Model Online / On-Site
Duration
Modules 5 modules

Course Objective & Scope

Parameter-Efficient Fine-Tuning (PEFT) updates less than one percent of a foundation model's parameters—slashing training hardware requirements while making it possible to serve dozens of task- or tenant-specific adapters from a single base model in GPU memory. This course explores the full PEFT landscape from mathematical foundations to multi-adapter serving.

Hands-on labs cover additive methods (prompt/prefix tuning), rescaling (IA3), and reparameterisation variants (LoRA, rsLoRA, DoRA, VeRA, LoftQ, PiSSA), exact VRAM footprint accounting, and dynamic multi-LoRA serving with vLLM and SGLang. Participants design and benchmark a shared-base multi-adapter platform.

Who Should Attend?

  • AI and ML engineers comparing and optimising PEFT algorithms
  • Platform engineers serving multi-tenant or multi-task custom models in production
  • Data scientists maximising adaptation quality under tight GPU memory budgets
  • Prerequisite: Python, PyTorch and foundational Transformer architecture knowledge

Key Takeaways

  • Select among LoRA, DoRA, rsLoRA, VeRA, IA3 and LoftQ based on task complexity and memory constraints
  • Calculate exact GPU VRAM consumption across weights, optimizer states, gradients and activations
  • Stabilise high-rank adapters with rsLoRA and boost quality with weight-decomposed DoRA
  • Serve multiple unmerged adapters concurrently on a single base model using vLLM and SGLang
  • Manage an adapter registry and lifecycle for multi-department or multi-tenant deployments

Curriculum

PEFT Taxonomy and GPU Memory Anatomy
  • Why full fine-tuning consumes 16–20x more VRAM than inference
  • Comparing Additive (Prefix/Prompt Tuning), Scaling (IA3) and Reparameterisation (the LoRA family)
  • Gradient checkpointing and activation memory trade-offs
Next-Generation Variants: DoRA, rsLoRA, PiSSA and LoftQ
  • Rank-Stabilized LoRA (rsLoRA) for scaling to higher ranks without gradient collapse
  • Weight-Decomposed Low-Rank Adaptation (DoRA): separating magnitude from direction
  • Principal singular-value initialisation (PiSSA) and quantisation-aware LoftQ
Advanced Configuration with Hugging Face PEFT
  • Targeting Attention and MLP projections and using modules_to_save for custom tokens
  • Loading, switching and combining multiple adapters in memory (add_weighted_adapter)
Multi-LoRA Production Serving with vLLM and SGLang
  • Merged single-model serving versus in-flight batching with Punica / S-LoRA CUDA kernels
  • Per-request adapter routing, GPU/CPU LRU adapter caching and latency profiling in vLLM
Workshop: Multi-Task PEFT Platform
  • Benchmarking LoRA, rsLoRA and DoRA on the same enterprise dataset
  • Serving three department-specific adapters concurrently from a single vLLM instance

Other Courses in This Category

  • Practical LLM Fine-Tuning with LoRA and QLoRA

    Hands-on training on Low-Rank Adaptation (LoRA) and 4-bit NormalFloat (NF4) QLoRA: rank/alpha tuning, target modules, double quantisation, paged optimizers and safe adapter merging.

    LLM Customisation & Model Adaptation Corporate
    Live Course
    5 modules
    View
  • Instruction Tuning: Following Complex Instructions, Schemas and Domain Constraints

    Course on designing and running instruction tuning so models reliably follow complex system prompts, strict JSON/schema constraints, corporate tone and refusal boundaries.

    LLM Customisation & Model Adaptation Corporate
    Live Course
    5 modules
    View
  • Preference Optimization: DPO, ORPO, KTO and LLM Alignment

    Advanced course on post-SFT alignment using DPO, IPO, KTO, ORPO and SimPO: building chosen/rejected preference datasets, tuning beta and KL constraints, and avoiding length bias and reward hacking.

    LLM Customisation & Model Adaptation Corporate
    Live Course
    5 modules
    View
  • Synthetic Data Generation: Data Synthesis and Distillation for LLM Training

    Hands-on course on generating high-diversity SFT and preference datasets in low-resource enterprise domains using teacher LLMs, Evol-Instruct, persona matrices, Distilabel and verification filters.

    LLM Customisation & Model Adaptation Corporate
    Live Course
    5 modules
    View
  • Training Data Design: Dataset Architecture for LLM Fine-Tuning

    Hands-on course on designing high-leverage SFT and multi-turn conversation datasets: the LIMA principle, task mixture ratios, expert annotation guidelines, schema design and dataset versioning.

    LLM Customisation & Model Adaptation Corporate
    Live Course
    5 modules
    View
  • Model Evaluation and Fine-Tuning Evals for Custom LLMs

    Course on evaluating fine-tuned LLMs for task accuracy, catastrophic forgetting regressions, hallucination and safety drift using lm-evaluation-harness, custom task suites and pairwise arena comparisons.

    LLM Customisation & Model Adaptation Corporate
    Live Course
    5 modules
    View

Get in touch