Thank you for sending your enquiry! One of our team members will contact you shortly.
Thank you for sending your booking! One of our team members will contact you shortly.
Course Outline
AI Sovereignty and Local Deployment of LLMs
- Risks associated with cloud LLMs: issues related to data retention, potential use of inputs for training, and foreign jurisdictional controls.
- Ollama architecture: understanding the model server, registry functionality, and OpenAI-compatible API structure.
- Comparative analysis with vLLM, llama.cpp, and Text Generation Inference.
- Overview of model licensing terms for Llama, Mistral, Qwen, and Gemma.
Installation and Hardware Configuration
- Setting up Ollama on Linux systems with CUDA and ROCm support.
- Implementing CPU-only fallback options and optimizing for AVX/AVX2 instructions.
- Deploying via Docker and configuring persistent volume mappings.
- Configuring multi-GPU environments and strategies for VRAM allocation.
Model Management
- Downloading models from the Ollama registry using commands like 'ollama pull llama3'.
- Importing GGUF format models sourced from HuggingFace and TheBloke repositories.
- Evaluating quantization levels, such as Q4_K_M, Q5_K_M, and Q8_0, to balance performance and size.
- Managing model switching and understanding limits on concurrent model loading.
Custom Modelfiles
- Syntax structure for writing Modelfiles: utilizing FROM, PARAMETER, SYSTEM, and TEMPLATE directives.
- Tuning parameters such as temperature, top_p, and repeat_penalty.
- Engineering system prompts to define role-specific behaviors.
- Creating and publishing custom models to the local registry.
API Integration
- Utilizing the OpenAI-compatible '/v1/chat/completions' endpoint.
- Handling streaming responses and enabling JSON mode.
- Integrating local LLMs with frameworks like LangChain, LlamaIndex, and bespoke applications.
- Implementing authentication and rate limiting via reverse proxy configurations.
Performance Optimization
- Optimizing context window sizing and managing KV cache efficiency.
- Executing batch inference and handling parallel requests effectively.
- Allocating CPU threads and maintaining NUMA awareness.
- Monitoring GPU utilization metrics and memory pressure levels.
Security and Compliance
- Establishing network isolation for endpoints serving models.
- Implementing input filtering and output moderation pipelines.
- Maintaining audit logs for prompts and generated completions.
- Verifying model provenance through hash verification.
Requirements
- Intermediate proficiency in Linux administration and container management.
- A high-level understanding of machine learning concepts and transformer model architectures.
- Familiarity with REST APIs and JSON data formats.
Target Audience
- AI engineers and developers looking to substitute cloud LLM APIs with on-premises solutions.
- Organizations handling sensitive data that restricts the use of cloud-based models.
- Government and defense teams necessitating air-gapped language model capabilities.
14 Hours
