Skill UI

Guides production-grade miles RL training for massive MoE models, covering FP8/INT4 low precision, train-inference alignment, and speculative rollout workflows to maximize throughput in enterprise deployments.

Mastering the full lifecycle of machine learning models from data ingestion to deployment. This expertise covers building robust, production-grade MLOps pipelines using orchestrators like Kubeflow, Airflow, or Prefect. Key skills include feature store implementation with Feast, rigorous data validation, experiment tracking (e.g., MLflow), model registry management, and automated deployment workflows. Essential for automating model training and ensuring reproducibility at scale.

Flash Attention optimizer delivers 2-4x speedup and 10-20x memory savings for transformer attention, ideal when training or running long-context models that hit GPU memory or latency limits; supports PyTorch native scaled_dot_product_attention, the flash-attn library, H100 FP8 precision, and sliding-window attention.

Guides adding PyTorch FSDP2’s fully_shard integration to training loops, covering correct distributed init, bottom-up sharding, mixed precision/offload policies, DTensor-aware optimizers, and distributed checkpointing so oversized models or DeviceMesh sharding goals work reliably.

PyTorch Lightning Toolkit streamlines PyTorch model training by organizing LightningModules, LightningDataModules, Trainer configs, callbacks, logging, and distributed strategies for scalable multi-GPU/TPU runs in research or production workflows.

This guide provides expert strategies for optimizing GPU inference performance on CoreWeave infrastructure. It covers GPU selection based on workload (LLM, image generation, training), advanced techniques like continuous batching (using vLLM), autoscaling setup (HPA), and benchmarking data. Use this to maximize GPU utilization, minimize latency, and optimize large-scale AI model deployment.

A comprehensive checklist for ensuring GPU workloads, such as inference services or model training pipelines, are fully prepared for production deployment on CoreWeave. It covers critical MLOps and DevOps aspects including autoscaling, resource limits, security policies, persistent storage validation, monitoring setup, and rollback procedures.

qzcli is a powerful command-line interface (CLI) tool designed for managing complex GPU compute jobs on the Qizhi platform. Modeled after industry standards like kubectl, it allows users to log in, discover available resources (nodes, workspaces), submit interactive or highly complex batch jobs (including matrix sweeps), and monitor job status efficiently. It is essential for handling high-performance computing (HPC) and distributed machine learning training workflows.

A comprehensive workflow guide for fine-tuning open-source Large Language Models (LLMs) using the Together AI API. This process covers preparing structured training data (JSONL format), uploading the file, initiating the fine-tuning job, monitoring the training status, and finally deploying the custom model for inference. Ideal for developers needing to customize models for specific enterprise tasks, such as customer support or specialized content generation.