speculative-decoding

Accélérer l'inférence des LLM en utilisant le décodage spéculatif, les têtes multiples Medusa et les techniques de décodage par anticipation. À utiliser lors de l'optimisation de la vitesse d'inférence (1,5 à 3,6 fois…

npx skills add https://github.com/firecrawl/ai-research-skills --skill speculative-decoding

Speculative Decoding: Accelerating LLM Inference

When to Use This Skill

Use Speculative Decoding when you need to:

  • Speed up inference by 1.5-3.6× without quality loss
  • Reduce latency for real-time applications (chatbots, code generation)
  • Optimize throughput for high-volume serving
  • Deploy efficiently on limited hardware
  • Generate faster without changing model architecture

Key Techniques: Draft model speculative decoding, Medusa (multiple heads), Lookahead Decoding (Jacobi iteration)

Papers: Medusa (arXiv 2401.10774), Lookahead Decoding (ICML 2024), Speculative Decoding Survey (ACL 2024)

Installation

# Standard speculative decoding (transformers)
pip install transformers accelerate

# Medusa (multiple decoding heads)
git clone https://github.com/FasterDecoding/Medusa
cd Medusa
pip install -e .

# Lookahead Decoding
git clone https://github.com/hao-ai-lab/LookaheadDecoding
cd LookaheadDecoding
pip install -e .

# Optional: vLLM with speculative decoding
pip install vllm

Quick Start

Basic Speculative Decoding (Draft Model)

from transformers import AutoModelForCausalLM, AutoTokenizer

# Load target model (large, slow)
target_model = AutoModelForCausalLM.from_pretrained(
    "meta-llama/Llama-2-70b-hf",
    device_map="auto",
    torch_dtype=torch.float16
)

# Load draft model (small, fast)
draft_model = AutoModelForCausalLM.from_pretrained(
    "meta-llama/Llama-2-7b-hf",
    device_map="auto",
    torch_dtype=torch.float16
)

tokenizer = AutoTokenizer.from_pretrained("meta-llama/Llama-2-70b-hf")

# Generate with speculative decoding
prompt = "Explain quantum computing in simple terms:"
inputs = tokenizer(prompt, return_tensors="pt").to("cuda")

# Transformers 4.36+ supports assisted generation
outputs = target_model.generate(
    **inputs,
    assistant_model=draft_model,  # Enable speculative decoding
    max_new_tokens=256,
    do_sample=True,
    temperature=0.7,
)

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

Medusa (Multiple Decoding Heads)

from medusa.model.medusa_model import MedusaModel

# Load Medusa-enhanced model
model = MedusaModel.from_pretrained(
    "FasterDecoding/medusa-vicuna-7b-v1.3",  # Pre-trained with Medusa heads
    torch_dtype=torch.float16,
    device_map="auto"
)

tokenizer = AutoTokenizer.from_pretrained("FasterDecoding/medusa-vicuna-7b-v1.3")

# Generate with Medusa (2-3× speedup)
prompt = "Write a Python function to calculate fibonacci numbers:"
inputs = tokenizer(prompt, return_tensors="pt").to("cuda")

outputs = model.medusa_generate(
    **inputs,
    max_new_tokens=256,
    temperature=0.7,
    posterior_threshold=0.09,  # Acceptance threshold
    posterior_alpha=0.3,       # Tree construction parameter
)

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

Lookahead Decoding (Jacobi Iteration)

from lookahead.lookahead_decoding import LookaheadDecoding

# Load model
model = AutoModelForCausalLM.from_pretrained(
    "meta-llama/Llama-2-7b-hf",
    torch_dtype=torch.float16,
    device_map="auto"
)
tokenizer = AutoTokenizer.from_pretrained("meta-llama/Llama-2-7b-hf")

# Initialize lookahead decoding
lookahead = LookaheadDecoding(
    model=model,
    tokenizer=tokenizer,
    window_size=15,    # Lookahead window (W)
    ngram_size=5,      # N-gram size (N)
    guess_size=5       # Number of parallel guesses
)

# Generate (1.5-2.3× speedup)
prompt = "Implement quicksort in Python:"
output = lookahead.generate(prompt, max_new_tokens=256)
print(output)

Core Concepts

1. Speculative Decoding (Draft Model)

Idea: Use small draft model to generate candidates, large target model to verify in parallel.

