Dashboard 模式需额外安装 rich（pip install -e ".[educational]"）；FlashAttention 需单独安装 flash-attn 且为可选依赖；首次运行会从 HuggingFace 下载模型权重。
git clone https://github.com/yourusername/nano-vllm.git
cd nano-vllm
pip install -e . 硬件要求按 README 说明
README 未说明具体显存要求；示例使用 TinyLlama-1.1B，Dashboard 示例显示 Device: CUDA、Dtype: float16，说明面向 GPU 运行。
纯 CPU 可运行需要 CUDA
教学版 vLLM 迷你实现，带可视化讲解，帮你搞懂 LLM 推理到底怎么跑
用 Python 从零实现的教学版 vLLM，帮助理解大模型推理引擎的调度与 KV Cache 机制，适合边读代码边学习。
核心能力
- 用 PagedAttention 按块管理 KV 缓存
- 支持连续批处理与迭代级调度
- 实现 GQA 共享 KV 头降低显存
- 提供 Narrator/X-Ray/Dashboard/Tutorial 四种教学模式
- 支持投机解码与优先级调度、抢占
容易踩的坑
内容由模型根据 README 生成，可能不完整或不准确，以官方文档为准。看原文 README
暂无 release 数据（可能没有发版，或尚未同步）。
用户点评
登录用户的真实使用记录，可按版本和硬件筛选共 0 条
还没有人点评。第一个说出真实体验的人最有价值。
nano-vllm
A minimalistic, educational implementation of a high-performance LLM inference engine, inspired by vLLM.
Learn how LLM inference really works - from tokenization to generation, with real-time visualizations and interactive tutorials.
┌─────────────────────────────────────────────────────────────────┐
│ LLaMA Architecture │
├─────────────────────────────────────────────────────────────────┤
│ Input: "The capital of France is" │
│ │ │
│ ▼ │
│ ┌─────────────────┐ │
│ │ Tokenizer │ "The" → 450, "capital" → 7483, ... │
│ └────────┬────────┘ │
│ ▼ │
│ ┌─────────────────┐ │
│ │ Embedding │ 450 → [0.12, -0.34, 0.87, ...] (4096-d) │
│ └────────┬────────┘ │
│ ▼ │
│ ╔═════════════════╗ │
│ ║ Decoder Layer ║ ×32 layers │
│ ║ ┌───────────┐ ║ │
│ ║ │ Attention │◄─╬──── KV Cache │
│ ║ └─────┬─────┘ ║ │
│ ║ ┌───────────┐ ║ │
│ ║ │ FFN │ ║ │
│ ║ └───────────┘ ║ │
│ ╚═════════════════╝ │
│ ▼ │
│ "Paris" (87%) │
└─────────────────────────────────────────────────────────────────┘
Features
Core Inference Engine
- Continuous Batching: Process multiple requests simultaneously with iteration-level scheduling
- PagedAttention: Memory-efficient KV cache management (like OS virtual memory)
- FlashAttention: IO-aware fused attention for faster computation
- Grouped Query Attention (GQA): Efficient attention with shared KV heads
Advanced Scheduling
- Priority Scheduling: Process high-priority requests first
- Preemption: Pause low-priority requests for urgent ones
- Chunked Prefill: Break long prompts into manageable chunks
- Prefix Caching: Share KV cache blocks for common prefixes
Optimizations
- Speculative Decoding: Use a draft model to speed up generation
- Block-based Memory Management: Zero memory waste with on-demand allocation
Educational Modes
- Narrator Mode: Real-time plain-English explanations
- X-Ray Mode: Visualize tensor shapes and math operations
- Dashboard Mode: Live terminal UI with progress tracking
- Tutorial Mode: Interactive step-by-step learning experience
Installation
# Clone the repository
git clone https://github.com/yourusername/nano-vllm.git
cd nano-vllm
# Install in development mode
pip install -e .
