ML Infrastructure Optimization & Troubleshooting
Debug ML infrastructure bottlenecks, fix distributed training failures, optimize performance
What You Can Do
You can systematically diagnose performance issues across your ML infrastructure, identify root causes of distributed training failures, and receive actionable optimization recommendations. This skill analyzes metrics, logs, and configuration data to pinpoint resource constraints, communication overhead, and configuration misalignments that impact training speed and reliability.
Features
Analyzes CPU, GPU, memory, and I/O utilization patterns to identify where your pipeline loses time
Traces collective communication failures, gradient synchronization issues, and node-to-node connectivity problems
Maps compute, data loading, and network stages to spot inefficiencies from data ingestion through model checkpointing
Reviews hyperparameters, batch sizes, gradient accumulation, and communication backends for alignment with your hardware
Recommends GPU/CPU/memory configurations based on model size, data throughput, and fault tolerance requirements
Extracts and correlates anomalies across training logs, system metrics, and error traces to surface hidden dependencies
Diagnoses clock skew, network latency, and synchronization issues across distributed clusters
Provides priority-ordered fixes with implementation guidance and expected performance gains
Example Output
Example 1: Bottleneck Analysis
- 🔴 CRITICAL: GPU utilization 32% (target 80%+)
├─ Root cause: Data loading blocking GPU pipeline
├─ Symptom: 4.2s train step (2.8s data load, 1.4s forward/backward)
└─ Fix: Enable async prefetching, increase DataLoader workers from 4 → 8
Expected improvement: 65% reduction in step time
- ⚠️ WARNING: AllGather communication overhead 18% of step time
├─ Symptom: Gradient synchronization across 8 GPUs takes 1.1s
└─ Recommendation: Switch NCCL backend to use NVLink + enable gradient compression
Example 2: Distributed Training Failure
- 🚨 Failure pattern detected across nodes:
├─ Node 3 shows hanging AllReduce at iteration 45
├─ Network trace: Packet loss 0.2% between Node 1 ↔ Node 3
├─ Root cause: Network congestion during weight synchronization
├─ Temporary fix: Increase NCCL timeout from 30s → 120s (masks symptom)
└─ Permanent fix: Route collective ops through dedicated network interface
What's Included
- SKILL.md: Full infrastructure analysis workflow with decision trees for diagnosis
- Performance profiling checklist: Metrics to collect, interpret, and correlate
- Configuration templates: DDP, FSDP, Horovod, Ray Train setup examples
- Log parsing patterns: Regular expressions and extraction rules for common ML frameworks
- Troubleshooting flowchart: Decision tree from symptom → root cause → fix
- Benchmark comparison tool: Script to baseline performance against known good configurations
Who It's For
- ML infrastructure engineers — Debug production training pipelines and scale distributed models
- Research scientists — Optimize hyperparameter sweeps and reduce iteration time on clusters
- DevOps/platform engineers — Right-size hardware allocation and improve cluster utilization
- MLOps teams — Implement monitoring and automated diagnosis for training pipeline reliability
- Data center operators — Identify network and resource constraints affecting ML workloads
Best For
- Diagnosing why distributed training hangs or crashes across multiple nodes
- Identifying data loading bottlenecks blocking GPU pipeline efficiency
- Analyzing collective communication overhead (AllGather, AllReduce) in multi-GPU setups
- Right-sizing batch size, gradient accumulation, and communication backends for your hardware
- Troubleshooting out-of-memory errors and resource allocation failures
- Profiling end-to-end pipeline performance from data ingestion through checkpoint save
You might also like
You can build production-grade vector search systems that retrieve semantically similar documents in sub-millisecond time using AgentDB's optimized HNSW indexing. The skill enables you to implement RAG pipelines, semantic search engines, and intelligent knowledge bases with configurable embedding dimensions, distance metrics (cosine, Euclidean, dot product), and similarity thresholds—all with built-in quantization and caching for massive performance gains.
by moonwell-fi
This skill transforms high-level feature requests into actionable implementation plans by conducting deep requirement interviews, analyzing existing codebases (or planning greenfield projects), and producing comprehensive handoff documents. It distinguishes between brownfield development (with existing code to integrate) and greenfield projects (starting from scratch), ensuring your technical strategy accounts for your actual constraints and patterns.
by moonwell-fi
Diagnose root causes of model performance failures, select optimal evaluation metrics based on your use case and data characteristics, and design comprehensive evaluation strategies. You'll get systematic workflows for analyzing failure modes, statistical significance testing, threshold optimization, and cross-validation planning—turning vague model problems into actionable fixes.
by moonwell-fi
This skill helps you systematically diagnose failures in distributed ML training and serving infrastructure. You provide system logs, metrics, and error traces, and Claude performs structured root-cause analysis to identify the underlying issue—whether it's resource exhaustion, distributed system deadlock, data pipeline corruption, or model serving misconfiguration. You get a detailed diagnosis with remediation steps ranked by likelihood and implementation effort.
by moonwell-fi
Systematically evaluate your Claude prompts using quantitative scoring frameworks to identify weaknesses in clarity, task alignment, and output quality. You'll receive detailed optimization recommendations with A/B testing guidance to improve prompt effectiveness and consistency across different use cases and input variations.
by moonwell-fi
You can prepare your domain-specific datasets, train fine-tuned Claude models optimized for your use case, benchmark performance against base models, and deploy production-ready custom models. This skill automates the entire fine-tuning pipeline from raw data to live inference, complete with cost-benefit analysis and version control.
by moonwell-fi
You get structured guidance for designing controlled experiments, selecting appropriate metrics, and analyzing results with statistical rigor. Claude helps you identify confounding variables, calculate required sample sizes, interpret statistical significance, and document methodology for reproducible benchmarking that stands up to scrutiny.
by moonwell-fi
You can confidently navigate complex statistical analyses by working through structured workflows that match your research question and data type to the right test. The skill guides you through assumption validation, helps you interpret results accurately without common pitfalls, and walks you through regression diagnostics with full transparency about what your findings actually mean.
by moonwell-fi