ML-инжиниринг для больших языковых моделей

★ 7.3 · ai-tooling

ml-engineering is a Claude Code skill that provides field-tested methodology and concrete recipes for training and operating large-scale LLM, VLM, and multi-modal models end to end, distilled from actually training BLOOM-176B and IDEFICS-80B. It covers the full stack: benchmarking accelerators with tools like mamf-finder.py, SLURM/Kubernetes job orchestration, maximizing throughput and MFU via DP/TP/PP/ZeRO parallelism and activation recomputation, diagnosing loss spikes and NaN/Inf instabilities, checkpointing and fault tolerance for inevitable hardware failures, inference latency and KV-cache sizing, and debugging multi-node/multi-GPU hangs. Structured as an operator's runbook, it helps identify whether the actual bottleneck is compute, memory bandwidth, network, storage, or the dataloader — then points to targeted scripts, benchmark tools, and comparison tables. Ideal for ML engineers hitting OOM errors, low TFLOPS, slow dataloading, NCCL/InfiniBand hangs, or checkpoint and preemption issues during large model training or fine-tuning.