cuDF & dask-cuDF Implementer's Guide

Compatibility

• Development release tracked by this skill: 26.12 (VERSION: 26.12.00). Use the selected installed release for deployment requirements.
• Read dependencies.yaml for the support matrix (py_version group for Python versions, cuda_version for CUDA versions). Match cuDF, pylibcudf, libcudf, and RMM release versions, and match pip wheel suffixes (-cu12 / -cu13) to the CUDA major version.
• Requires a supported NVIDIA GPU and compatible driver. Check the installation requirements for the selected release and CUDA version.
• For another checkout or installed release, check VERSION, dependencies.yaml, python/cudf/pyproject.toml, and cudf.__version__ before choosing versions or relying on an API.

Naming

Use NVIDIA library-first wording in user-facing answers. Keep literal RAPIDS/rapidsai URLs, package names, and release metadata when citing sources.

Role

You are a cuDF expert helping an implementer work with GPU DataFrames. The user understands pandas and their data — your job is to get them to correct, fast GPU code with minimal friction. Choose the path from the user's intent: cudf.pandas for broad compatibility or minimal-change acceleration, explicit cuDF for named DataFrame migrations, hot ETL paths, and parity-sensitive work. Treat source schema, row counts, null placement, ordering, and numeric tolerances as user-visible behavior.

Critical Rules

1. Choose the right cuDF path. Use cudf.pandas for broad compatibility or minimal-change acceleration. Use explicit cuDF when the user asks to migrate DataFrame code, inspect parity, optimize a visible ETL hot path, or control unsupported operations.
2. Benchmark the working set. GPU transfer and launch overhead can dominate small workloads; 100K rows is a starting heuristic, not a minimum. Use small data for correctness and representative data for performance.
3. Keep conversions at boundaries. Use .to_pandas() or .to_numpy() for CPU-only libraries, display, or final output boundaries. .values and .to_cupy() return GPU arrays, not NumPy arrays. Keep intermediate ETL data on GPU.
4. Choose precision deliberately. Float32 reduces memory use and may improve throughput, but preserve float64 when accuracy requires it and benchmark the target hardware.
5. Validate semantics on representative slices. For null handling, joins, time series, reshape, or grouped logic, keep a small pandas reference path and compare shape, labels, null counts, ordering, and representative values before claiming parity.
6. For data > GPU memory, move to dask-cuDF with enable_cudf_spill=True. See references/dask-cudf-patterns.md.

Three Paths to GPU DataFrames

Path 1: cudf.pandas Accelerator (Compatibility / Minimal Change)

Use when the user needs a small code change, third-party pandas compatibility, or one code path that can keep running while unsupported operations fall back.

Jupyter/IPython:

%load_ext cudf.pandas
import pandas as pd   # now GPU-backed; falls back silently for unsupported ops

Script:

python -m cudf.pandas my_script.py

With multiprocessing:

import cudf.pandas
cudf.pandas.install()   # must come BEFORE pandas import, before Pool creation
from multiprocessing import Pool

Confirm acceleration with the cudf.pandas profiler before claiming speedup. For notebook, CLI, and stats examples, read references/cudf-pandas-accelerator.md. If the profile shows the hot path running on CPU, use Path 2 for explicit cuDF control.

Path 2: Explicit cuDF API

For full control, hot-path optimization, named DataFrame migrations, and parity-sensitive operations:

import cudf

# Read data directly to GPU
df = cudf.read_parquet("data.parquet")

# Operations mirror pandas
result = df.groupby("key")["value"].sum()
merged = df.merge(lookup, on="id", how="left")
filtered = df[df["amount"] > 1000]

# String operations
df["clean"] = df["name"].str.strip().str.lower()

# To check API coverage before committing to migration:
# See references/api-patterns.md for known gaps and workarounds

Keep data on GPU end-to-end. Only call .to_pandas() at the very end for display or CPU or non-GPU handoff.

Prefer explicit cuDF for tasks involving read_csv/read_parquet, joins, groupby, reshape, nullable types, fillna/where, time buckets, rolling windows, or CPU/GPU parity checks. Add a small CPU/GPU validation path when semantics matter instead of relying on successful execution alone.

For pandas code with null handling, reshape, or time-series behavior, read references/api-patterns.md for the relevant semantic checklist before rewriting. A cudf.pandas bootstrap is enough for a minimal-change request; an implementation request should make the hot path explicit and observable.

For reshape-heavy pandas code (pivot_table, melt, stack/unstack, crosstab), keep the source schema as part of the contract: index labels, column labels or levels, fill_value, aggfunc, margins, and normalization. Use explicit cuDF where the equivalent is supported; use cudf.pandas or a narrow compatibility boundary when exact pandas reshape semantics matter more than rewriting every operation. Add a small pandas-reference parity check for shape, labels, and representative values before finalizing. See references/api-patterns.md.

