Refine eight notebooks for DLI readiness: restructured sections and added more context #120

brycelelbach · 2025-12-12T00:59:40Z

No description provided.

… added more context

brycelelbach · 2025-12-12T21:43:40Z

tutorials/accelerated-python/notebooks/fundamentals/01__numpy_intro__ndarray_basics.ipynb

+      "source": [
+        "## 4. Aggregations and Axes\n",
+        "\n",
+        "When performing aggregations (like $\\text{sum}$, $\\text{mean}$, $\\text{max}$), you must specify the **Axis** you want to collapse (or reduce) the array along.\n",


Instead of using \text for code like ndarray, why not just use code font, e.g. backticks?

I think this would especially make sense as it follows what the other notebooks do.

brycelelbach · 2025-12-12T21:47:44Z

tutorials/accelerated-python/notebooks/fundamentals/03__numpy_to_cupy__ndarray_basics.ipynb

+        "if xp == cp:\n",
+        "    cp.cuda.Stream.null.synchronize()\n",
+        "\n",
+        "t1 = time.perf_counter()\n",


I'm a littel concerned about using %%timeit, %time, and perf_counter. They require us to add stream synchronizes, which means we have to explain them, and it's easy to forget them.

Why not use CuPy's benchmarking facility?

brycelelbach · 2025-12-12T21:48:51Z

tutorials/accelerated-python/notebooks/fundamentals/03__numpy_to_cupy__ndarray_basics.ipynb

+        "%%time\n",
+        "# GPU SVD\n",
+        "x_gpu = cp.random.random((1000, 1000))\n",
+        "u, s, v = cp.linalg.svd(x_gpu)"


This is missing a stream synchronize. I think it may be best to use cupyx benchmark instead of %%time.

brycelelbach · 2025-12-12T21:50:04Z

tutorials/accelerated-python/notebooks/fundamentals/05__memory_spaces__power_iteration.ipynb

+        "    max_steps: int = 400     # Maximum iterations\n",
+        "    check_frequency: int = 10 # Check for convergence every N steps\n",
+        "    progress: bool = True    # Print progress logs\n",
+        "    residual_threshold: float = 1e-10 # Stop if error is below this"


Please align all the comments so they begin at the same column.

brycelelbach · 2025-12-12T21:53:45Z

tutorials/accelerated-python/notebooks/fundamentals/05__memory_spaces__power_iteration.ipynb

+        "# Uncomment to test your implementation:\n",
+        "# A_test = generate_device_exercise()\n",
+        "# print(f\"Matrix shape: {A_test.shape}\")\n",
+        "# print(f\"Matrix is on GPU: {isinstance(A_test, cp.ndarray)}\")"


The original notebook was designed to highlight that NumPy and CuPy random number generation give you completely different results. I don't see that preserved in this version. Is there a place where we show that you get different answers for generate_host and generate_device? I think it's an important point to make.

brycelelbach · 2025-12-12T21:57:07Z

tutorials/accelerated-python/notebooks/fundamentals/06__asynchrony__power_iteration.ipynb

+      "provenance": []
+    },
+    "kernelspec": {
+      "display_name": "Python 3 (RAPIDS 25.10)",


This is the wrong kernelspec for this notebook. It should be "display_name": "Python 3 (ipykernel)". We don't have a RAPIDS 25.10 kernel in the Docker image. Please make sure this is fixed for all of the notebooks.

brycelelbach · 2025-12-12T21:59:30Z

tutorials/accelerated-python/notebooks/libraries/20__cudf__nyc_parking_violations.ipynb

@@ -1,2436 +1,667 @@
 {
- "cells": [


Can you ask @shwina to review this? He made this notebook.

brycelelbach · 2025-12-12T22:00:09Z

tutorials/accelerated-python/notebooks/libraries/23__cuda_cccl__customizing_algorithms.ipynb

@@ -1,1492 +1,1528 @@
 {
- "cells": [


Can you ask @shwina to review this? He made this notebook.

brycelelbach

These changes look good. I left some comments on feedback where I think changes may be needed, please take a look.

One broader remark: I noticed that you haven't updated the solution notebooks. The solution notebooks are copies of the exercise notebooks that have the exercises filled in. We use them to demonstrate to the class and to test the content in CI. Can you update them as well?

shwina

Nice work - I reviewed the notebooks on cuDF and cuda.cccl.

