Skill UI

Orchestrates PyTorch-native TorchTitan 4D parallelism (FSDP2, TP, PP, CP) for large-scale LLM pretraining, scaling from 8 to 512+ GPUs with Float8, torch.compile, and distributed checkpointing to boot Llama 3.1, DeepSeek V3, or custom models.

This skill provides comprehensive guidance on training Mixture of Experts (MoE) models, including architectures like Mixtral and DeepSeek-V3. It details techniques for achieving large-scale model capacity with significantly reduced computational costs (e.g., 5x reduction vs. dense models). Topics covered include top-k routing, load balancing, expert parallelism, and integration with frameworks like DeepSpeed and HuggingFace for efficient, resource-constrained training.

The Neon AI Gateway provides a single API endpoint and one credential to access a diverse catalog of frontier and open-source LLMs (including OpenAI, Anthropic, Google Gemini, Meta, Alibaba, and DeepSeek). This unified layer simplifies integration by eliminating the need for separate provider accounts and managing model-specific SDK complexities. It allows developers to switch between models seamlessly by changing only the model name, while also providing crucial branch-scoped isolation for development and CI environments.

This skill manages complex, stateful proof projects as a local-first pipeline. It guides the user through drafting, rigorous verification, and source bookkeeping. It systematically handles proof continuation across runs, maintaining a full audit trail. When local attempts stall, it orchestrates handoffs to external AI models (like GPT Pro) or performs optional deep audits (DeepSeek), ensuring a structured, traceable, and verifiable output suitable for academic or technical submission.

One-shot chat completion against DeepSeek's deepseek-chat model via the OpenAI-compatible chat completions endpoint. Designed for non-reasoning tasks such as summarization, extraction, and quick classification, with JSON output, environment-based API key handling, and graceful degradation when the API is unavailable.

Wraps DeepSeek's deepseek-reasoner model to expose chain-of-thought reasoning separately from the final answer. Enables multi-step reasoning tasks such as proofs, planning, root-cause analysis, and bug triage with explicit thinking output for audit or verification. Handles API key via environment, degrades gracefully when unavailable.

BlockRun enables Claude Code and Google Antigravity to access external AI capabilities like image generation and real-time X data via x402 micropayments. It uses a wallet system, supports budget control, and integrates with providers like OpenAI, xAI, and DeepSeek.

Runs DeepSeek Harness headlessly for bounded repository tasks. Delegates analysis or implementation to DeepSeek Harness, checks availability, handles authentication, and manages workspace modifications.