DALI dynamic-mode — написание и миграция кода

★ 7.6 · content

dali-dynamic-mode is a Claude Code skill that guides AI agents in writing, reviewing, and migrating code that uses DALI's imperative dynamic-mode API (`nvidia.dali.experimental.dynamic`, `ndd`). It enforces direct `ndd` operator calls from ordinary Python control flow, explicitly prohibiting pipeline-mode constructs such as `Pipeline`, `@pipeline_def`, `pipe.build()`, and `pipe.run()`. The skill covers core data types — `Tensor` and `Batch` with `batch.tensors[]` for sample selection and `batch.slice[]` for per-sample slicing — stateful readers that pass `batch_size` to `next_epoch()`, `device="gpu"` instead of the pipeline-mode `"mixed"` keyword, and PyTorch conversion via `.torch()` with `pad=True` for variable-shape batches. It is the right tool when building or porting GPU data-loading and preprocessing pipelines with `ndd`.