Add a new HF MoE model to the transformersmodelingbackend experiment. Use when the user provides a HuggingFace model ID and wants to integrate it with Titan's MoE, or invokes /addmoemodel.
EnglishProvides PyTorch-native distributed LLM pretraining using torchtitan with 4D parallelism (FSDP2, TP, PP, CP).
EnglishProfiler-driven hill-climbing to close the inference throughput gap between TorchTitan's unified model (running inside vLLM) and vLLM's native model. Benchmark with generate.py --benchmark, climb optimization rungs (compile / cudagraph / fused kernels), profile torchtitan vs the native target, then patch the single biggest gap at a time and re-measure. Use when the user wants to benchmark or optimize RL inference generation speed, reproduce previous hill climbing study, or invokes /inferenceperfhillclimb.
EnglishCapture and compare per-op activations between two TorchTitan runs to spot numerics divergence (eager vs aotfxtrace, FSDP vs no-FSDP, before vs after a refactor). Use when the user wants to debug bitwise / numeric drift in training, or invokes /numericsdebugging.
EnglishBisect PyTorch commits to find the regression that breaks TorchTitan. Use when the user wants to bisect PyTorch or invokes /torchbisect.
English
AGENT SKILLS ATLASIncludes Add Moe Model, Distributed LLM Pretraining Torchtitan, Inference Perf Hillclimb, and other workflows.
Claude Code
Codex
Cursor
GitHub Copilot
Gemini CLISupported by the Atlas install prompt. Skill-specific requirements vary.
Pack stats
What’s included
Add Moe Model
Add a new HF MoE model to the transformersmodelingbackend experiment. Use when the user provides a HuggingFace model ID and wants to integrate it with Titan's MoE, or invokes /addmoemodel.
Distributed LLM Pretraining Torchtitan
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 model
Inference Perf Hillclimb
Profiler-driven hill-climbing to close the inference throughput gap between TorchTitan's unified model (running inside vLLM) and vLLM's native model. Benchmark with generate.py --benchmark, climb optimization rungs (compile / cudagraph / fused kernels), profile torchtitan vs the native target, then patch the single biggest gap at a time and re-measure. Use when the user wants to benchmark or optimize RL inference generation speed, reproduce previous hill climbing study, or invokes /inferenceperfhillclimb.
Numerics Debugging
Capture and compare per-op activations between two TorchTitan runs to spot numerics divergence (eager vs aotfxtrace, FSDP vs no-FSDP, before vs after a refactor). Use when the user wants to debug bitwise / numeric drift in training, or invokes /numericsdebugging.
Torch Bisect
Bisect PyTorch commits to find the regression that breaks TorchTitan. Use when the user wants to bisect PyTorch or invokes /torchbisect.
Quick start
- Choose your skills. Select individual skills above, or choose the entire pack.
- Copy the install prompt. Use Copy install prompt above the list.
- Send it to your agent. Paste into your coding agent’s chat. Your agent reviews the source and guides the setup.
Try it with your agent
After installation, start with a request like:
Use the Add Moe Model skill to help with my task.
Ask me for the context you need before getting started.Requirements & license
- A coding agent that can access your project files and install skills.
- Your agent checks each skill’s instructions, dependencies, and permissions during setup.
- Reported licenses for the skills shown: BSD-3-Clause, MIT. The author’s terms apply.