Algorithm:

  1. Draft model generates K tokens speculatively
  2. Target model evaluates all K tokens in parallel (single forward pass)
  3. Accept tokens where draft and target agree
  4. Reject first disagreement, continue from there
def speculative_decode(target_model, draft_model, prompt, K=4):
    """Speculative decoding algorithm."""
    # 1. Generate K draft tokens
    draft_tokens = draft_model.generate(prompt, max_new_tokens=K)

    # 2. Target model evaluates all K tokens in one forward pass
    target_logits = target_model(draft_tokens)  # Parallel!

    # 3. Accept/reject based on probability match
    accepted = []
    for i in range(K):
        p_draft = softmax(draft_model.logits[i])
        p_target = softmax(target_logits[i])

        # Acceptance probability
        if random.random() < min(1, p_target[draft_tokens[i]] / p_draft[draft_tokens[i]]):
            accepted.append(draft_tokens[i])
        else:
            break  # Reject, resample from target

    return accepted

Performance:

  • Speedup: 1.5-2× with good draft model
  • Zero quality loss (mathematically equivalent to target model)
  • Best when draft model is 5-10× smaller than target

2. Medusa (Multiple Decoding Heads)

Source: arXiv 2401.10774 (2024)

Innovation: Add multiple prediction heads to existing model, predict future tokens without separate draft model.

Architecture:

Input → Base LLM (frozen) → Hidden State
                                ├→ Head 1 (predicts token t+1)
                                ├→ Head 2 (predicts token t+2)
                                ├→ Head 3 (predicts token t+3)
                                └→ Head 4 (predicts token t+4)

Training:

  • Medusa-1: Freeze base LLM, train only heads
    • 2.2× speedup, lossless
  • Medusa-2: Fine-tune base LLM + heads together
    • 2.3-3.6× speedup, better quality

Tree-based Attention:

# Medusa constructs tree of candidates
# Example: Predict 2 steps ahead with top-2 per step

#         Root
#        /    \
#      T1a    T1b  (Step 1: 2 candidates)
#     /  \    / \
#  T2a  T2b T2c T2d  (Step 2: 4 candidates total)

# Single forward pass evaluates entire tree!

Advantages:

  • No separate draft model needed
  • Minimal training (only heads)
  • Compatible with any LLM

3. Lookahead Decoding (Jacobi Iteration)

Source: ICML 2024

Core idea: Reformulate autoregressive decoding as solving system of equations, solve in parallel using Jacobi iteration.

Mathematical formulation:

Traditional:  y_t = f(x, y_1, ..., y_{t-1})  (sequential)
Jacobi:       y_t^{(k+1)} = f(x, y_1^{(k)}, ..., y_{t-1}^{(k)})  (parallel)

Two branches:

  1. Lookahead Branch: Generate n-grams in parallel

    • Window size W: How many steps to look ahead
    • N-gram size N: How many past tokens to use
  2. Verification Branch: Verify promising n-grams

    • Match n-grams with generated tokens
    • Accept if first token matches
class LookaheadDecoding:
    def __init__(self, model, window_size=15, ngram_size=5):
        self.model = model
        self.W = window_size  # Lookahead window
        self.N = ngram_size   # N-gram size

    def generate_step(self, tokens):
        # Lookahead branch: Generate W × N candidates
        candidates = {}
        for w in range(1, self.W + 1):
            for n in range(1, self.N + 1):
                # Generate n-gram starting at position w
                ngram = self.generate_ngram(tokens, start=w, length=n)
                candidates[(w, n)] = ngram

