All skills
dali-dynamic-mode
Write, review, or migrate DALI dynamic-mode (ndd) code using direct imperative operator calls.
- Part of NVIDIA/skills
- 2.1k installs
- Machine Learning · Training frameworks
When to use it
You have DALI pipeline code using @pipeline_def and pipe.run() and want to convert it to plain Python loops, or you're writing new ndd code and need correct reader, batching, and device handling.
For example
You say
Convert this DALI pipeline-mode image classification script to dynamic mode (ndd): it uses @pipeline_def, fn.readers.file, fn.decoders.image with device="mixed", and pipe.run() with batch_size=64.
You get
- Rewritten ndd script
- Reader and batch_size fixes
- device="gpu" decode change
How it works
- Readers are stateful: create once and call next_epoch(batch_size=N) each epoch instead of rebuilding.
- Batch has no indexing; use batch.tensors[i] to pick samples and batch.slice[...] to slice within samples.
- Checkpointing covers only readers and RNG objects, and readers must opt in with enable_checkpointing=True.
Before installing
NVIDIA DALI must be installed with nvidia.dali.experimental.dynamic importable; GPU decode and GPU operators need a CUDA-capable DALI build and an NVIDIA GPU/driver.
This skill includes a script your assistant may run on your computer.
More Training frameworks skills
See allAlso in this pack
- accelerated-computing-cudf3.2kGuides GPU DataFrame work with cuDF, including pandas acceleration, joins, groupby, file I/O and multi-GPU jobs.
- cuopt-developer2.8kModifies, builds, tests, and debugs NVIDIA cuOpt source code and prepares contributions.
- cuopt-install2.6kInstalls cuOpt for Python, C, or server via pip, conda, or Docker and verifies it works.
- data-designer2.6kGenerates a synthetic dataset from your description of the data you need.