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 cuDF Accelerated Computing Skill
NVIDIA's own guidance for moving pandas workloads onto GPU DataFrames with cuDF and dask-cuDF — when it pays off, and how to keep results identical.
Install
Other ways to install
About
- Chooses between cudf.pandas, explicit cuDF and dask-cuDF from the user's intent
- Enforces a 100K-row size gate so small data is not moved to GPU pointlessly
- Keeps intermediate data on GPU and converts only at output boundaries
- Validates null handling, ordering and numeric parity against a pandas reference
- Covers CSV/Parquet I/O, joins, groupby and multi-GPU workloads
- Tracks cuDF release 26.04 with explicit CUDA and driver requirements
An official, NVIDIA-authored agent skill that turns a coding agent into a competent cuDF implementer. Most "port this to GPU" attempts fail in one of two ways: the dataset is too small for the transfer cost to ever pay back, or the port silently changes null handling, row ordering or numeric precision and nobody notices until a downstream report is wrong. This skill encodes the guardrails for both.
What it teaches the agent
- Which of three paths to take.
cudf.pandasfor near-zero-change acceleration of an existing pandas codebase, explicit cuDF for deliberate migrations and hot ETL paths where parity matters, and dask-cuDF once the working set exceeds GPU memory (withenable_cudf_spill=True). - A size gate. Below roughly 100,000 rows the host-to-device transfer usually costs more than the speedup buys; the skill tells the agent to stop rather than produce a slower program.
- Where conversions belong. Keep intermediate ETL data on the GPU and convert at boundaries only — display, plotting, CPU-only libraries and final output — via
.to_pandas(),.valuesor.numpy(). - Parity discipline. For null semantics, joins, time series, reshapes and grouped logic, keep a small pandas reference path and compare shape, labels, null counts, ordering and representative values before claiming equivalence.
- Cheap wins, such as casting to float32 early where precision allows, since float64 operations are slower on GPU.
Bundled reference files cover dask-cuDF patterns, multi-GPU workloads, and CSV/Parquet I/O in more depth than the main file.
Compatibility
Tracks cuDF release 26.04. Requires NVIDIA Volta or newer on CUDA 12, or Turing or newer on CUDA 13; CUDA 12.2–12.9 needs driver 535+, CUDA 13.0–13.1 needs driver 580+. Python 3.11–3.14.
Installing it
Part of the NVIDIA agent-skills catalogue, which is mirrored daily from NVIDIA's product repositories and carries OMS signatures verifiable against nv-agent-root-cert.pem. Install the whole set with npx skills add nvidia/skills, or copy this one folder into your agent's skills directory. Works with Claude Code, Codex, Cursor and any runtime that follows the Agent Skills standard. Source is Apache-2.0, skill text is CC-BY-4.0.
Try it
Once installed, ask your agent something like:
"This nightly ETL job reads a 40M-row Parquet file and does three groupby aggregations in pandas. Can cuDF help, and what would change about the output?"
SKILL.md
The instructions your agent loads, as published by NVIDIA.
Related Skills
See all NVIDIA cuDF Accelerated Computing 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.