Uncensored AI: Unleashed

Deploy, Run and Master Unfiltered LLMs in the Cloud

Introduction

1. The Alignment Tax
2. Why Self-Host?
3. What This Book Will Teach You
4. Who Should Read This

Chapter 1: Inside the LLM: Foundational Architecture

1. Attention Is All You Need: The Transformer Core
2. Autoregressive Generation and Autoregressive Loss
3. Parameters, Layers and Compute Budget
4. Tokens, Tokenization and Why They Matter
5. Context Windows and Memory

Chapter 2: The Model Zoo: Families, Sizes and Tradeoffs

1. The Llama Lineage: Meta and Its Forks
2. Mistral, Mixtral and Sparse MoE Architectures
3. Qwen and the Chinese Model Renaissance
4. DeepSeek and Open Weights for Reasoning
5. Smaller Models and Efficiency Champions
6. Choosing a Model Family for Your Use Case

Chapter 3: Alignment, Refusal and Safety: How It Works

1. Supervised Fine-Tuning and Instruction Tuning
2. Reinforcement Learning from Human Feedback
3. Direct Preference Optimization and Alignment Loss
4. The Anatomy of Refusal: How Models Learn to Say No
5. Alignment Tax and Capability Degradation
6. Measuring Refusal Rates and Overrefusal

Chapter 4: Uncensored, Abliterated and Unaligned Models

1. Base Versus Instruct Versus Uncensored
2. Fine-Tuning for Refusal Removal
3. What “Uncensored” Means in Practice
4. License Implications and Permissiveness
5. Ethical and Legal Considerations
6. Finding the Right Level of Uncensorship

Chapter 5: Abliteration: The Science and Practice

1. The Concept of Activation Steering
2. Finding Refusal Directions in Weight Space
3. Gradient Abliteration Explained
4. Implementing Abliteration: Step by Step
5. Limitations, Fail Modes and Safety
6. Evaluating Abliterated Models

Chapter 6: Model Formats, Precision and Quantization

1. Full Precision: FP32, FP16 and BF16
2. The Rise of FP8 and Mixed Precision
3. Integer Quantization: INT8 and INT4 Fundamentals
4. GGUF and the llama.cpp Ecosystem
5. GPTQ, AWQ and EXL2 for GPU Deployment
6. bitsandbytes and 4-bit Loading
7. Choosing the Right Format for Your Hardware

Chapter 7: Hugging Face: Discovery, Downloads and Model Cards

1. Navigating the Hub for Uncensored Models
2. Reading Between the Lines of Model Cards
3. Tags, Metrics and Quality Signals
4. Licensed, Permissive and Restricted Weights
5. Download Strategies: Git, CLI and Direct Links
6. Verifying Model Integrity and Safety

Chapter 8: The GPU Hardware Landscape

1. NVIDIA Architecture Generations and Tensor Cores
2. Consumer GPUs: RTX 4090 and RTX 5090
3. Datacenter Workhorses: T4, A10, L4, L40S
4. Flagship GPUs: A100, H100 and H200
5. The Blackwell Era: B200 and GB200
6. Hardware Selection by Model Size and Workload

Chapter 9: Cloud GPU Providers: A Comprehensive Comparison

1. Pricelist and Availability Comparison
2. Security and Privacy Models
3. Storage, Networking and Persistence
4. Reliability and Community Trust
5. Acceptable Use Policies and Model Restrictions
6. Decision Matrix: Which Provider for What Workload

Chapter 10: Deploying on RunPod: Secure Cloud and Serverless

1. Account Setup and Template Selection
2. Choosing Your GPU and Configuring Storage
3. Launching a Container with vLLM or Ollama
4. Exposing Your Endpoint with Port Configuration
5. Persistent Storage and Model Caching
6. Serverless Inference for Cost Efficiency

Chapter 11: Deploying on Vast.ai: Market-Driven GPU Access

1. Understanding the Marketplace Model
2. Evaluating Hosts and Instance Quality
3. Provisioning and Connecting to an Instance
4. Deploying Models with Prebuilt and Custom Templates
5. Network Security and SSH Configuration
6. Cost Control and Instance Management

Chapter 12: Deploying on CoreWeave and AWS

1. CoreWeave: The Specialized GPU Cloud
2. CoreWeave Account Setup and Instance Deployment
3. AWS EC2 GPU Instances: p5, p4, g5 and g6 Families
4. Configuring AWS Instances and Installing Drivers
5. Managing AWS GPU Costs and Spot Instances
6. Choosing Between CoreWeave and AWS

Chapter 13: Deploying on GCP, Azure, Lambda and Paperspace

1. GCP GPU Instances: A2, A3 and L2 Families
2. Azure GPU VMs: NV, ND and NC Series
3. Lambda Cloud: High-Density Direct-Connect GPUs
4. Paperspace and Gradient: Notebooks and GPU Instances
5. Cross-Provider Cost and Performance Comparison
6. Multi-Provider Strategies for Availability

Chapter 14: Inference Frameworks: The Engine Room

1. Hugging Face Transformers and Accelerate
2. vLLM: PagedAttention and Continuous Batching
3. SGLang: Structured Generation and RadixAttention
4. llama.cpp: CPU and Quantized GPU Inference
5. Ollama: Simplicity and Model Management
6. Text Generation Inference and ExLlamaV2
7. Framework Selection Matrix

Chapter 15: End-to-End Deployment: Building Your Inference Stack

1. Architecture Overview and Component Selection
2. Provisioning Your GPU Instance
3. Installing the Stack: Drivers, Python, Framework
4. Loading the Model and Starting Inference
5. Exposing the API: Nginx, TLS and Auth
6. Connecting Clients and Verifying Performance

Chapter 16: Building OpenAI-Compatible APIs and Chat Interfaces

1. OpenAI-Compatible API Endpoints
2. vLLM and Ollama OpenAI Compatibility
3. Open WebUI and Frontend Integration
4. Connecting Local Applications via API
5. Automation and Workflow Integration
6. Multi-Model Routers and Load Distribution

Chapter 17: Advanced Inference Optimization

1. Multi-GPU and Distributed Inference
2. Tensor Parallelism and Pipeline Parallelism
3. KV-Cache Management and Prefilling
4. Continuous Batching and Request Scheduling
5. Speculative Decoding and Draft Models
6. FlashAttention and Memory-Optimized Kernels

Chapter 18: Custom Uncensored Models: Fine-Tuning and Modification

1. Dataset Preparation for Uncensored Alignment
2. LoRA and QLoRA Fine-Tuning
3. Merging LoRA Adapters and Merging Models
4. Applying Abliteration to Your Model
5. Evaluating Refusal Behavior Changes
6. Saving, Versioning and Deploying Custom Checkpoints

Chapter 19: Benchmarking, Metrics and Performance Tuning

1. Key Metrics: TPS, TTFT, Throughput and Latency
2. Reproducible Benchmarking Scripts
3. Profiling VRAM and GPU Utilization
4. Tuning Parameters for Speed and Quality
5. Cost Per Token and Operational Economics
6. Comparative Benchmarking Across Frameworks

Chapter 20: Production Deployment: Scaling, Security and Operations

1. Production Architecture Patterns
2. Autoscaling and Load Balancing
3. Monitoring, Logging and Observability
4. Security Hardening and Network Isolation
5. Secrets Management and Access Control
6. Backups, Snapshots and Disaster Recovery

Conclusion: The Future of Uncensored Inference

References