dfastllm 0.0.4

High-performance inference engine for Diffusion Language Models - 3x faster with advanced optimizations

DFastLLM

🚀 High-Performance Inference Engine for Diffusion Language Models

Quick Start • Features • Performance • Documentation • Contributing

---

🎯 What is DFastLLM?

DFastLLM is a production-ready inference engine optimized for Diffusion Language Models (LLaDA, Dream, MDLM). Unlike autoregressive models that generate tokens sequentially, diffusion LLMs generate multiple tokens in parallel through iterative denoising — enabling massive throughput gains.

Traditional LLM:     Token → Token → Token → Token (sequential)
Diffusion LLM:       [████████] → [████████] → Done! (parallel)

---

⚡ Quick Start

Installation

pip install dfastllm

Generate Text

import torch
from transformers import AutoModelForCausalLM, AutoTokenizer
from dfastllm.engine.diffusion import DiffusionEngine

# Load model
model = AutoModelForCausalLM.from_pretrained(
    "GSAI-ML/LLaDA-8B-Instruct",
    torch_dtype=torch.bfloat16,
    device_map="auto",
    trust_remote_code=True,
)
tokenizer = AutoTokenizer.from_pretrained("GSAI-ML/LLaDA-8B-Instruct", trust_remote_code=True)

# Create engine and generate
engine = DiffusionEngine(model, tokenizer)
output = engine.generate("What is artificial intelligence?", max_tokens=64)
print(output)

With Quantization (2x Memory Savings)

pip install dfastllm[quantization]

from dfastllm import load_quantized_model

# Load 8B model in ~8GB instead of ~16GB
model = load_quantized_model("GSAI-ML/LLaDA-8B-Instruct", "int4")

Batch Processing (4x Throughput)

prompts = ["What is AI?", "What is ML?", "What is DL?", "What is NLP?"]
outputs = engine.generate(prompts, max_tokens=64)
# Throughput: 1,000+ tok/s

---

🔥 Features

Feature	Description	Status
Diffusion Generation	Parallel token unmasking	✅
Batch Processing	Process multiple requests	✅
INT4/INT8 Quantization	2-4x memory reduction	✅
torch.compile	JIT compilation for 2x speedup	✅
FlashAttention	Memory-efficient attention	✅
Multi-GPU	Tensor parallelism	✅
OpenAI API	Drop-in compatible server	✅
Streaming	Real-time token streaming	✅
CUDA Graphs	Zero-overhead inference	✅
Kubernetes	Production deployment	✅

---

📊 Performance

Benchmarked on NVIDIA L40S (46GB) with LLaDA-8B:

Batch Size	Throughput	Latency	Speedup
1	265 tok/s	241 ms	1.0x
2	484 tok/s	132 ms	1.8x
4	786 tok/s	81 ms	3.0x
8	1,056 tok/s	61 ms	4.0x

Memory Usage

Configuration	Memory	Notes
BF16	16.8 GB	Default
INT8	~10 GB	1.7x reduction
INT4	~6 GB	2.8x reduction

---

🐳 Docker

# GPU image
docker run --gpus all -p 8000:8000 ghcr.io/dfastllm-project/dfastllm:gpu

# CPU image
docker run -p 8000:8000 ghcr.io/dfastllm-project/dfastllm:latest

---

☸️ Kubernetes

apiVersion: apps/v1
kind: Deployment
metadata:
  name: dfastllm
spec:
  replicas: 1
  template:
    spec:
      containers:
      - name: dfastllm
        image: ghcr.io/dfastllm-project/dfastllm:gpu
        resources:
          limits:
            nvidia.com/gpu: 1
        ports:
        - containerPort: 8000

---

🌐 OpenAI-Compatible API

Start the server:

dfastllm-serve --model GSAI-ML/LLaDA-8B-Instruct --port 8000

Use with OpenAI client:

from openai import OpenAI

client = OpenAI(base_url="http://localhost:8000/v1", api_key="not-needed")

response = client.chat.completions.create(
    model="GSAI-ML/LLaDA-8B-Instruct",
    messages=[{"role": "user", "content": "What is AI?"}],
    stream=True,
)

for chunk in response:
    print(chunk.choices[0].delta.content, end="")

---

🛠️ Supported Models

Model	Parameters	Status
LLaDA-8B-Instruct	8B	✅ Full Support
LLaDA-8B-Base	8B	✅ Full Support
Dream	7B	⚠️ Experimental
MDLM	Various	⚠️ Experimental

---

📚 Documentation

• 📖 Documentation Website
• 🚀 Getting Started
• ⚙️ Configuration
• 📊 Benchmarks
• ☸️ Kubernetes Deployment

---

🤝 Contributing

We welcome contributions! Here's how to get started:

# Clone the repo
git clone https://github.com/dfastllm-project/dfastllm.git
cd dfastllm

# Install dev dependencies
pip install -e ".[dev]"

# Run tests
pytest tests/

# Run linting
ruff check dfastllm/
black --check dfastllm/

See CONTRIBUTING.md for detailed guidelines.

---

📄 License

Apache 2.0 - See LICENSE for details.

---

🙏 Acknowledgments

• LLaDA - The primary diffusion LLM we support
• HuggingFace Transformers - Model loading infrastructure
• PyTorch - Deep learning framework

---

Made with ❤️ by the DFastLLM Team