# For dashboard mode (optional)
pip install -e ".[educational]"
Requirements
- Python 3.10+
- PyTorch 2.0+
- transformers
- safetensors
- rich (optional, for dashboard mode)
- flash-attn (optional, for FlashAttention)
Quick Start
Basic Generation
# Single prompt
python -m nano_vllm.cli --model TinyLlama/TinyLlama-1.1B-Chat-v1.0 \
--prompt "The capital of France is"
# Multiple prompts (continuous batching)
python -m nano_vllm.cli --model TinyLlama/TinyLlama-1.1B-Chat-v1.0 \
--prompt "The capital of France is" \
--prompt "The largest planet is" \
--prompt "Python is a"
Priority Scheduling
python -m nano_vllm.cli --model TinyLlama/TinyLlama-1.1B-Chat-v1.0 \
--prompt "Low priority task" --priority 1 \
--prompt "High priority task" --priority 10
Speculative Decoding
python -m nano_vllm.speculative.cli \
--target-model TinyLlama/TinyLlama-1.1B-Chat-v1.0 \
--draft-model TinyLlama/TinyLlama-1.1B-Chat-v1.0 \
--prompt "The future of AI is" \
--num-speculative-tokens 5
Educational Modes
nano-vllm includes four educational modes that help you understand what happens inside an LLM during inference. It's like watching surgery with an expert explaining each step!
1. Narrator Mode (--narrate)
Real-time plain-English commentary explaining each step as it happens.
python -m nano_vllm.cli --prompt "Hello world" --narrate
What you'll see:
═══════════════════════════════════════════════════════════════════
🎓 INFERENCE ANATOMY - Educational Mode
Understanding what happens inside an LLM
═══════════════════════════════════════════════════════════════════
📖 Prompt: "The capital of France is"
🤖 Model: TinyLlama/TinyLlama-1.1B-Chat-v1.0
🎬 ACT 1: TOKENIZATION
┌─────────────────────────────────────────────────────────────┐
│ Converting your prompt into numbers the model understands... │
└─────────────────────────────────────────────────────────────┘
"The capital of France is"
↓ Tokenizer (BPE algorithm)
[The] [capital] [of] [France] [is] → [450, 7483, 310, 3444, 338]
💡 WHY TOKENS? LLMs don't read text - they process numbers.
🎬 ACT 2: PREFILL PHASE (The "Reading" Phase)
┌─────────────────────────────────────────────────────────────┐
│ The model reads your entire prompt at once... │
└─────────────────────────────────────────────────────────────┘
🎬 ACT 3: DECODE PHASE (Token-by-Token Generation)
Step 1: Predicting token #6
│ Top 5 predictions:
│ ┌──────────────┬───────────┐
│ │ Paris │ 87.3% ██████████ │
│ │ the │ 4.2% █ │
│ │ Lyon │ 1.8% ░ │
│ └──────────────┴───────────┘
└── 🎲 Sampled: "Paris"
The narrator explains:
- Tokenization: How text becomes numbers
- Embedding: How token IDs become vectors
- Prefill: How the model "reads" the prompt
- Attention: How tokens communicate
- KV Cache: Why caching matters
- Decode: How tokens are generated one by one
- Sampling: How the next token is chosen
2. X-Ray Mode (--xray)
Shows the actual mathematics and tensor operations happening during inference.
python -m nano_vllm.cli --prompt "Hello" --xray
What you'll see:
╔══════════════════════════════════════════════════════════════╗
║ X-RAY: Layer 1 - Q/K/V Projections ║
╚══════════════════════════════════════════════════════════════╝
Step 1: Project to Q, K, V
────────────────────────────────
hidden_states: [1, 5, 4096]
(batch=1, seq=5, hidden=4096)
↓ Linear projections (learned weights)
Q: [1, 32, 5, 128] (32 heads, each 128 dims)
K: [1, 8, 5, 128] (8 KV heads for GQA)
V: [1, 8, 5, 128]
💡 GQA: 32 Q heads share 8 KV heads (4:1)
Memory saving: 4x less KV cache!
Step 2: Compute Attention Scores
─────────────────────────────────
scores = Q @ K^T / √128
[1,32,5,128] @ [1,32,128,5] → [1,32,5,5]
Q K^T scores
X-Ray mode shows:
- Tensor shapes at each step
- Matrix multiplication dimensions
- Memory calculations
- GQA head sharing
- Numerical examples from running inference
3. Dashboard Mode (--dashboard)
A rich terminal UI with live-updating visualizations (requires rich library).
