Skill UI

Provides expert guidance for DeepSpeed distributed training, covering ZeRO optimizations, pipeline parallelism, mixed precision, DeepNVMe I/O, and best practices for debugging or deploying large-scale models.

TorchTitan delivers PyTorch-native distributed LLM pretraining with composable 4D parallelism (FSDP2, TP, PP, CP), Float8, torch.compile, and checkpointing so you can scale Llama 3.1, DeepSeek V3, or custom models across 8 to 512+ GPUs.

Runs build, test, lint, serve, and other tasks in Nx workspaces via run, run-many, and affected flows, honoring the detected package manager and target metadata while exposing filters, parallelism, and caching controls for CI or local workflows.

Guides orchestrating multiple Task tool subagents with run_in_background for concurrent work, covering when to trigger parallelism, how to craft prompts, launch in one message, synchronize results, and integrate TodoWrite status updates.

Blueprint turns a single-line goal into a multi-session, multi-agent engineering plan by decomposing the work into context-rich, self-contained steps, applying dependency graphs, parallelism detection, adversarial reviews, and PR/CI workflows so new agents can execute safely.