megatron-memory-estimator

This repository provides expert-level skills for AI infrastructure engineering tasks. Each skill packages domain knowledge, code examples, and best practices to transform Claude into a specialized developer for specific frameworks and tools.

Construction Methodology (Unless Otherwise Specified)

1. Knowledge Gathering: Use Gemini DeepResearch to collect comprehensive, up-to-date information on target frameworks
2. Skill Development: Transform research into structured skills using skill-creator in Claude Code
3. Validation: Test skill-generated code examples to ensure correctness
4. Maintenance: Regular updates based on latest official documentation

TileLang Developer

Write high-performance GPU kernels using TileLang for NVIDIA, AMD, and Ascend hardware.

Capabilities:

• Matrix multiplication (GEMM) kernels
• FlashAttention implementations
• DeepSeek MLA operators
• Performance optimization (swizzle layouts, pipelining, warp specialization)
• Cross-platform kernel development

Status: ✅ Complete

Megatron Memory Estimator

Estimate GPU memory usage for Megatron-based MoE and dense models. Built upon [megatron_memory_estimator](https://huggingface.co/spaces/ISEEKYAN/megatron_memory_estimator).

Capabilities:

• Estimate memory from HuggingFace configs
• Support for MoE models (DeepSeek-V3, Qwen, etc.)
• Parallelism strategy comparison (TP/PP/EP/CP)
• Memory optimization recommendations

Status: ✅ Complete

SLIME User

Guide for using SLIME (LLM post-training framework for RL Scaling). Built upon [THUDM/slime](https://github.com/THUDM/slime).

Capabilities:

• RL training setup and configuration (GRPO, GSPO, PPO, Reinforce++)
• Multi-turn tool calling and agent workflows
• Custom reward models and generation functions
• Megatron and FSDP backend configuration
• SGLang integration and optimization
• Dynamic sampling and partial rollout
• Multi-node distributed training

Status: ✅ Complete

Prompt to create this skill, with Sonnet 4.5:

```

Use skill-creator to create a skill called slime-user at this repo. slime is an LLM

post-training framework for RL Scaling. Its repo is https://github.com/THUDM/slime.

Skill creation procedure:

1. Git clone the latest repo
2. Analyze docs/en, understand basic structure and write a doc navigation guide for user

getting started or finding docs for advanced usage

1. Gather valuable examples from the docs and examples dir, write key ideas and script

path down for quick reference

1. Checkout some important source code, for example slime/slime/utils/arguments.py and

slime/rollout/sglang_rollout.py, provide its path and functions for a quick find.

```

TikZ Flowchart

Create professional flowcharts and architecture diagrams using LaTeX TikZ with standardized styles.

Capabilities:

• Professional flowcharts with Google Material-like color palette
• Standardized node types (data, memory, operation, kernel boxes)
• Architecture diagrams and process flows
• Grouping and layout best practices
• Clean orthogonal edges and relative positioning

Status: ✅ Complete

SGLang Developer

Development skill for SGLang (Structured Generation Language) runtime and optimization.

Planned capabilities:

• SGLang runtime configuration
• Custom sampling strategies
• Performance tuning for LLM inference
• Multi-GPU serving optimization

Status: 🚧 Planned

vLLM Developer

Skill for vLLM engine development and deployment.

Planned capabilities:

• PagedAttention implementation
• Custom scheduler development
• Multi-LoRA