pip install rich
python -m nano_vllm.cli --prompt "Hello" --dashboard
What you'll see:
╭─────────────────── nano-vllm Inference Dashboard ───────────────────╮
│ │
│ Model: TinyLlama-1.1B Device: CUDA Dtype: float16 │
│ │
│ ┌─ Progress ──────────────────────────────────────────────────┐ │
│ │ Phase: 📖 PREFILL │ │
│ │ Prefill: ████████████████████████████████████████░░░░ 85% │ │
│ │ Layer: ████████████████░░░░░░░░░░░░░░░░░░░░░░░░ 12/32 │ │
│ │ Decode: ░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░ 0/50 │ │
│ └──────────────────────────────────────────────────────────────┘ │
│ │
│ ┌─ Memory ──────────┐ ┌─ Throughput ─────────────┐ │
│ │ KV Cache: 24.5MB │ │ Prefill: 1,234 tok/s │ │
│ │ ████████░░ 80% │ │ Decode: 42 tok/s │ │
│ │ Blocks: 156/200 │ │ Total: 847 tok/s │ │
│ └───────────────────┘ └──────────────────────────┘ │
│ │
│ ┌─ Generated Text ──────────────────────────────────────────────┐ │
│ │ The capital of France is Paris.█ │ │
│ └────────────────────────────────────────────────────────────────┘ │
│ │
│ ┌─ Top Predictions ─────────────────────────────────────────────┐ │
│ │ Paris 87.3% ████████████████████████████████████ │ │
│ │ the 4.2% ████ │ │
│ │ a 2.1% ██ │ │
│ └────────────────────────────────────────────────────────────────┘ │
╰──────────────────────────────────────────────────────────────────────╯
Dashboard displays:
- Progress: Phase, layer, and decode progress bars
- Memory: KV cache usage and block allocation
- Throughput: Tokens per second metrics
- Generated Text: Live token generation
- Predictions: Top-k token probabilities
4. Tutorial Mode (--tutorial)
An interactive, step-by-step guided tour with explanations and quizzes.
python -m nano_vllm.cli --tutorial
What you'll experience:
═══════════════════════════════════════════════════════════════════
🎓 NANO-VLLM INTERACTIVE TUTORIAL
Understanding LLM Inference from the Inside Out
═══════════════════════════════════════════════════════════════════
This tutorial has 12 chapters covering:
Chapter 1: Tokenization
Chapter 2: Embedding Lookup
Chapter 3: Self-Attention
Chapter 4: Prefill Phase
Chapter 5: KV Cache
Chapter 6: Decode Phase
Chapter 7: Sampling
Chapter 8: PagedAttention
...
[Press Enter to continue...]
───────────────────────────────────────────────────────────────────
CHAPTER 3: Self-Attention
───────────────────────────────────────────────────────────────────
Self-attention lets each position 'attend' to all previous positions...
🧠 POP QUIZ:
Why is the attention mask 'causal' (triangular)?
[A] To save memory
[B] So tokens can only see past tokens, not future
[C] To make computation faster
[D] Because GPUs prefer triangular matrices
Your answer: _
The tutorial covers:
- Model architecture overview
- Tokenization and embeddings
- Self-attention mechanism
- Prefill vs decode phases
- KV cache optimization
- Sampling strategies
- PagedAttention memory management
Combining Modes
You can combine educational modes for a richer learning experience:
# Narrator + X-Ray: See explanations AND tensor math
python -m nano_vllm.cli --prompt "Hello" --narrate --xray
# All three visualization modes
python -m nano_vllm.cli --prompt "Hello" --narrate --xray --dashboard
Project Structure
nano_vllm/
├── __init__.py
├── config.py # ModelConfig dataclass
├── cache.py # KV cache implementations
├── sampler.py # Token sampling (greedy, top-k, etc.)
