skymcp

Analyzes SkyPilot cloud spending and optimizes GPU costs. Monitors active clusters, identifies waste (idle clusters, over-provisioned resources, missing autostop), and suggests concrete savings strategies with dollar amounts. Use proactively when expensive jobs are running or when reviewing cloud costs, or reactively when users ask about spending.

Validates SkyPilot YAML configurations before launch. Checks for known gotchas, estimates costs, verifies resource feasibility, and suggests optimizations. Use proactively before any sky launch or sky jobs launch command, or reactively when a user asks to review a config. Triggers when YAML files are being written or when launch commands are about to execute.

Designs and runs controlled ML experiments -- hyperparameter sweeps, architecture ablations, scaling law studies, and A/B comparisons. Use when comparing configurations, searching for optimal settings, or when the user discusses experimentation, ablation, or comparison. Triggers proactively after a training run completes to suggest next experiments.

Diagnoses and fixes ML training failures -- OOM crashes, NaN loss, loss plateaus, gradient explosion, slow throughput, data corruption, and preemption recovery. Use when training encounters problems or produces unexpected results. Triggers proactively when error patterns are detected in logs, or reactively when users report training issues.

Autonomous ML training lifecycle manager. Use when the user wants to train a model end-to-end -- framework selection, YAML generation, launch, monitoring, failure recovery, and iteration. Triggers proactively when training tasks are detected, such as dataset preparation discussions, model fine-tuning requests, or when training scripts exist but have not been launched.

Use when saving or resuming training checkpoints, merging LoRA adapters, converting between model formats (safetensors, GGUF, PyTorch), quantizing models, using mergekit for model merging, or managing checkpoint lifecycle on cloud storage

Use when reducing cloud GPU costs, choosing spot vs on-demand, estimating training expenses, managing budgets, configuring auto-stop, or optimizing spend across clouds - covers SkyPilot optimizer, spot failover, pricing tiers, and budget management

Use when designing a data pipeline, curating pretraining data, choosing between NeMo Curator and other tools, configuring tokenization, running deduplication (exact or fuzzy), applying data filtering or quality classifiers, working with FineWeb or Dolma or RedPajama, or assessing data quality for ML training - the production data pipeline design reference

Use when setting up multi-GPU or multi-node training, configuring torchrun or deepspeed launchers, choosing parallelism strategy (FSDP, DeepSpeed ZeRO, tensor parallel, pipeline parallel), tuning NCCL, or enabling high-performance networking with SkyPilot

Use when selecting a training framework, comparing NeMo vs Axolotl vs torchtune vs TRL vs DeepSpeed vs Megatron, choosing between FSDP and DeepSpeed, deciding how to fine-tune or pretrain a model, or configuring LoRA/QLoRA/full fine-tuning - the definitive framework selection guide for ML training at any scale

3 hooks across 3 events