        # Verification branch: Find matching n-grams
        verified = []
        for ngram in candidates.values():
            if ngram[0] == tokens[-1]:  # First token matches last input
                if self.verify(tokens, ngram):
                    verified.append(ngram)

        # Accept longest verified n-gram
        return max(verified, key=len) if verified else [self.model.generate_next(tokens)]

Performance:

  • Speedup: 1.5-2.3× (up to 3.6× for code generation)
  • No draft model or training needed
  • Works out-of-the-box with any model

Method Comparison

MethodSpeedupTraining NeededDraft ModelQuality Loss
Draft Model Speculative1.5-2×NoYes (external)None
Medusa2-3.6×Minimal (heads only)No (built-in heads)None
Lookahead1.5-2.3×NoneNoNone
Naive Batching1.2-1.5×NoNoNone

Advanced Patterns

Training Medusa Heads

from medusa.model.medusa_model import MedusaModel
from medusa.model.kv_cache import initialize_past_key_values
import torch.nn as nn

# 1. Load base model
base_model = AutoModelForCausalLM.from_pretrained(
    "lmsys/vicuna-7b-v1.3",
    torch_dtype=torch.float16
)

# 2. Add Medusa heads
num_heads = 4
medusa_heads = nn.ModuleList([
    nn.Linear(base_model.config.hidden_size, base_model.config.vocab_size, bias=False)
    for _ in range(num_heads)
])

# 3. Training loop (freeze base model for Medusa-1)
for param in base_model.parameters():
    param.requires_grad = False  # Freeze base

optimizer = torch.optim.Adam(medusa_heads.parameters(), lr=1e-3)

for batch in dataloader:
    # Forward pass
    hidden_states = base_model(**batch, output_hidden_states=True).hidden_states[-1]

    # Predict future tokens with each head
    loss = 0
    for i, head in enumerate(medusa_heads):
        logits = head(hidden_states)
        # Target: tokens shifted by (i+1) positions
        target = batch['input_ids'][:, i+1:]
        loss += F.cross_entropy(logits[:, :-i-1], target)

    # Backward
    optimizer.zero_grad()
    loss.backward()
    optimizer.step()

Hybrid: Speculative + Medusa

# Use Medusa as draft model for speculative decoding
draft_medusa = MedusaModel.from_pretrained("medusa-vicuna-7b")
target_model = AutoModelForCausalLM.from_pretrained("vicuna-33b")

# Draft generates multiple candidates with Medusa
draft_tokens = draft_medusa.medusa_generate(prompt, max_new_tokens=5)

# Target verifies in single forward pass
outputs = target_model.generate(
    prompt,
    assistant_model=draft_medusa,  # Use Medusa as draft
    max_new_tokens=256
)

# Combines benefits: Medusa speed + large model quality

Optimal Draft Model Selection

def select_draft_model(target_model_size, target):
    """Select optimal draft model for speculative decoding."""
    # Rule: Draft should be 5-10× smaller
    if target_model_size == "70B":
        return "7B"  # 10× smaller
    elif target_model_size == "33B":
        return "7B"  # 5× smaller
    elif target_model_size == "13B":
        return "1B"  # 13× smaller
    else:
        return None  # Target too small, use Medusa/Lookahead instead

# Example
draft = select_draft_model("70B", target_model)
# Returns "7B" → Use Llama-2-7b as draft for Llama-2-70b

Best Practices

1. Choose the Right Method

# New deployment → Medusa (best overall speedup, no draft model)
if deploying_new_model:
    use_method = "Medusa"

# Existing deployment with small model available → Draft speculative
elif have_small_version_of_model:
    use_method = "Draft Model Speculative"

# Want zero training/setup → Lookahead
elif want_plug_and_play:
    use_method = "Lookahead Decoding"

2. Hyperparameter Tuning

Draft Model Speculative:

# K = number of speculative tokens
K = 4  # Good default
K = 2  # Conservative (higher acceptance)
K = 8  # Aggressive (lower acceptance, but more when accepted)

# Rule: Larger K → more speedup IF draft model is good

Medusa:

# Posterior threshold (acceptance confidence)
posterior_threshold = 0.09  # Standard (from paper)
posterior_threshold = 0.05  # More conservative (slower, higher quality)
posterior_threshold = 0.15  # More aggressive (faster, may degrade quality)

# Tree depth (how many steps ahead)
medusa_choices = [[0], [0, 0], [0, 1], [0, 0, 0]]  # Depth 3 (standard)

Lookahead:

# Window size W (lookahead distance)
# N-gram size N (context for generation)

# 7B model (more resources)
W, N = 15, 5

# 13B model (moderate)
W, N = 10, 5

# 33B+ model (limited resources)
W, N = 7, 5

3. Production Deployment

# vLLM with speculative decoding
from vllm import LLM, SamplingParams

# Initialize with draft model
llm = LLM(
    model="meta-llama/Llama-2-70b-hf",
    speculative_model="meta-llama/Llama-2-7b-hf",  # Draft model
    num_speculative_tokens=5,
    use_v2_block_manager=True,
)

# Generate
prompts = ["Tell me about AI:", "Explain quantum physics:"]
sampling_params = SamplingParams(temperature=0.7, max_tokens=256)

outputs = llm.generate(prompts, sampling_params)
for output in outputs:
    print(output.outputs[0].text)

Resources

See Also

  • references/draft_model.md - Draft model selection and training
  • references/medusa.md - Medusa architecture and training
  • references/lookahead.md - Lookahead decoding implementation details

Plus de skills de firecrawl

firecrawl-research-index
firecrawl
Trouvez les articles qui répondent à une requête de recherche avec Firecrawl Research, en utilisant la recherche sémantique, l'expansion sémantique et structurelle, et la vérification dans le corps du texte. Utilisez toujours cette compétence pour toute tâche de recherche d'articles ou de récupération de documents — qu'il s'agisse de consulter un seul article ou un ensemble complet de plusieurs articles.
data-analysisresearchweb-scraping
oracle
firecrawl
Meilleures pratiques pour utiliser l'interface en ligne de commande oracle (invite + regroupement de fichiers, moteurs, sessions et modèles de pièces jointes).
pinecone
firecrawl
Base de données vectorielle gérée pour les applications d'IA en production. Entièrement gérée, mise à l'échelle automatique, avec recherche hybride (dense + sparse), filtrage par métadonnées et espaces de noms.…
wpds
firecrawl
À utiliser lors de la création d'interfaces utilisateur exploitant le WordPress Design System (WPDS) et ses composants, tokens, motifs, etc.
audiocraft-audio-generation
firecrawl
Bibliothèque PyTorch pour la génération audio incluant texte-vers-musique (MusicGen) et texte-vers-son (AudioGen). À utiliser lorsque vous avez besoin de générer de la musique à partir de texte…
skypilot-multi-cloud-orchestration
firecrawl
Orchestration multi-cloud pour les charges de travail ML avec optimisation automatique des coûts. À utiliser lorsque vous devez exécuter des tâches d’entraînement ou par lots sur plusieurs clouds, tirer parti de…
firecrawl-seo-audit
firecrawl
Auditer le SEO d'un site web avec Firecrawl. À utiliser lorsque l'utilisateur demande un audit SEO, une revue des métadonnées et des titres, une analyse du sitemap/de la structure du site, des opportunités de mots-clés, une comparaison des SERP concurrents, ou des recommandations priorisées d'optimisation pour la recherche.
data-analysisresearchweb-scraping
gh-issues
firecrawl
Récupérer les issues GitHub, générer des sous-agents pour implémenter des correctifs et ouvrir des PR, puis surveiller et traiter les commentaires de révision des PR. Utilisation : /gh-issues [owner/repo] [--label…]