LLM Inference Principles

What is LLM Inference

LLM (Large Language Model) inference refers to the process where, given an input text (Prompt), the model predicts the next token through self-attention mechanisms and probability distribution, generating output text.

Inference Flow

┌────────────────────────────────────────────────────────┐
│                    LLM Inference Flow                  │
├────────────────────────────────────────────────────────┤
│                                                        │
│   Input Text ──→ Tokenize ──→ Embed + Pos Enc ──→ Transformer │
│       │                                              │
│       │                                              │
│       ◀──────────── Sampling ◀─────────────── Output Token │
│                                                        │
│   Output Text ◀── Detokenize ◀── Vocabulary Proj ◀── Transformer │
│                                                        │
└────────────────────────────────────────────────────────┘

Stage Descriptions

Stage	Input	Output	Description
Tokenize	Text	token ID sequence	Split text into model-processable units
Embed	token ID	Vector	Map token to high-dimensional vector
Positional Encoding	Vector	Position info added	Let model perceive token position relationships
Transformer	Vector sequence	Hidden states	Core computation, multi-layer self-attention
Vocabulary Projection	Hidden states	Vocabulary probability	Predict next token
Sampling	Probability distribution	token ID	Select output token based on strategy
Detokenize	token ID	Text	Convert token sequence back to text

Core Technical Components

1. Tokenizer

Tokenizer converts text into token sequences that the model can process.

Chinese Tokenization Example:

Input: "今天天气很好"
Output: [192, 3847, 2093, 3847, 452, 2398]  # 6 tokens

English Tokenization Example:

Input: "hello world"
Output: [15339, 1917]  # 2 tokens

2. Self-Attention

Self-attention is the core of Transformer, enabling the model to "see" relationships between any positions in the text:

Query: "I want to go"
Key:   "Beijing weather"
Value: "Beijing temperature info"

Attention score = softmax(Q × K^T / √d)

3. Sampling Strategies

Strategy	Description	Use Case
Greedy	Select token with highest probability	Deterministic output
Temperature	Control randomness via temperature	Creative writing
Top-p	Sample from set with cumulative probability p	Balance quality and diversity

Temperature Effects:

Temperature	Effect
T < 1.0	More deterministic, conservative
T = 1.0	Balanced
T > 1.0	More random, creative

Inference Efficiency Optimization

KV Cache

Cache computed Key and Value to avoid redundant calculations:

Without KV Cache: Recalculate all historical tokens for each generation
With KV Cache: Only compute KV for new tokens, reuse history

Batching

Process multiple requests simultaneously to improve GPU utilization:

Single request: [User 1's input]
Batching: [User 1, User 2, User 3, ...] inputs

Quantization

Use lower precision (e.g., INT8) to reduce memory usage:

Precision	Memory	Quality Loss
FP32	100%	None
FP16	50%	Minimal
INT8	25%	Acceptable
INT4	12.5%	Task-dependent

Platform Optimization Strategies

ai.TokenHub provides efficient inference through:

• Smart Routing: Automatically select optimal model and node
• Distributed Inference: Multi-node collaboration for large requests
• Cache Reuse: Return cached results for identical requests
• Traffic Scheduling: Balance load through peak-shaving