At a high level regarding cuDF, we may want to check with RAPIDS engineering/product folks whether we want tutorials to focus on the base cuDF API or the simpler cudf.pandas - or both.

shwina · 2025-12-18T13:43:40Z

tutorials/accelerated-python/notebooks/libraries/20__cudf__nyc_parking_violations.ipynb

+      "source": [
+        "## 1. Introduction\n",
+        "\n",
+        "In this notebook, we will build a foundation in data manipulation using **Pandas**, the industry standard for Python data analysis. Then, we will transition to **cuDF**, which allows us to run standard Pandas-like code on the GPU.\n"


I would keep the tone here a bit more neutral. Instead of industry standard, let's just call it a popular tool.

shwina · 2025-12-18T14:04:54Z

tutorials/accelerated-python/notebooks/libraries/20__cudf__nyc_parking_violations.ipynb

+        "\n",
+        "In Pandas, `.apply()` works because the CPU can execute your Python function one element at a time. On the GPU, this model does not work: a GPU cannot interpret Python bytecode. To make custom functions run on the GPU, cuDF uses Numba to compile your Python function into GPU machine code (PTX). That compilation step imposes strict rules:\n",
+        "\n",
+        "- The function must be Numba-compilable (pure math only; no Python objects).\n",


This is somewhat inaccurate. cuDF supports more than pure math operations in UDFs. For example, operations on strings are supported.

Perhaps we can point to the docs above regarding the features/limitations of apply() in cuDF.

shwina · 2025-12-18T14:08:51Z

tutorials/accelerated-python/notebooks/libraries/20__cudf__nyc_parking_violations.ipynb

+        "- faster\n",
+        "- simpler\n",
+        "- more readable\n",
+        "- the intended way to use GPUs\n"


Suggested change

"- the intended way to use GPUs\n"

.apply() isn't inherently a bad way to use GPUs

shwina · 2025-12-18T14:14:18Z

tutorials/accelerated-python/notebooks/libraries/23__cuda_cccl__customizing_algorithms.ipynb

+      "id": "bf385d3d",
+      "metadata": {},
+      "source": [
+        "## 6.3 Transform with Iterators for Memory Efficiency\n",


We introduce the concept of iterators later in this notebook, so we may want to move this example to that section or after it.

ncclementi · 2025-12-18T16:58:04Z

+1 to all the comments that @shwina left regarding cuDF.

The content is a bit out of date too, I have worked on more up to date content regarding cudf and cuml too, not sure if you'll want to include cuml in there too. Here is the altest iteration https://github.com/rapidsai-community/tutorial

Now, I'm not representing cudf product, so I'm not sure what do they would like to focus on, getting @btepera's opinion here would be valuable.

btepera · 2025-12-18T20:41:30Z

It'd be good to include cudf.pandas in this tutorial, particularly since the notebook is already showing how to perform all of these operations in pandas as a starting point. As cudf.pandas has matured in terms of feature completeness, we have leaned more heavily on it (rather than cudf classic) as an entry point for users. If the target audience for this notebook is "users new to GPU-accelerated data science" then I think cudf.pandas makes a lot of sense.

ncclementi · 2025-12-18T20:48:25Z

@nv-kriehl based on @btepera comment above. I can happily help modifying the notebook to be more focused on cudf.pandas. If you think this notebook plus script (to learn about profiling and how to use cudf.pandas with scripts too) I'm happy to contribute this notebook (or a version of it if needed) instead here.

https://github.com/rapidsai-community/tutorial/blob/main/2.cudf_pandas.ipynb

Refining eight notebooks for DLI readiness: restructured sections and…

0337d44

… added more context

brycelelbach requested a review from nv-kriehl December 12, 2025 00:59

This was referenced Dec 12, 2025

CUDA Python DLI content #118

Closed

DLI updates #117

Closed

brycelelbach commented Dec 12, 2025

View reviewed changes

shwina reviewed Dec 18, 2025

View reviewed changes

Refine eight notebooks for DLI readiness: restructured sections and added more context #120

Are you sure you want to change the base?

Refine eight notebooks for DLI readiness: restructured sections and added more context #120

Uh oh!

Conversation

brycelelbach commented Dec 12, 2025

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

brycelelbach left a comment

Choose a reason for hiding this comment

Uh oh!

shwina left a comment

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

ncclementi commented Dec 18, 2025 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

btepera commented Dec 18, 2025

Uh oh!

ncclementi commented Dec 18, 2025

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

6 participants

ncclementi commented Dec 18, 2025 •

edited

Loading