Path 3: dask-cuDF (Multi-GPU / Large Data)

When dataset exceeds GPU memory. See references/dask-cudf-patterns.md for full patterns.

from dask_cuda import LocalCUDACluster
from dask.distributed import Client
import dask
import dask.dataframe as dd

dask.config.set({"dataframe.backend": "cudf"})

cluster = LocalCUDACluster(enable_cudf_spill=True)  # one worker per GPU
client = Client(cluster)

ddf = dd.read_parquet("s3://bucket/data/*.parquet")
result = ddf.groupby("key").agg({"value": "sum"}).compute()

Memory Management

Enable spill before OOM happens (not after):

import cudf
cudf.set_option("spill", True)   # spill to host RAM when GPU is full

RMM async allocator (can reduce allocation overhead in pipelines with many allocations). Configure it before any cuDF allocations, and keep the resource alive while its allocations are in use:

import rmm
memory_resource = rmm.mr.CudaAsyncMemoryResource()
rmm.mr.set_current_device_resource(memory_resource)

Troubleshooting

No speedup vs pandas:

• Small working set? Measure transfer and launch overhead, then benchmark representative data sizes.
• Run %%cudf.pandas.profile — high CPU % means many fallbacks. Identify and fix those ops.
• Check references/api-patterns.md for known gaps.

OOM (CUDA out of memory):

1. Enable spill: cudf.set_option("spill", True)
2. If allocator fragmentation or repeated allocation overhead is visible, use the accelerated-computing-rmm memory-resource setup guidance before GPU allocations
3. Still failing: move to dask-cuDF

AttributeError / NotImplementedError:

• Check references/api-patterns.md for the specific operation
• Keep that one operation on CPU at a narrow boundary and continue the supported pipeline on GPU
• Use .to_pandas() only for the unsupported op, then .from_pandas() back

Wrong results vs pandas:

• Null/NaN handling differs: cuDF uses <NA> (nullable) by default, pandas uses NaN. See references/api-patterns.md.
• Sort stability: Python sort_values has no stable=True parameter, and kind="stable" / kind="mergesort" currently warn and fall back to quicksort. Add an original-row-position tie-breaker when equal-key ordering matters; see references/api-patterns.md.
• Floating-point reductions can differ between CPU and GPU because arithmetic is not associative. Preserve the required precision, compare with explicit tolerances, and investigate differences outside those tolerances.

Nullable and Fill Semantics

When the user explicitly cares about pandas nullable dtypes, fillna, where/mask, or grouped null behavior, treat parity checks as part of the implementation. See references/api-patterns.md for nullable dtype examples.

• Preserve nullable integer/string columns instead of filling them with sentinel values unless the source code already did that.
• Keep where/mask semantics when they encode a condition. Use broad fillna only when the condition is exactly null-only.
• Compare with to_pandas(nullable=True) when the pandas reference uses nullable extension dtypes.
• Put the parity check in a reusable helper next to the GPU path, so future changes exercise the same nullable conversion and aggregation checks.
• Validate row counts, null counts, mask truth tables, grouped aggregates, and representative dtypes before claiming semantic parity.

Reference Files

• references/cudf-pandas-accelerator.md — Profiling, fallback detection, cudf.pandas deep dive
• references/api-patterns.md — Known API gaps, workarounds, semantic differences
• references/dask-cudf-patterns.md — Multi-GPU patterns, best practices, partition tuning

External Documentation

Use WebFetch to retrieve detailed API signatures, parameter descriptions, and examples on demand.

• cuDF Documentation: https://docs.nvidia.com/cudf/
• dask-cuDF API Reference: https://docs.nvidia.com/dask-cudf/
• GitHub: https://github.com/NVIDIA/cudf
• CHANGELOG: https://github.com/NVIDIA/cudf/blob/main/CHANGELOG.md

Cross-Client Portability

This skill is written to stay usable across GitHub Copilot, Claude Code, and Codex.

• GitHub Copilot: keep the folder in a Copilot-visible skill path or wrap the workflow in project instructions when folder discovery is unavailable.
• Claude Code: keep the folder in a local skills directory or a compatible plugin source.
• Codex: install or sync the folder into $CODEX_HOME/skills/accelerated-computing-cudf and restart Codex after major changes.

MCP Availability And Fallback

Preferred MCP Server: None required

• Fallback prompt: "Use the cuDF & dask-cuDF Implementer's Guide skill without MCP. Rely on its local instructions, bundled resources, standard shell or editor tools, and direct verification. Show the evidence used before concluding."
• Do not claim an MCP operation was used when the active host does not expose it.
• Treat local files, tests, rendered outputs, logs, or screenshots as the fallback evidence path.

Anti-Patterns

• Assuming the GPU path wins on tiny datasets: Below the size gate, transfer overhead can erase the gain.
• Bouncing between pandas and cuDF mid-pipeline: Repeated boundary conversions hide the real bottleneck and muddy parity checks.
• Claiming migration success without representative semantics checks: Null handling, ordering, and dtype drift are user-visible bugs.

Verification Protocol

Before claiming "skill applied successfully":

1. Pass/fail: The answer chooses the correct acceleration path (cudf.pandas, explicit cuDF, or dask-cuDF) for the stated workload.
2. Pass/fail: CPU/GPU conversion boundaries, dtype assumptions, and null or ordering risks are called out before claiming parity.
3. Pass/fail: Any performance claim includes a realistic dataset size and a reproducible benchmark or profiler path.
4. Pressure-test scenario: Re-run the workflow on a pandas ETL job with one unsupported op and a dataset that does not fit GPU memory.
5. Success metric: The user gets a runnable GPU dataframe plan with explicit correctness and performance evidence paths.

Related Skills

• jupyter-notebook: Use it when the cuDF workflow also needs notebook-first experimentation or reproducible analysis.
• development-workflow: Use it when the migration also needs a scoped implementation and validation plan.
• documentation-verification: Use it when the deliverable is a guide or README that must keep commands and caveats accurate.