Diagnoses wrong gradients in differentiable NVIDIA Warp programs by measuring first — comparing autodiff against finite differences on a shrunk reproduction before proposing any fix.
NVIDIA DALI Dynamic Mode Skill
NVIDIA's official skill for DALI's imperative dynamic-mode API — write GPU data loading as ordinary Python, or migrate an existing pipeline-mode graph across.
Install
Other ways to install
About
- Writes DALI data loading as imperative Python with no pipeline graph
- Blocks pipeline-mode APIs that silently misbehave in dynamic mode
- Covers the Tensor and Batch data model, slicing and zero-copy wrapping
- Handles PyTorch hand-off via .torch(), including pad=True for ragged batches
- Guides migration of existing pipeline-mode code to ndd
- Supports DLPack, __cuda_array_interface__ and __array__ interop
An official, NVIDIA-authored agent skill for nvidia.dali.experimental.dynamic (imported as ndd), the imperative API that lets DALI operators be called from normal Python control flow instead of being wired into a pipeline graph. The skill exists because the two modes look similar and are not interchangeable: pipeline-mode habits carried into dynamic mode produce code that fails in confusing ways.
The rules it enforces
- No pipeline-mode APIs.
Pipeline,@pipeline_def,pipe.build()andpipe.run()have no place in dynamic-mode code; the agent is told to write directnddcalls instead. - Readers are stateful. Create them once, reuse them across epochs, and pass
batch_sizetonext_epoch(...)rather than rebuilding. - Batch size is explicit. There is no pipeline-level batch size to inherit, so random operators need it passed in.
- Convention differences that bite.
device="gpu"replaces pipeline mode's"mixed";Batch.tensors[...]selects samples andBatch.slice[...]slices per sample. - Framework hand-off.
.torch()converts a tensor or batch to PyTorch, withpad=Truefor batches of variable shape.
The data model it covers
Tensor is a single sample supporting ndd.tensor(data) (copy) and ndd.as_tensor(data) (wrap without copying where possible), .cpu() / .gpu() moves, slicing, np.asarray on CPU, and the __dlpack__, __cuda_array_interface__ and __array__ protocols plus arithmetic operators. Batch is a collection of samples that tolerates variable shapes.
Who it is for
Teams writing or reviewing GPU data-loading and augmentation code, and anyone migrating an existing DALI pipeline to the imperative API. The skill deliberately stands down on pipeline-only tasks rather than volunteering advice that does not apply.
Requirements
NVIDIA DALI installed with dynamic mode importable as nvidia.dali.experimental.dynamic. GPU decode and GPU operators need a CUDA-capable DALI build and an available NVIDIA GPU and driver; the PyTorch conversion examples need PyTorch installed.
Installing it
Part of NVIDIA's agent-skills catalogue, mirrored daily from NVIDIA's product repositories. npx skills add nvidia/skills installs the set, or npx skills add nvidia/skills --skill dali-dynamic-mode --yes fetches just this one. Apache-2.0.
Try it
Once installed, ask your agent something like:
"Migrate this @pipeline_def image-augmentation pipeline to DALI dynamic mode, keep the reader alive across epochs, and hand the batches to PyTorch without an extra copy."
SKILL.md
The instructions your agent loads, as published by NVIDIA.
Related Skills
See all NVIDIA DALI Dynamic Mode alternativesSkill: NVIDIA TAO RT-DETR
by NVIDIA
Trains, distils, quantises, evaluates, exports and runs inference for RT-DETR real-time object detection in NVIDIA TAO, routing training through AutoML hyperparameter search by default.
Skill: NVIDIA Holoscan SDK Setup
by NVIDIA
Inspects the host's hardware, OS, CUDA driver and existing tooling, recommends one Holoscan SDK install method with a reason, then hands off to the matching install skill.
Skill: NVIDIA NV-Reason-CXR
by NVIDIA
Runs NVIDIA's NV-Reason-CXR-3B chest X-ray reasoning model through a documented wrapper, locally on GPU or via the public Hugging Face Space, and returns the model's full reasoning trace as engineering evidence.