├── engine.py # Main inference engine
├── cli.py # Command-line interface
│
├── core/
│ ├── sequence.py # Sequence abstraction
│ ├── scheduler.py # Priority scheduler with preemption
│ ├── block.py # Block abstraction for PagedAttention
│ └── block_manager.py # KV cache block management
│
├── attention/
│ ├── paged_attention.py # PagedAttention implementation
│ └── flash_attention.py # FlashAttention integration
│
├── speculative/
│ ├── speculative_decoding.py # Speculative decoding engine
│ └── cli.py # CLI for speculative decoding
│
├── educational/ # 🎓 Educational visualization system
│ ├── narrator.py # Plain-English explanations
│ ├── xray.py # Tensor/math visualizations
│ ├── dashboard.py # Rich terminal UI
│ ├── tutorial.py # Interactive tutorial
│ ├── visualizers.py # ASCII art generators
│ ├── explanations.py # Educational text content
│ └── educational_engine.py # Engine with educational hooks
│
└── model/
├── loader.py # HuggingFace weight loading
└── llama.py # LLaMA model implementation
Key Concepts Explained
PagedAttention
Traditional inference pre-allocates KV cache for maximum sequence length, wasting memory:
Traditional: Pre-allocate max_length for each sequence
┌────────────────────────────────────────────────────┐
│ Seq A: █░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░ (1/32 used) │
│ Seq B: █████████░░░░░░░░░░░░░░░░░░░░░░░░ (9/32 used) │
└────────────────────────────────────────────────────┘
↓
PagedAttention: Allocate blocks on-demand
┌────┬────┬────┬────┬────┬────┬────┬────┐
│ B0 │ B1 │ B2 │ B3 │ B4 │ B5 │ B6 │ B7 │
│████│████│░░░░│████│████│░░░░│████│░░░░│
└────┴────┴────┴────┴────┴────┴────┴────┘
▲ ▲ ▲ ▲ ▲
└────┴─────────┴────┴─────────┘
Seq A Seq B
(2 blocks) (3 blocks)
Continuous Batching
Unlike static batching that waits for all sequences to finish:
Static Batching:
[Seq A ████████████████████]
[Seq B ██████████ ] ← Seq B waits for A
[Seq C ████ ] ← Seq C waits for A
Continuous Batching:
[Seq A ████████████████████]
[Seq B ██████████][Seq D ██████████] ← D joins when B finishes
[Seq C ████][Seq E ██████████████████] ← E joins when C finishes
KV Cache
Without caching, attention recomputes K,V for all tokens every step (O(n²)). With caching, we only compute K,V for new tokens (O(n)):
WITHOUT CACHE:
Step 1: Process [The] → 1 token
Step 2: Process [The][capital] → 2 tokens
Step 3: Process [The][capital][of] → 3 tokens
Total: 1+2+3+...+n = O(n²)
WITH CACHE:
Prefill: Process all tokens once → n tokens
Decode: Only process NEW token → 1 token each
Total: n + 1 + 1 + ... = O(n)
CLI Options
python -m nano_vllm.cli --help
| Option | Description |
|---|---|
--model | HuggingFace model ID or local path |
--prompt | Input prompt (can specify multiple) |
--max-tokens | Maximum tokens to generate (default: 50) |
--device | Device to run on (cuda/cpu) |
--dtype | Data type (float16/float32/bfloat16) |
--priority | Request priority (higher = more important) |
--scheduling-policy | fcfs or priority (default: priority) |
--max-prefill-tokens | Chunked prefill budget (default: 512) |
--no-paged-attention | Disable PagedAttention |
--no-flash-attn | Disable FlashAttention |
--no-prefix-caching | Disable prefix caching |
--show-memory-stats | Show memory statistics |
| Educational | |
--narrate | Enable narrator mode |
--xray | Enable X-Ray mode |
--dashboard | Enable dashboard mode |
--tutorial | Run interactive tutorial |
Development
Running Tests
pytest tests/
Running Benchmarks
Advanced Scheduling Benchmarks - Tests priority scheduling, preemption, chunked prefill, and prefix caching (requires GPU):
# Run all advanced scheduling benchmarks
python -m benchmarks.benchmark_advanced_scheduling --all
# Run specific benchmarks
python -m benchmarks.benchmark_advanced_scheduling --priority
python -m benchmarks.benchmark_advanced_scheduling --preemption
python -m benchmarks.benchmark_advanced_scheduling --chunked-prefill
python -m benchmarks.benchmark_advanced_scheduling --prefix-caching
Scheduler Unit Benchmarks - Fast benchmarks for scheduler logic (no GPU required):
python -m benchmarks.benchmark_scheduler_unit
Comparison Benchmarks - Compare nano-vllm vs HuggingFace Transformers:
python -m benchmarks.benchmark_comparison
References
- vLLM Paper: Efficient Memory Management for Large Language Model Serving
- FlashAttention Paper
- PagedAttention Blog Post
- vLLM Repository
License
MIT License - feel free to use this for learning and experimentation!
Contributing
Contributions are welcome! This is an educational project, so clarity and documentation are prioritized over raw performance.
Areas for contribution:
- Additional educational explanations
- More visualization modes
- Support for more model architectures
- Performance optimizations with educational documentation