Tag: pytorch

Why Apple Silicon is useful for local LLM prototyping and LoRA experiments, but still has sharp boundaries compared with CUDA-scale NeMo or Hugging Face training.

Outcome: Separated Mac-local MPS and MLX fine-tuning paths from NVIDIA-only training features so local experiments can start with realistic hardware expectations.

A production-friendly pattern for pairing scikit-learn preprocessing graphs with PyTorch models so training and inference use the same feature contract.

Outcome: Defined an artifact contract that keeps column preprocessing, feature order, model weights, metadata, and inference behavior synchronized across batch and serving environments.

A repeatable triage routine — the three-run baseline, DataLoader diagnosis, five profiler signatures, and a copy-paste scaffold — for finding where training time actually goes before touching the model.

Outcome: Identified and resolved training bottlenecks in under an hour by running the three-run baseline and reading profiler signatures before changing any model code.

torch.compile, mixed precision, gradient accumulation, DDP vs FSDP, and the profiler — the five levers I reach for before rethinking the model architecture.

Outcome: Cut training wall-clock time and GPU memory pressure by applying compile, AMP, and accumulation patterns in sequence before ever touching model architecture.