Draw Sebastian-Raschka-gallery-style TikZ architecture diagrams for any HuggingFace decoder-only LLM, with per-block parameter formulas and concrete numbers. Supports MHA, GQA, MLA, DeepSeek-V4-Flash (Hyper-Connections + Sparse Attention with learned indexer), dense and MoE FFNs (incl. hash routing), and MTP heads. Use when the user asks to visualize / diagram / illustrate a transformer or LLM architecture (DeepSeek, Qwen, Llama, Mistral, gpt-oss, etc.), wants a Raschka-style figure, or wants a TikZ/LaTeX rendering of an HF model.
Creates professional TikZ flowcharts with standardized themes, including Google Material-like and Anthropic-inspired options.
Create and revise pure HTML/CSS flowcharts using an Anthropic-inspired design language. Use when Codex needs to produce process diagrams, decision trees, pipelines, or system flows that should share warm ivory backgrounds, transparent dashed grouping containers, pastel node fills, SF Pro-style sans-serif labels, smaller rounded corners, quiet orthogonal connectors, and theme-tinted text hierarchy in standalone `.html` outputs.
Create presentation slides using Material You (Material Design 3) style. Generates 1280x720 HTML slides with M3 color tokens, Roboto typography, rounded cards, flow diagrams, metric cards, code blocks, and structured layouts. Use when the user asks to create slides, presentations, or decks and wants a clean, modern Material Design 3 aesthetic.
Guide for using SLIME (LLM post-training framework for RL Scaling). Use when working with SLIME for reinforcement learning training of language models, including setup, configuration, training execution, multi-turn interactions, custom reward models, tool calling scenarios, or troubleshooting SLIME workflows. Covers GRPO, GSPO, PPO, Reinforce++, multi-agent RL, VLM training, FSDP/Megatron backends, SGLang integration, dynamic sampling, and custom generation functions.
Estimate GPU memory usage for Megatron-based MoE (Mixture of Experts) and dense models. Use when users need to (1) estimate memory from HuggingFace model configs (DeepSeek-V3, Qwen, etc.), (2) plan GPU resource allocation for training, (3) compare different parallelism strategies (TP/PP/EP/CP), (4) determine if a model fits in available GPU memory, or (5) optimize training configurations for memory efficiency.
Write, optimize, and debug high-performance AI compute kernels using TileLang (a Python DSL for GPU programming). Use when the user requests: (1) Writing custom GPU kernels for AI workloads (GEMM, Attention, MLA, etc.), (2) Optimizing existing TileLang code for NVIDIA, AMD, or Ascend hardware, (3) Implementing non-standard operators (like DeepSeek MLA, FlashAttention variants), (4) Debugging TileLang compilation or runtime errors, or (5) Cross-platform kernel development targeting multiple GPU vendors.