Tokenizer (BPE, SentencePiece, tiktoken).
- Embedding + positional encoding (RoPE, ALiBi 2026 표준).
- Self-attention / cross-attention.
- Teacher forcing for training, autoregressive decoding for inference.

### 매 Decoding 전략
- Greedy / Beam search — 매 deterministic task.
- Sampling (temperature, top-p, top-k, min-p) — 매 creative.
- Speculative / Medusa — 매 inference 가속.
- Constrained / structured (JSON schema) — 매 tool use.

### 매 응용
1. Machine translation (NLLB, M2M-100).
2. Summarization (BART, Pegasus).
3. Code generation (Claude Code, Copilot).
4. Speech (Whisper encoder + decoder).
5. Image captioning, VQA (multimodal seq2seq).

## 💻 패턴

### Tiny Transformer encoder-decoder
```python
import torch.nn as nn
class Seq2Seq(nn.Module):
    def __init__(self, vocab, d=256, nhead=4, nl=4):
        super().__init__()
        self.emb_s = nn.Embedding(vocab, d)
        self.emb_t = nn.Embedding(vocab, d)
        self.tx = nn.Transformer(d, nhead, nl, nl, batch_first=True)
        self.out = nn.Linear(d, vocab)
    def forward(self, src, tgt):
        return self.out(self.tx(self.emb_s(src), self.emb_t(tgt)))
```

### HF Transformers (T5)
```python
from transformers import T5ForConditionalGeneration, T5Tokenizer
tok = T5Tokenizer.from_pretrained("t5-base")
m = T5ForConditionalGeneration.from_pretrained("t5-base")
inp = tok("translate English to German: Hello world", return_tensors="pt").input_ids
print(tok.decode(m.generate(inp)[0], skip_special_tokens=True))
```

### Decoder-only generation (Claude API)
```python
import anthropic
c = anthropic.Anthropic()
msg = c.messages.create(
    model="claude-opus-4-7",
    max_tokens=1024,
    messages=[{"role": "user", "content": "Summarize: ..."}],
)
print(msg.content[0].text)
```

### Beam search decode
```python
out = model.generate(input_ids, num_beams=4, length_penalty=0.6,
                     no_repeat_ngram_size=3, max_new_tokens=128)
```

### Streaming
```python
with c.messages.stream(model="claude-opus-4-7", max_tokens=512,
                       messages=msgs) as s:
    for text in s.text_stream:
        print(text, end="", flush=True)
```

### KV cache reuse
```python
out = model(**inputs, use_cache=True, past_key_values=pkv)
pkv = out.past_key_values  # 매 next step 에 재사용
```

## 매 결정 기준
| 상황 | Approach |
|---|---|
| General LLM task | decoder-only (Claude, GPT) |
| Specific translation/summarization fine-tune | T5/BART encoder-decoder |
| Embedding / classification | encoder-only (BERT family) |
| Speech-to-text | Whisper-style enc-dec |
| Long sequences, low cost | Mamba / Hybrid seq2seq |

**기본값**: decoder-only LLM via API.

## 🔗 Graph
- 부모: [[Deep Learning]] · [[NLP]]
- 변형: [[Transformer]] · [[Selective State Space Models (Mamba)]] · [[Encoder-Decoder]]
- 응용: [[Summarization]] · [[Code-Generation]]
- Adjacent: [[Attention Mechanism]] · [[Tokenization]]

## 🤖 LLM 활용
**언제**: input → output 변환 task 정의 가능. 매 API call 로 충분.
**언제 X**: pure classification — encoder + head 가 매 더 cheap.

## ❌ 안티패턴
- **Greedy for creative**: repetition. 매 sampling 사용.
- **No cache**: O(L²) inference. 매 KV cache 필수.
- **Train from scratch**: 매 거의 항상 잘못된 선택. Fine-tune 또는 prompt.

## 🧪 검증 / 중복
- Verified (Sutskever 2014, Vaswani 2017, HF Transformers docs).
- 신뢰도 A.

## 🕓 Changelog
| 날짜 | 변경 |
|---|---|
| 2026-05-08 | Phase 1 |
| 2026-05-10 | Manual cleanup — full seq2seq family 2026 |