AI Engineering

Implements and trains LLMs using LitGPT with 20+ architectures like Llama, Gemma, Phi. Use for clean model implementations, understanding architectures, or production fine-tuning with LoRA/QLoRA. Single-file, no abstraction layers.

Implements Mamba state-space models achieving O(n) complexity for sequence modeling. Enables 5× faster inference, million-token sequences, and no KV cache. Ideal for streaming and long-context applications.

A minimalist, educational GPT implementation in ~300 lines of code. Ideal for learning transformer architectures from scratch; supports training on Shakespeare (CPU) or reproducing GPT-2 on OpenWebText (multi-GPU).

A skill for using RWKV, the RNN-Transformer hybrid model that offers linear-time inference, infinite context, and constant memory. Covers installation, text generation, long context processing, fine-tuning, and comparison with Transformers.

Provides PyTorch-native distributed LLM pretraining using torchtitan with 4D parallelism (FSDP2, TP, PP, CP). Use when pretraining Llama 3.1, DeepSeek V3, or custom models at scale from 8 to 512+ GPUs with Float8, torch.compile, and distributed checkpointing.

Rust-based tokenization library for NLP that processes 1GB of text in under 20 seconds. Supports BPE, WordPiece, and Unigram algorithms, custom vocabulary training, and seamless transformers integration for research and production.

SentencePiece is a language-independent tokenizer that works on raw Unicode text without pre-tokenization. It supports BPE and Unigram algorithms for fast, deterministic vocabulary training, ideal for multilingual NLP models and CJK languages.

Provides expert assistance for configuring and running LLM fine-tuning with Axolotl, covering YAML configs, 100+ model architectures, LoRA/QLoRA, DPO/KTO/ORPO/GRPO, and multimodal support.

Expert guidance for fine-tuning LLMs with LLaMA-Factory: a no-code WebUI, support for 100+ models, QLoRA in 2-8 bits, and multimodal capabilities.

Fine-tune large language models (7B–70B) with limited GPU memory using LoRA and QLoRA. Train less than 1% of parameters with minimal accuracy loss, enabling efficient multi-adapter serving.

Expert guidance for fast fine-tuning with Unsloth, achieving 2-5x faster training and 50-80% less memory using LoRA/QLoRA optimization.

Guidance for interpreting and manipulating neural network internals using nnsight, with optional remote execution via NDIF. Use for interpretability experiments on massive models without local GPUs, or with any PyTorch architecture.